Last Tuesday, Google held its annual I/O conference, and I’m one of those people who watch the entire keynote. I came away both overwhelmed and underwhelmed. The new models are genuinely impressive, with real agentic capability and a new fast-and-cheap model, Gemini 3.5 Flash, that does frontier-level work at a fraction of the cost. Charts sped by showing AI usage exploding across the Google ecosystem. The technology is real, and it is improving ahead of schedule. What frustrated me was not the technology. It was the way Google chose to show us what they want us to use it for.

A handful of lines also stood out, particularly from Sundar Pichai, Google’s CEO, and Demis Hassabis, the CEO of Google DeepMind and the company’s resident visionary. Pichai casually folded in two examples of how people are using AI: students studying for final exams, and his own use of it to “make sense of my parents’ doctor visits.” Later, Hassabis stated, essentially as fact, that we are just a “few years away” from AGI. I know these conferences are victory laps, with their obligatory pauses for applause, like a State of the Union address. Even so, Pichai reached for two of the most contested uses of AI in the culture right now - student work and medical diagnosis - and presented both as if the debate were over and decided in their favor. Hassabis has moved his AGI timeline up to the point where he’ll say it flatly, on a stage he knows the world is watching. Each of these lines struck me wrong. But it was the demos where they really lost me.

A year ago I wrote about a significant development most people missed that rolled out at the same conference, and we’ve all been living with the consequences ever since. AI Mode, built directly into the search bar, is now just part of the Google experience. This made AI unavoidable for the billions of people who use Google search, including our students. This year, Google is clearly doubling down on its AI infrastructure investment. To what end?

All That Intelligence, for This?

The demos, one after another, were mostly small in a way that’s oddly hard to describe. Gemini Spark, the company’s shiny new personal agent, was put to work organizing a block party. In Josh Woodward’s demo it tracked down RSVPs, updated a Google Sheet, drafted emails to the neighbors who hadn’t replied, and built a hype deck in Slides, all of it running in the background. A second presenter, Robby Stein, showed something different: generative UI in Search, where it built a custom app on the fly, in this case a weekend planner for his family, which he then shared with his wife for her seal of approval. A little later came the universal shopping cart that follows you across the web, and that was the clincher for me. All that money, all that intelligence, and it was mostly pitched as a better way to throw a barbecue and a faster way to check out after an online shopping spree.

The promise of all this is that AI does things for you. The demos kept showing me how AI could build things, even directly within the search box.1 I laughed at the weekend planner, because I know exactly what would happen if I built an app to plan our weekend and sent it to my wife. She would probably throw her phone at me. Who actually wants this? When I use Google Maps to find a coffee shop, I want the coffee shop, not to spend time fiddling with a custom-built coffee-shop-finding dashboard I can come back to later.

Over and over, they handed me solutions to problems I don’t have. So much of it felt less like help and more like homework, an endless proliferation of apps and dashboards for things families have managed on their own for decades. Google’s bet is that making these will become so fast and so frictionless that we’ll reach for them without thinking. Maybe they’re right. I’m skeptical.2

It’s Shopping All the Way Down

The part that left me most cynical is how obviously Google wants us to shop more. People already shop across Google more than a billion times a day, and the company intends to make that even more effortless. The new shopping agents work around the clock. Your cart follows you everywhere, hunts for deals while you sleep, and can be authorized to buy on your behalf with you out of the loop. One slide called it “shopping with superpowers.” The commerce segment ended on the words “happy shopping.” Then, just in case the message hadn’t landed, they scored the transition to Depeche Mode’s “Just Can’t Get Enough,” hitting Gen X squarely in the solar plexus.

I’ve no doubt some people may love this. Still, a discount on something you’d never have bought without a shopping agent doesn’t count as “savings.” It’s just a more efficient way to spend more. “Shopping with superpowers” is really a demand that we all shop with superhuman discipline, which is not a thing most of us are good at. We’ve already built a world where we can want something at eleven at night and have it on the doorstep the next day. That last layer of friction, between the impulse and actually clicking Buy, was arguably the one thing protecting some of us from ourselves, and Google’s plan is to remove even that. In a spring when gas is closing in on five dollars a gallon, inflation is climbing, and consumer confidence just hit an all-time low, do young people, and the rest of us for that matter, really need to be trained to turn every passing whim into a transaction?

Glasses That Can Whisper the Answers

Near the end came the demo I’d been dreading, and the one that should scare teachers most. Google’s first audio glasses, which speak information privately into your ear, arrive this fall; the display glasses, which show it on a lens in front of your eye, are still in limited testing. The demo was kept deliberately mundane, because the general creepiness of AI headgear is the part Google would rather not dwell on. Shahram Izadi, Google’s XR lead, and Nishtha Bhatia walked through ordering a cold brew through the frames, summarizing muted texts, and turning the crowd into a cartoon on her watch. Cool, I suppose, but do consumers really need this?

For a teacher, a device built to feed you information invisibly, whispered in your ear or shown on a lens only you can see, is an assessment-integrity problem on a scale we haven't faced, and it is not waiting for Google's fall launch. You can already buy prescription glasses with AI built in. Meta has sold millions, and a student in my class had a pair last year. Can we ban prescription AI glasses? The blue book and the in-class exam, the defenses we've fallen back on this year, do not survive glasses that give students answers in private. Beyond cheating, the glasses are just the next layer of AI envelopment Google keeps working on to make it all feel inevitable, arriving at a moment when the backlash is real and a vocal public is asking everyone to slow down. The whole keynote felt tone-deaf to it.

The Argument They Skipped

The debate about AI and education is the one the keynote treated as a closed case. The references that did appear were all upbeat - Pichai's line about students prepping for final exams, and Josh Woodward's quick run through Gemini's student features, guided learning, practice tests, and a nod to NotebookLM.3 The closest the keynote came to demonstrating actual learning was Robby Stein's astrophysics moment, where Search built a custom interactive visual explaining how black holes affect space-time. Even that was staged as an exciting feature of what students can do. Yet the question that keeps teachers up at night, whether these tools are actually helping students learn, never came up. Google ignored the most contested issue in my world and declared it settled while the rest of us are still figuring it out. It won the argument by declining to have it.

After the keynote, Dan Fitzpatrick (no relation), a highly sought-after voice in the AI-and-education world, sat down with Chris Phillips, the man who runs education at Google, and came away with a much more nuanced, teacher-centered case the keynote never bothered to make. His piece is worth reading in full.

Two of his takeaways landed, for different reasons. The first described one of the most impressive demos in which Antigravity 2.0, Google's revamped agentic coding platform, built a working operating system from scratch, mostly on its own, for under a thousand dollars. Building something that complex for that little has real ramifications for developers, but it is not likely to impact teachers any time soon. It is just another example of the gap between what the frontier can do and what the 2026 classroom is actually concerned with.

His larger point is that Gemini 3.5 Flash is now good enough to do almost everything schools are asking for, and the new price point will lower the cost barrier keeping AI out of districts willing to make the switch but unable to commit the resources. I suspect it will make AI in schools more likely, not less.

Though Dan’s piece might be interpreted as more bullish than I am on where AI in schools is headed, it did not offset my visceral and mostly negative reaction to the keynote itself.

Even though it’s clear from his reporting that Google has a thoughtful account of what its tools do for learning, it’s telling that it chose not to put that on the stage.

The Promise vs. the Product

The keynote opened on a sizzle reel of possibility - a young boy on a tablet doing something undefined but clearly interesting with AI, a teacher mid-music-lesson, a few small-business owners, over a voiceover about solving disease and “making something that matters.” There were legitimate use cases: a Korean immigrant who used these tools to build a hiring platform for her neighbors, and creative AI tools for musicians and filmmakers.4 It closed on Hassabis invoking the “foothills of the singularity.”

Hassabis is the AI paradox in one man. He came out early to announce that AGI is just a few years away, and he returned at the very end as the closer, touting AlphaFold, Isomorphic Labs chasing the goal of one day “solving all disease,” and WeatherNext, the forecasting model that called Hurricane Melissa’s path days early, in time to get people out of the way. These are all real ways AI has advanced scientific development. He promises a “new golden age of scientific discovery” that improves lives everywhere. The demo that ran just before that close was the cold brew, ordered through a pair of glasses. From a cold brew to the singularity in a single cut. Everything between his two appearances, the apps, the shopping, the glasses, was the version of AI its skeptics fear most. Which future is Google actually building?

To What End?

Pichai mentioned that Google expects to spend something like $180 to $190 billion on infrastructure this year, roughly six times what it spent in 2022. Spark, the advanced agentic tool covered in the keynote that works while your laptop is closed, requires a Google AI Ultra plan, which runs between $100 and $200 a month. What's free for everyone is the new search box, the on-the-fly generative results, and the cart. Google would tell you the cart is what pays for the moonshot. Maybe. AGI is held out as a tantalizing promise, but the cart is what turns a billion shopping searches a day into the money that pays for it.

There were genuine standouts - the aforementioned Antigravity build and Gemini Omni, the new video model, were both impressive. With Google Flow’s editing tools I could see both the fear and the possibility in it for anyone who fools around with photos, film, and music. We will likely absorb Spark, the in-search app builds, and the universal cart the same way we absorbed AI Overviews, without too much fanfare. Maybe I’m just grumpy, but a lot of it felt more like gimmicks than upgrades.

This all tells me that with agentic AI now baked into how we use Google, the fantasy of “banning” AI in schools should be over, if it was ever a serious position to begin with. Yet we still barely know whether, or how, to teach with these tools at all. For the first time, as someone usually curious and fascinated by this stuff, I’m feeling more AI fatigue than excitement. Maybe I am just suffering from AI exhaustion.

Still, what stuck with me most was the shopping. The direction Google is pointing, as best I can tell, is to get me to buy more, build disposable apps to help me do it, and surrender to a world where every experience is something to digitize, optimize, and monetize. AGI promised something bigger than this. Do I really need a customized shopping experience, or bespoke apps for using the web? And do I want a pair of glasses talking into my ear or digitally interpreting my world so I never have to engage with it? The 2026 keynote was a tale of two futures, and the one that feels likeliest after watching it is the one I want least.

Connect With Me

Beyond this newsletter, I work directly with schools, educators, and other institutions about pedagogical questions raised by AI. Take a look at my website and reach out - I’d love to hear what you’re working on.

It’s mid-May and there is light at the end of the tunnel. Our school year ends in three weeks. Instead of anticipation, what I feel is something closer to exhaustion. The exhaustion is partially due to an especially brutal workload, but it also comes from spending nine months watching the gap between what AI actually is and how schools engage with it grow wider, not narrower. I feel like the conversation itself is stuck.

Below I’ll share a few observations as we wind down the academic calendar. None are especially surprising or groundbreaking. Taken together, though, they suggest significantly more work ahead for teachers, students, and institutions. My key takeaway is that the 2025 - 2026 school year will be remembered as the one where AI became ambient, unavoidable, and still mostly unaddressed.

Ambient AI

At the risk of stating the obvious, AI is everywhere. It’s still mostly associated with ChatGPT, but as of May 2026, unless you live under a rock, almost everyone is aware that several major companies have powerful AI models. More importantly, regardless of whether you know the names, understand the difference between Claude Code and Codex, or have experimented with any agentic capabilities - everyone on the internet has encountered AI images, audio, and sophisticated AI-generated content.

This was not true a year ago. The deep reasoning models had only emerged at the end of 2024 and I recall shocking teachers with demonstrations of what they could do last spring. The speed with which add-ons, plug-ins, and upgrades have rolled out means we are well past the ability to shock.

The implication for educators is simple. Unless we are prepared to keep our students from accessing Google, they will always be able to use a powerful frontier AI model in the form of Gemini through AI mode. I’ve argued since last June that teachers who claim they “forbid” AI are engaging in semantic theatre once a student is on the internet - we all encounter AI overviews whether we like it or not. Google’s AI mode allows for the critical cognitive off-loading educators fear, with the ability to have an LLM write, analyze, and work with uploaded documents baked into search. Literally the same capabilities students can access through ChatGPT. They do not need paid models.

Any notion that AI will not have a significant impact on teaching and learning is pure nostalgia.1

Schools Are Finally Reacting, But The Gap Is Still Growing

The good news is that schools are finally taking AI seriously beyond writing policies. More administrators and district leaders are grappling with what’s actually happening rather than merely drafting documents that are increasingly opaque, difficult to enforce, and not designed to withstand the speed with which the technology is moving.

The bad news is there is no consensus on what the reaction should be. The recent trend is to zero in on devices themselves - a movement grounded in legitimate concerns but one that also conveniently feels like it “solves” the AI problem. If students don’t have their phones or access to laptops, there’s no technology to turn to. Essentially, it asks us to return to an idealized past where kids all wrote by hand and played outside. As a father of a 6th grader, I see the appeal. I really do.

But the majority of school districts already have technology so embedded in their infrastructure that an unraveling is not nearly as easy as it looks. Many are going in the opposite direction, having invested heavily in a future where AI is a significant part of their mission. With over 13,000 districts in the U.S., we will see almost every permutation along the technology and AI adoption spectrum. At least these experiments will give us more data and more opportunities to evaluate what works.2

The deeper problem is that even the conversations finally taking place are still mostly happening in the context of 2023 AI. Few schools that I know of are seriously engaging with where the models are now and where they are heading. Over the past year, AI has moved well beyond the text generation that prompted most school policies - it now builds functional applications, conducts multi-step research, and produces professional-quality documents and presentations, capabilities that most school conversations haven’t begun to address. And these are the paid frontier models - what the average teacher or administrator is actually encountering is likely already a generation behind.

I agree with those who argue that mastering current models shouldn’t be the goal of AI literacy - they will continue to evolve and what we use in May 2026 will not be the same as September 2026, let alone May 2030. But I depart from those who claim that learning AI is “easy” and that our responsibilities as educators do not include preparing students for the world they are actually entering.

The reason I disagree is the same reason I would have pushed back in the early 2000s when internet access arrived at scale. We had responsibilities then to demonstrate best practices, explain how to evaluate information, and model effective use. Some of those early digital literacy strategies didn’t hold up over time - and that will likely be the case for some of what we teach about AI today. But talking about AI in the abstract without showing students what these tools do, how they work, and what kinds of output they produce is not serving them well. Especially when we know almost all of them are using it outside the classroom.

The Big Questions Are Finally Landing

Perhaps the most meaningful development of the year is that AI has forced educators to confront questions about purpose. More people - even those most viscerally resistant to AI - are recognizing that these systems will likely be a significant part of our lives going forward. Skeptics are also wrestling with something harder to dismiss: that AI can organize, synthesize, and generate information in ways that people engaged in knowledge work find genuinely useful whether or not those opposed to it use AI themselves.

These capabilities are forcing hard, but necessary conversations. Not simply about cheating, which dominated the past three years, but about purpose. What are schools actually for when obtaining and presenting information - in virtually any format we require - is approaching zero cost? What does a research assignment mean when AI can search, synthesize, and organize information faster than the student writing it? These are existential questions, no longer the province of the futurists and technologists who have been raising them for years. They are finally landing in faculty meetings and department conversations - perhaps under duress, but landing nonetheless.

Maybe it was never about AI in the first place. It’s always been about the value of what we assign and what we ask students to do - questions we could avoid as long as the products students handed in were plausibly their own. It’s dawning on most of us, if we hadn’t realized it already, that the skills hardest to automate are the ones the humanities have always tried to teach: empathy, curiosity, the ability to ask good questions, and the patience to sit with ones that don't have clear answers. It's worth noting that many in the industry are arriving at the same conclusion.3

What It Cost Me This Year

I want to add one observation that is entirely anecdotal but may resonate with other teachers. I covered less content this year. The constant vigilance about AI, the reduced number of assignments I gave outside of class, the time spent on research projects completed under supervision rather than at home - it all added up. I simply could not cover as much material as I have in previous years.

I don’t know if this was specific to my circumstances or something more widespread.4 But it strikes me as one of the hidden costs of this transition that we are not yet accounting for. AI has added more monitoring, more in-class assessment, and more rethinking of assignments to a curriculum that was never conceived with it in mind. If schools don't change their approach, something has to give - and right now, for me, it's content. Maybe that’s not the worst thing when virtually unlimited access to information and the ability for a motivated student to learn on their own are higher than they’ve ever been. But Sal Khan himself recently acknowledged that Khanmigo was a 'non-event' for most students - they simply didn't use it.5 Students are constantly using AI, just not the way adults design it for them.

As I Head Into Summer …

2025 - 2026 is ending more or less where we started, but with even more uncertainty - and significantly more work ahead within our schools, with our students, our parents, and our colleagues, to catch up to the reality of what AI has done to the academic project. I can’t imagine 2026 - 2027 won’t bring more of the same.

I’m exhausted by it and I suspect many teachers are as well. But the questions won’t disappear just because we’re afraid of asking them. If anything, we’re just getting started.

Connect With Me

Beyond this newsletter, I work directly with schools, educators, and other institutions about pedagogical questions raised by AI. Take a look at my website and reach out - I’d love to hear what you’re working on.

In March of 2020, millions of teachers and students across the country pivoted online for distance learning because of the pandemic. Districts that did not have 1:1 laptop programs swiftly invested in Chromebooks for every child. Despite the imperfect nature of the experience, we would not have been able to teach or connect with our students without those devices. Chromebooks became a lifeline between many students and school, particularly those in the least resourced communities. Yet, even for schools with 1:1 devices before 2020, most never carefully designed pedagogy around providing every child with one. Recent cell phone bans and the rise of generative AI have created a perfect storm for an anti-screen movement. The impulse is both understandable and rooted in legitimate concern. My fear is that the same lack of forethought that led to bringing laptops into the classroom in the first place will result in an overreactive push to eliminate them at exactly the moment students most need to learn how to navigate an AI-infused digital landscape.

Recent mainstream media articles report an accelerating trend.1 Parents and policy makers are demanding technology rollbacks in Los Angeles, New York, and elsewhere. Some middle schools have limited student access to Chromebooks. Others are trying to get rid of them altogether. The reporting describes negative reactions to the prevalence of screens in classrooms, with AI as a visceral catalyst that has helped transform long-standing discomfort into a national movement. The backlash is both understandable and real, and the responses range widely - from sensible restrictions on where, how, and for whom the devices are used, to outright calls for removal.

Bans are not the same thing as reforms - the distinction matters.

Any teacher who has spent time in a classroom over the past decade knows that screens can be the death of attention. I am sympathetic to the occasional urge to remove them. Yet, for lessons and activities designed well, digital devices that connect to the internet are one of the most powerful tools ever devised. Everyone reading this piece is either using a cell phone, a tablet, or a computer. Internet access is a necessity to participate in modern life.

The clash of these two realities - the difficulty of maintaining attention in the presence of a phone or computer that can deliver and create unlimited content - is at the heart of the conflict. At what age and in what circumstances can students be expected to responsibly use a device? How can teachers leverage intentional digital instruction and what can schools do to best ensure those devices are used appropriately? What does it mean when every child has their own device in the current paradigm, and what would it mean if we suddenly took these away?

The fundamental question around the issue is: Are the devices inherently the problem, or is the fact that we never developed real pedagogy around them - despite over a decade of trying - the real culprit? Some in the screen-removal movement increasingly claim it’s the laptops themselves. I've seen Chromebooks and classroom technology work for kids. Like almost everything when it comes to teaching, the results are context and classroom dependent.

Chromebooks were issued almost as a wartime emergency measure in 2020 and 2021. The devices have subsequently let in, among other things, first YouTube and now AI as a Trojan Horse, causing many of the problems outlined in the articles. Rather than ditching devices entirely, a better strategy is to pinpoint where the technology is doing more harm than good. Pulling the football away now, after a decade of telling kids and teachers digital and media literacy is essential, signals one thing above all: the adults have no idea what they are doing.

What the Movement Gets Right

Most of the horror stories in the recent coverage focus on the K-8 case, where the concern is clearly warranted. Before the pandemic, only about forty percent of districts had given devices to every elementary student. By March 2021, that number had jumped to eighty-four percent.2 The pandemic's distinctive contribution was pushing screens down to the youngest students at scale, before anyone had figured out what those devices were supposed to be doing in elementary classrooms. Putting screens in the hands of seven and eight year olds as a primary means of learning is not pedagogically defensible (though I suspect most elementary school teachers would chafe at the characterization of what they are doing). Nevertheless, the bombshells in the reporting - a seventh-grader in Wichita accessing more than 13,000 YouTube videos over three months on his school-issued account, and a Los Angeles 4th grade classroom where an AI art assignment produced sexualized images of a children's book character - are damning evidence that issuing 1:1 devices to younger students is walking a tightrope at best. It doesn’t take a neuroscientist to support eliminating or, at the very least, significantly restricting screen time for kids still learning how to read.

The Larger Failure We’ve Yet to Grapple With

The stories about Chromebook misuse and unlimited access to YouTube treat the screens themselves as the problem. A more nuanced view is that most districts that issued devices never built the institutional capacity to manage them. Many schools that gave kids the keys to the kingdom in 2020 did not plan for what classroom use would look like after the emergency. For example, as the WSJ reported, internal Google documents showed that by 2019 the company itself acknowledged the YouTube experience in schools was broken and that filters were trivially easy for students to bypass. Schools have known this for years. Most have done very little. But some have used a scalpel rather than a hammer: Watertown Public Schools in Massachusetts blocked YouTube districtwide in January. They did not eliminate Chromebooks.

What the Reversal Would Actually Cost

Banning devices from older students as well as younger ones also does not realistically engage with what a full reversal would require. More than 88% of public schools now operate 1:1 device programs, according to government survey data cited by the Wall Street Journal. Schools have rebuilt their pedagogy around digital delivery for more than a decade. Teachers have rewritten units, redesigned assessments, and built classroom workflows that assume students can access shared documents, online texts, video resources, all layered on top of learning management systems.3 Standardized assessments are now digital. For many students, a school-issued device is their only means of completing their work in the upper grades. The infrastructure is not removable on a whim. It is replaceable, but not in a year, and not without enormous transitional cost. We rolled these devices out under emergency conditions in 2020 because there was no time to plan. Even though we have more time now, nothing suggests we are using it any better. Recent research on cell phone bans in the United States and on Australia's under-16 social media ban shows how hard it is for bans to deliver the outcomes their advocates promise.4

The K-8 Case for a Ban Is Not the 9-12 Case

The screen-removal movement is least defensible where it shifts from real concerns about K-8 into a blanket case for total bans. The districts at the forefront of the movement are pulling or scaling back 1:1 for elementary, and a few are now extending that to middle school. McPherson Middle School in Kansas collected back student Chromebooks in December and now keeps them in classroom carts, available for specific assignments. Bend, Oregon blocked YouTube for lower grades and is reviewing high school access where it remains the most-used app. Granville County, North Carolina is phasing out 1:1 for elementary while keeping it for upper grades. Even the most aggressive recent moves are almost all grade-specific.5 The rhetoric, however, is not making the developmental distinction explicit. More significantly, if we overcorrect and ban all devices without thinking through the consequences, re-introducing them later will be that much harder once the political moment passes. Better to do the slower work of assessing what we have than to make rash decisions that cannot easily be undone.

Chromebooks for second graders are entirely different from high school seniors learning to research with AI in preparation for college. Younger students still primarily need handwriting practice and tactile, deliberate work with physical tools in the classroom. Twelfth graders are about to enter an environment in which AI is already embedded across every common platform, and most of their academic experience will be mediated through devices. These are not the same question. A screen-removal movement that treats them as one risks landing somewhere that will likely need revisiting within a few years. Too wide a pendulum swing will compound the mistakes.

A More Thoughtful Path

Most of what is actually being reported and adopted is not a ban. It is a rollback. The districts pulling back are restricting YouTube, eliminating devices for the youngest students, scaling back from 1:1 to classroom carts, and developing screen time limits for higher grades. These are scaled-back approaches that require careful management but preserve the option to use the devices when it makes sense.

The same type of approach could be applied more broadly. Restrict use in lower grades. Conduct better and more rigorous research in middle and high school on device usage. Lock down YouTube and other sites of maximal distraction. Give parents real opt-outs. Tech-Free Tuesdays and Thursdays is a more elegant solution than immediate removal. The screen-use ratio should reverse over the course of an academic career - nearly 100% analog in lower grades, moving to more realistic usage by senior year of high school. As with so many debates, framing the issue as a binary - ban laptops or keep them - is what makes the conversation so contentious.

A Year Later: What’s the Message?

What message would laptop bans send to students? A year ago, the New York Times was profiling Miami-Dade for rolling out Google’s Gemini to 105,000 high school students. To my knowledge, the Times has not run a follow-up on how that experiment is going.6 Notably, the same reporter who profiled Miami-Dade's rollout, Natasha Singer, has since written two of the articles driving the backlash narrative - but not a follow-up on whether the largest AI deployment in American public schools delivered results.

In March, the Times also reported that NYC Public Schools, the largest district in the country, released formal guidance designating AI lesson planning, brainstorming, and drafting as “green” applications. Two of the nation’s largest districts, plus a growing number of others including Chicago, Denver, and Charlotte-Mecklenburg, are leaning in carefully. Now the mainstream media is running cover stories on the screen backlash, with at least one arguing for wholesale device removal. Across the country, many states are finally catching up by developing AI literacy plans while others are pursuing rollbacks and investigating outright device bans. These messages are more contradictory than complementary.

This is whiplash for teachers as much as for students. Many state legislatures, education experts, and consultants have spent the past two years building AI literacy frameworks and integration guidelines on the assumption that students would have access to devices. The screen-removal movement now wants to remove those devices. The signal to everyone is that, after fifteen years of use, we still have no consensus on what these tools are for, how to use them, or whether they belong in classrooms at all.7 Students are at the mercy of the adults in charge, and the adults in charge keep changing their minds.

How Can We Best Serve Our Students?

My fear is that education's response to AI will be to retreat rather than to engage. AI literacy taught in the abstract is not really AI literacy at all. That is not the future many states have spent the past two years building toward. That does not mean integrating AI into elementary classrooms. It does mean addressing AI head-on, which includes explaining, demonstrating, and recognizing where students are using it. That requires them to have access to devices at some stage of their academic career.

Older students, especially those in the years prior to college and university, need hands-on experience and guidance with teachers about both practical and ethical AI use. What that precisely looks like, how often and in what context, will be debated for the foreseeable future. To pull the rug out from under both students and teachers after almost a decade of training on digital content delivery would send a powerful message that we are moving backwards.

For some schools and communities, a flat out device ban in certain grades may best reflect their values. Claiming that 1:1 laptop programs have been the problem all along, however, when so many of our challenges in education pre-date the pandemic is to make the same reactive mistake over and over again. Students will encounter screens, devices, and AI outside of school regardless of what their schools decide. For students with limited home access, school may be the only place they will encounter these tools under adult guidance. Removing the devices does not remove the technology from their lives. It just shifts where, when, and how they encounter it, and which students get any meaningful instruction at all.

The K-8 screen reckoning is overdue, and if AI is the reason it finally arrived, thoughtfully considered rollbacks happening now are largely the right response. But banning screens at every grade level, in reaction to a technology that has barely had time to be understood, would throw away almost fifteen years of pedagogical infrastructure to fix a problem most schools have not yet seriously tried to manage. Students need to learn to master the machines at some point in their education. When and how to do that in the face of emerging and powerful technologies will be an enduring question. Let schools and teachers make these decisions based on their own experiences, not in response to a reactive and contradictory national conversation simultaneously demanding AI literacy while calling for the removal of the devices students would need to develop it. We need to get this right.

I am especially interested in hearing from anyone making the case for high school bans. Tell me what I am missing.

Connect With Me

Beyond this newsletter, I work directly with schools, educators, and other institutions about pedagogical questions raised by AI. Take a look at my website and reach out - I’d love to hear what you’re working on.

A slew of frontier model updates and new capabilities have recently dropped in quick succession.1 One, Claude Design, got lost in the shuffle. While playing around with it over the past week, something became clear. Now that code generation rather than text is the leading edge of AI capability, it comes with a cost. AI is getting more complicated to use. Vibe coding was a thing a year ago2 - non-coders using natural language to make AI write code for them. What’s emerging now is something new: native interfaces specifically tailored to create things - a document, a presentation, an app, a web tool. Most people are not taking advantage. The barrier to entry to really leverage what the most advanced AIs are capable of now requires thinking like a coder.3

Unfortunately, I’ve never been a coder. Most teachers reading this probably aren’t either. An entirely new vocabulary, one I've been learning mostly through newsletters geared toward entrepreneurs, is leaking into everyday conversation: APIs, tokens, context windows, agents, MCPs, harnesses, CLIs. You can’t read deeply in the AI space now without encountering hyperventilating claims about new workflows that produce all sorts of new digital artifacts and prototypes. Software engineers, marketers, and designers are rapidly absorbing these new tools into their jobs. That’s the hype. But hype doesn’t mean things aren’t actually changing. My experience is ahead of most teachers', but the wall is real - and humanities teachers will hit it hardest.

For most of my professional life I knew, in some abstract way, that everything I typed and clicked online was built by someone in a backroom who knew how to make it work. That was the deal humanities educators made with the digital age. We taught our students how to read critically, write cogently, and think deeply, all while seamlessly using digital tools - the internet, computers, and software. The people who knew the secret language of the machines handled the machines. All that happened somewhere else.

That deal is changing.

The original promise of generative AI was democratization through natural language. Anyone able to type can access expertise. That was true for the first wave of chatbots and it’s still true today. A teacher can quickly draft a rubric or a lesson plan. Someone who knows nothing about auto mechanics can teach themselves rudimentary maintenance tasks. AI quickly became an interactive YouTube tutorial. The friction between having a question and getting a coherent answer collapsed almost overnight.

But the simple chatbot is now the floor. The most powerful uses of AI sit several levels above where most people are currently working, which means we are drifting back toward a real expertise gap. To build something really useful with Claude Design - or Cowork, or Codex, or any of the platforms sprouting up everywhere - you have to think like someone who knows what a digital artifact really is, when to deploy versus refine, how to prompt at a granular level of specificity, and when to ask for code rather than revision. Hardest of all: how to recognize when the AI is offering you a design choice instead of a final product. A mistaken entry has consequences - in token use, in time spent waiting for an output, and in the rate limits you'll hit faster than you expect.

You don’t have to write any code. You do have to think the way coders think, make the kinds of decisions they make, and understand the language they use. A new threshold needs to be crossed.

Models, Apps, and Harnesses

If you want to see the issue up close, read Ethan Mollick’s latest post on GPT-5.5. Mollick is one of the most accessible writers about AI and he works hard to translate for non-technical readers. But even his current framing includes assumptions about what general readers are tracking. If you haven’t kept up, his posts are liable to be overwhelming.

He encourages us to think about AI as three interlinked concepts: models, apps, and harnesses. Models are things like Opus 4.7 and GPT-5.5. Apps are the products that let models do real work - claude.ai, chatgpt.com, Claude Code, and Codex. Harnesses are the tools the AI uses and how it is hooked up to them. That last one is probably the deal breaker for most teachers. How many educators do you think are linking APIs to GitHub? Setting up MCP servers? Are you still with me? This is where the real power resides for serious AI users, and it is currently beyond the reach and expertise of most educators.

Things are moving so fast that even the most vigilant cannot keep up. That is becoming more true by the week. A humanities teacher reading the GPT-5.5 piece is not reading about whether students can cheat on an essay. Mollick describes generating a near-PhD-quality academic paper from four prompts using Codex and an interactive 3D simulation of an evolving town across 6,000 years from a single prompt. These are now ordinary outputs from someone who knows how to integrate models, apps, and harnesses into a repeatable workflow.

Even Mollick has acknowledged this gap. In a post from late March, writing about the most powerful tools currently available, he admitted:

These tools are terrific, but they are really built for programmers. They assume you know Python and Git. Their interfaces look like a 1980s computer lab. For the 99% of knowledge workers who are not developers, these powerful AI tools are not optimized for them.

Claude Dispatch and the Power of Interfaces

Tools like Claude Design are the industry's attempt to fix this. The capability is there but what is new is the scaffolding meant to let non-technical users reach it. The catch, based on my experience, is that the new scaffolding is itself technical. The interface is better, but the conceptual demands are still out of reach for most. Maybe that will change, but right now it’s an impediment.

Hitting A Wall

Last December I appeared on the leaderboard of a brand new platform called AI Cred4, and the founder’s follow-up post name checked me as one of the surprises - a 30-year history teacher with no coding background showing up alongside engineers and consultants. In February I wrote that Claude Cowork “just worked” for me, and I meant it. Building online tools took some figuring out, but I did it. I suspect I'll get there eventually with Claude Design, but I'm finding myself out of my depth more often than I once was.

I need to reread every sentence twice. I’m having flashbacks to my days fumbling through FORTRAN and Pascal in the 80s. The skills that got me on that leaderboard last fall haven’t gone away. The problem is everything that’s piled on top of them since. Keeping up with the models and new releases is practically a full-time job.

The tantalizing thing is that I do see the value. I understand why these tools are a breakthrough and eventually worth learning. But taking the time to master them is an opportunity cost I can’t currently spend. It all makes me nervous about what comes next.

Last week I asked Claude Design to revise a one-page professional summary I’ve used for years. The draft came back elegant and the formatting impeccable. I was inserting updates from the last few years and needed to know how best to export the file. It recommended a standalone HTML file - not PDF and not Word. Then it offered an unsolicited note: “Start a new chat to save 107k tokens of context. This uses your rate limits more effectively.”

I knew what “standalone HTML” meant because of the Cowork projects I've done. I also knew that saving “107k tokens of context” was a good thing. A teacher who recently upgraded to Claude Pro to try Claude Design might not have known what either phrase meant.

It is not that learning these capabilities is hard, exactly. You can ask the AI when you get stuck and it will generally walk you through it. What the companies are doing, however, is putting the newer models out of reach for all but the most dedicated users with the resources and time necessary to devote to it. It’s understandable why companies will invest in this. It’s not yet clear why schools should.

The Eternal Skills Question

The best argument against using AI in classrooms goes like this. Teach students the eternal skills - writing, judgment, critical thinking, the ability to refine ideas - and AI fluency will follow. Kids need to know how to think before ever touching a chatbot. Concrete elements of AI instruction, like prompting, can be learned later. I tend to agree with all of this in the abstract. The strongest version of the argument is that schools exist to build durable human capacities, not to chase whatever the workforce happens to need this year.

But the eternal skills argument assumes students will not be using AI during that time. That is impossible unless we lock them in a room without technology. Some parents would willingly make that tradeoff. Most schools cannot.

It also assumes the AI students encounter at twenty-three will resemble the AI they are using at fifteen. It won’t. Students graduating this May and June will matriculate to college and eventually workplaces where the entry-level fluency expectation is already past the chatbot tier. The AI of 2034 will be unrecognizable to them, and the judgment required to use it well is not something that can be snapped on at the end of college. AI is the first tool that talks back in any meaningful sense - it argues, flatters, redirects, and sometimes deceives. Nothing in our experience tells us how long it takes to develop fluency with something like that. Withholding AI from students until they’re “ready” assumes a static technology which currently doesn’t exist.

I'm not convinced the eternal skills will be enough on their own, or at least, they are going to have to be developed alongside the reality of student use. AI literacy is on track to become a baseline professional expectation and the cost of being late is likely to compound quickly.

Class of 2032

My middle-school daughter graduates high school in 2032. Even if you accept the argument that high school is not workplace training, not exposing her to AI and what it is capable of during her high school years feels like an abdication of our educational responsibilities. We know they're using it already - every conversation I've had with actual students confirms it. Teachers cannot use AI themselves while refusing to explain it to them. The question no one has been able to answer yet is how to do that well. Can we learn this with them rather than ahead of them? The challenge is how to teach a tool we’re still figuring out ourselves.

This is uncomfortable territory because it implies that the humanities teachers leaning into the AI conversation in schools - and I am one of them - need to learn things we did not sign up to learn. Not exactly how to code, but how to adopt the mindset and understand what computer code is. We need to recognize when a tool should be used as an artifact builder rather than a chatbot and, most importantly, when it should not be used at all. Even more daunting is we may need to adapt the vocabulary of the software engineer to the vocabulary of the classroom.

The Home Stretch

As we head into the home stretch of the school year, I have been reflecting on what has changed for me over the past twelve months. “Everyone is Cheating Their Way Through College” was published on May 7th, 2025. That is old news now, and a tired take, though I do not know exactly how much has actually changed in classrooms. There may yet be a wave of success stories about blue books and oral exams. My experience has been mixed.

The real story for me is how I have moved from a purely text-based use of AI to one predicated on the ability to generate code to do things. Yes, it is partly the “agentic layer” that was promised last year and is now taking root. But the more important point is that it is going to require an entirely new set of skills for non-coders to do the kinds of things that come naturally to coders.

When I spent a summer as a paralegal, one of the partners explained to me that every profession has its own secret vocabulary, its own language that sets it apart. Generative AI gave us, briefly, the illusion that we could keep swimming solely in the world of words we understood. The truth is we have been living in someone else's domain all along. We just did not have to see it. Humanities teachers may have to learn to read in a language we did not train for, if we intend to keep up.

Connect With Me

Beyond this newsletter, I work directly with schools, educators, and other institutions about pedagogical questions raised by AI. Take a look at my website and reach out - I’d love to hear what you’re working on.

It's late April and I'm running on fumes. Students and teachers everywhere can see the finish line, but it's still far enough away to be more distraction than relief. Overextended and buried under a to-do list a mile long, if ever there was a moment when a time-saving tool would be critical, this is it. And yet AI is not that tool - at least not in the way it’s frequently pitched.

The availability of powerful generative AI makes it very easy to do things that probably should not be done at all. A core skill in the coming years will be determining when, why, and for what purpose AI should be used. Because, make no mistake, it will continue to be used by millions of people, and many of those people will be students and teachers.

Six Weeks of Time Savings? Really?

If there is one message that unites nearly every report, conference keynote, and vendor claim about AI in education, it is that AI can save teachers time. The January Brookings report on AI and students summarized the consensus:

AI’s benefits extend to teachers: by reducing time on tasks, AI allows teachers to focus on individualized student attention and enhance curriculum and instruction.

NPR’s coverage of the report cited a U.S. study claiming teachers who use AI save an average of nearly six hours a week and about six weeks over a full school year. That number deserves scrutiny.

The "U.S. study" in question is a Walton Family Foundation/Gallup survey from June 2025 - predating the Brookings report, which cites it. The methodology was key -teachers who already used AI weekly were asked to estimate, to the nearest half-hour, how much time AI saved them on individual tasks. Those self-reported estimates were then summed. The result was 5.9 hours per week or, cumulatively, six weeks per year.

I don't buy it. The Brookings report itself hedges the claim in a section titled "Time savings only occur if teacher workloads remain steady" (Box 05 on p. 40 of the full report). More importantly, neither the Brookings report nor the Walton study can specify what is actually being done with the time saved. As Tom Daccord noted on his Substack:

Even if AI is saving teachers time, the benefits only accrue if that time goes back into individualized attention and deeper instruction. AI time savings might simply mean that teachers pack up and go home earlier. It might mean the time is absorbed into more administrative and ancillary tasks. The difference matters because if one of the major selling points for teachers using AI is to “save time,” then it is imperative that the time saved is used wisely. But I question the entire narrative that AI is saving teachers time in the first place.

The Walton/Gallup survey's own data complicates the six-week headline. Look at the frequency chart. Eighty percent of teachers have never used AI for one-on-one tutoring. Fifty-nine percent have never used it to make assessments. Seventy-three percent have never used it to supplement instruction. The "six weeks" figure comes from frequent users - a group that represents, at most, a third of the teaching population. Which means, for most teachers, AI time savings remain entirely theoretical. It’s also worth asking - if the time savings is so dramatic, why aren’t more teachers using it?1

The majority of tasks for which teachers are using AI are either tasks they were doing already or, in the case of something like modifying materials to meet student needs, ones which AI enables in a way that wasn’t feasible before.

This exposes a subtler problem with how time savings get measured. Consider a teacher who uses AI to build an elaborate differentiated worksheet in 20 minutes. She estimates that creating it on her own would have taken an hour and reports 40 minutes saved. But without AI, that worksheet never gets made. The actual time impact is plus 20 minutes, not minus 40. In that instance, the survey measures savings against a hypothetical task that would never have been undertaken. Multiply that logic across nine task categories and you get 5.9 hours of savings that potentially exists only on paper. The worksheet may well be superior. That's a quality argument, and a potentially legitimate one. But it is not a time-savings argument.

What the Workplace Data Actually Shows

Regardless of the inconsistency of self-reports, there is data available that undercuts the notion that AI time savings in the workplace translate into actual efficiency gains at all. Recent studies measuring behavior rather than perception tell a different story.

The ActivTrak 2026 State of the Workplace report, released in March 2026, analyzed 443 million hours of behavioral data across 1,111 organizations and more than 163,000 employees. Their conclusion was blunt: “The data is unambiguous: AI does not reduce workloads.” Among users comparing 180 days before and after AI adoption, time spent across every measured work category increased - ranging from 27% to 346%. No category decreased. Meanwhile, the average focused, uninterrupted work session fell to just over 13 minutes, down 9%, and continuing a three-year downward trend. The optimal band of AI usage turned out to be extraordinarily narrow - employees spending 7-10% of their work hours in AI tools showed the highest productivity of any group, yet only 3% of workers fell within that range.

A separate BCG study of nearly 1,500 full-time U.S. workers, published in the Harvard Business Review, also in March 2026, identified a phenomenon they called “AI brain fry.” The primary driver wasn’t AI use itself but the cognitive burden of overseeing its output. Workers with high AI oversight loads reported 14% more mental effort, 12% greater mental fatigue, and 19% more information overload. Those experiencing the condition made 39% more major errors and were significantly more likely to consider leaving their jobs.

These are not education-specific studies. But I’m not sure why we would assume the findings would be significantly different for teachers. Evaluating what AI produces is mentally demanding work, and the overall pattern across industries is consistent: AI adds new layers of work rather than reducing existing ones.

The Novice Teacher Problem

Imagine you are new to teaching. You’ve completed your training and you are organizing your course materials for the first time. What lesson should you start with? How should you set up your classroom? An AI chatbot will happily provide answers to all of these questions and some might be generically helpful. But which ones are right for your students, in your room, given your strengths? You have no way to know.

The ability to navigate the hundreds of decisions that go into teaching comes with experience - the kind that requires following your own instincts, getting real-time feedback from actual students, and adjusting the next day to adapt to your circumstances.

LLMs are great at generating ideas. Many are creative and potentially useful. An experienced teacher can scan them quickly and locate the one or two that might lead to a productive tweak. Newer teachers don’t have that luxury. They haven’t made the hundreds of mistakes that allow them to grab the one useful suggestion among dozens.

This is a judgment problem, and chatbots cannot impart judgment.

The “brain fry” finding is particularly relevant here. The primary driver of cognitive strain wasn’t using AI itself but overseeing the output. And that presupposes a major assumption – that teachers are actually carefully reviewing what AI produces. As models become more advanced and more verbose, the cumulative effect of reading and evaluating AI-generated text can easily take longer than starting from scratch.

Veteran Teachers Have A Different Problem

Even for experienced teachers, the time-savings pitch doesn’t hold up, though for different reasons.

Over the past year, I’ve written about several ways I use AI in my own practice - building a Claude Skill to audit lessons using the Understanding by Design framework, and creating interactive web modules for specific activities in my classes. Each of these was a significant investment that I believe paid off (or will pay off) in better student outcomes. But none of them saved time.

The lesson audit produced useful, actionable feedback. It identified misalignment between what I wanted students to learn and what my activities actually assessed. But evaluating every suggestion required 30 years of teaching experience. It was helpful but not “efficient” – at least not in the way the Brookings report envisions “time savings for teachers.” Performing an AI audit for every lesson during the school year is simply not realistic. That’s useful summer work.

The interactive modules are, I think, genuinely effective teaching tools. They are also time-consuming to build and require significant design thinking before AI can execute anything usable. My abiding mantra when creating or redesigning a lesson has been: what is the value add, and is it worth doing? AI hasn't changed that question.

Even for teachers who aren’t building custom tools, AI is still unlikely to save them time. Generating a quiz is easy. Proofreading questions, verifying outputs, and ensuring the product meets your classroom standards is not. Creating new worksheets from AI-generated ideas may be rewarding and great for your students. But the measure of “time savings” is always against the alternative – using the original lesson. If AI leads you to do something different and do it well, it’s a worthy goal, ideally one that leads to better teaching and learning. It’s not a shortcut.

Depth, Not Speed

The argument I am making is not that AI cannot help teachers teach better. Used effectively, AI is exceptionally good at helping teachers improve their own work. The newer models and capabilities offer possibilities for teachers with the curiosity and - most critically - the time to learn about and play with them.

All of the best use cases require significant up-front investment. Using AI effectively may also increasingly require access to the top-tier frontier models - an entirely different issue.2 These two points cannot be sugarcoated. AI opens up new possibilities precisely because it enables work that wasn’t possible before - not because it makes existing work faster.

The ActivTrak data confirmed what I’ve found in practice. AI doesn’t replace what we’re already doing. It amplifies it. But the claim that it saves time the way a dishwasher saves time or a calculator saves time is not what I've experienced.

Which brings me back to Drucker. AI makes it efficient to do things that shouldn’t be done at all - churning out generic lesson plans no teacher has examined, producing feedback no student will read, and generating worksheets that exist because they’re easy to create rather than because they serve a clear learning goal. If that’s the primary way in which teachers are using AI, that’s the Drucker problem in a nutshell. Efficiency without judgment is worse than inefficiency.

Time is the most important resource a teacher has. Be wary of anyone who tells you AI will give you more of it. What AI can do - for teachers willing to go deeper - is open up choices that didn't exist before. Those choices mean more decisions, and every one costs time and attention.

I’d love to hear from other teachers about their experience with AI as a time-saving tool. Am I wrong? What am I missing? Let me know your thoughts in the comments.

Connect With Me

Beyond this newsletter, I work directly with schools, educators, and other institutions about pedagogical questions raised by AI. Take a look at my website and reach out - I’d love to hear what you’re working on.

Schools should screen the new AI Doc - teachers, students, and parents should all watch this film. Preferably together. The reason is not because the documentary is especially illuminating on education. It isn’t. The current controversial and pressing questions about AI and schools are not the focus. This movie is not about AI and writing or cognitive offloading.

The reason the film should be viewed is to provide common vocabulary for us to have conversations about a technology predicted to be the most transformative in history. More importantly, it serves as an excellent primer for viewers unfamiliar with the existential debates surrounding AI acceleration and gives them exposure to both the apocalyptic and messianic positions those closest to the technology hold. Citizens, especially students, cannot challenge positions they don't know exist.

I recently saw The AI Doc: Or How I Became An Apocaloptimist with a colleague. The documentary arrives at a critical juncture. I recommend the documentary to everyone because it details the range of extreme positions taken by AI’s creators, risk assessors, and others in the field who are setting the framework for the debates.

What the Film Does

The basic conceit of the film is that co-director Daniel Roher1, an artist and Oscar-winning documentary filmmaker who serves as a stand-in for the audience, and his wife Caroline (also a filmmaker), are expecting their first child. Both are smart and likable - but Daniel's anxiety grows the more and more he learns about AI. He decides to investigate and make a documentary.

Some of the most recognizable names in tech agreed to sit and be interviewed in his makeshift studio, including Sam Altman, Demis Hassabis, and Dario Amodei, the CEO of OpenAI, CEO and Co-Founder of Google DeepMind, and CEO of Anthropic, respectively. The reaction shots of Daniel as he listens to one extraordinary claim after another are priceless - first the doomers predicting catastrophic outcomes due to unregulated AI development followed by a parade of AI optimists who tell Daniel this is absolutely the best time ever to be born. The doomers talk about AI leading to the collapse of humanity (or, as Eliezer Yudkowsky politely interjects, “abrupt extermination”) while the optimists describe a utopia where no one has to work, cancer is cured, and AI ushers in a world of abundance such as we’ve never seen.

For viewers who are hearing tech folks speak this way for the first time, Daniel is an everyman who articulates what the audience is thinking. Why are we doing this if these horrific outcomes are possible? How much and who am I supposed to believe about any of this? He falls into despair. Mid-film, there is an amusing and creative discussion with his wife via a cartoon flip-page notebook in which she begs him to become more hopeful as she carries their first child to term.

To watch the film is to accept the premise that AI is going to be a very, very, big deal.

But the hyperbolic venting of all these people is the point. In the back half of the movie, Roher comes to the quite reasonable conclusion that most of the issues associated with these fantastical prognostications are coming from the founders and developers themselves.

Altman comes off the strangest here in my opinion, partially due to the incredibly long pauses and cryptic answers he gives to some of the most direct questions. When asked if there are protocols in place for what happens if one of his models goes rogue and starts to do tremendous damage he says yes, but fails to elaborate on what they are. The fact that he is more than aware and clearly planning for the possibility is disturbing enough.

Why It Matters for Schools

Students are understandably tired of only hearing about AI in the context of cheating. Teachers are tired of harping on it. Education will be dealing with the fallout of AI for years, but the conversation is going to expand and only get more divisive and confusing as the technology continues to improve. I’ve argued previously that schools need to engage the larger issues about AI and The AI:Doc offers precisely that entry point. I can easily envision assemblies, advisory discussions, and guest speakers brought in to elaborate and answer questions about issues which, while they are going to affect everyone, will have a disproportionate impact on students enrolled right now. The purpose is not to accept the views of the film’s participants as gospel but to serve as a jumping-off point: for additional readings, other interviews, and opportunities to hear missing pieces and critical commentary.

Students will be rattled by the film. Public viewings will spur useful conversations that rise above the simple context of cheating and cognitive offloading that currently swamps the debate about AI and education. If this documentary is any indication, those issues are going to quickly become supplanted by much larger ones and I think everyone, but especially the younger generation, should gain access to the different sides of the debates.

These are the actual conversations a very small number of people are having when it comes to AI. What The AI Doc does is put forward a series of arguments and positions in front of the audience who may not be inclined to realize exactly how much and why Silicon Valley is focused on AI to the exclusion of almost everything else. There are statements made in the film that give the sense that the people who are closest to AI development lie awake at night contemplating the end of humanity. Others insist we're standing at the threshold of unprecedented human flourishing. Those are radically different perspectives.

There is an amusing sequence as Roher does a series of Google searches investigating how to survive off the grid if things go kaput, basically modeling how and why some people are turning into “preppers” for the world’s collapse.

The primary argument of the documentary is what might happen if we end up getting AI wrong, even if it approaches only a fraction of what’s being predicted, and how the general public should respond. The issues are significant enough that politicians, policy makers, educators, and anyone with an interest in the future for their children, need to be aware of what the people building these models think is at stake.

Yes, the Middle Position is Missing

Critics of the film will rightly point out that the missing middle conversation is left out. Yann LeCun and Gary Marcus, both well-known skeptics of imminent AGI, are not featured (Emily Bender gets a single quote).2 There is no exploration of “AI as Normal Technology.” Exponential AI scaling is taken as a given and the assumed timelines for AGI by most of the speakers are on the aggressive end of the spectrum. But I actually think that’s a benefit. It will be instructive to look back in 5 years to see whether this film was prescient or overblown.

One participant has invoked similarities to The Day After and I would also liken it to An Inconvenient Truth. The AI:Doc is a wake-up call. The people with the most direct knowledge of where AI is headed are genuinely alarmed, and quite a few are begging for government regulation - that includes the ones building it, like Amodei.

The sooner more people learn about these predictions, investigate these claims, and learn for themselves what’s likely to be true and what isn’t, the better. If screening the AI:Doc in schools has that effect, it can’t come soon enough.

Connect With Me

Beyond this newsletter, I work directly with schools, educators, and other institutions about pedagogical questions raised by AI. Take a look at my website and reach out - I’d love to hear what you’re working on.

One of the best things about being a teacher is that work literally stops over break. Unlike other jobs, nobody needs someone to cover for you back at the office. I spent the better part of the past week away from a screen and it was exactly what I needed. It also showed me something I’ve known intuitively but hadn’t experienced directly - normal people are not following the AI news cycle with any kind of regularity. I am not normal.

So when I came back online this weekend, one story caught my eye. As of this posting, the story about Claude Mythos is a little more than 72 hours old. If you haven’t heard about it, join the club. While it will likely be lost in the deluge of the AI news cycle, it foreshadows an issue that will only become more relevant in the coming months - the staggering cost of running the most powerful AI models and who gets access to them.

How I Found the Story

My substack feed led me to Mythos through Carlo Iacono’s most recent post. It sparked my curiosity and gave me an opportunity to engage in the same kind of source analysis I model for students. How do you verify a story with so many moving parts and so little actual information to go on?

To find out, I turned to the SIFT framework, developed by digital literacy researcher Mike Caulfield.1 It stands for Stop, Investigate the source, Find better coverage, and Trace claims to their origin. Caulfield and Stanford researcher Sam Wineburg expanded on this approach in their book Verified.

What SIFT Looks Like in Practice

Here’s what I found when I traced the story back.

The original reporting came from Fortune’s Beatrice Nolan. Every hot take and every inflated claim on X flows downstream from those articles.2 What I would tell my students, however, is that even Fortune's reporting was based on documents independently discovered by two cybersecurity researchers - Roy Paz of LayerX Security and Alexandre Pauwels of the University of Cambridge - in a publicly searchable data store that Anthropic had failed to secure. Those draft blog posts and internal assets were the real primary source material. Anthropic locked down the data store after Fortune's inquiry, though by then cached copies had already begun circulating.3

Fortune’s reporting establishes the following: Anthropic has trained a new model currently called “Claude Mythos” (also referred to as Capybara), which is the next tier above its current flagship, Opus. The company confirmed it represents “a step change” in capabilities with “meaningful advances” in reasoning, coding, and cybersecurity.

The leaked draft is slightly more assertive than the public statement - it describes the model as “far ahead of any other AI model in cyber capabilities” and warns of “an upcoming wave of models that can exploit vulnerabilities in ways that far outpace the efforts of defenders.” Because of this risk, the planned rollout gives organizations early access to protect their systems before wider release.

The leak itself was caused by a content management system that made uploaded files public by default. Anthropic called it “human error” and said it was unrelated to any of its AI tools. It’s worth noting the irony that the company that has been warning about AI-powered cyberattacks lost control of its own documents through the kind of vulnerability its model is designed to detect.4

The article is based on internal draft “blog posts” which are essentially marketing copy written for a product launch. These are not internal memos or unfiltered emails. We're getting what Anthropic intended to say publicly, just before they were ready to say it. That's quite different from actually reading what they say privately and led to some commenters claiming the “leak” was intentional and simply a way to gin up hype. It wouldn’t be the first time a company engaged in such tactics, though Anthropic does not have that reputation.

That’s the gist of the verified story. Lots of people took this and ran with it.

Separating the Signal From the Noise

The Mythos story is a perfect example of a pattern all students and readers should know well by now. One verified story often generates a dozen unverified extrapolations. Within hours, new and unconfirmed details - ten trillion parameters, ten billion dollars to train - were circulating alongside claims that came directly from Fortune's reporting. One widely shared post mixed the two so seamlessly that a casual reader couldn't tell which claims were verified and which were invented.

That’s how it works - inferences, embellishments, and flat-out fabrications get shared more widely than the original reporting and errors get lost in the shuffle. Social media turbocharges the phenomenon and by the time most people encounter a story like this, the verified facts are indistinguishable from the noise.

The number one question students should ask when confronted with a claim that requires confirmation: can I trace it to a named source with direct access to the evidence? For the Mythos leak, the answer separates a surprisingly short list of confirmed facts from a much longer list of speculation.

One confirmed detail that matters most for educators was buried underneath the cybersecurity headlines. The model will be “very expensive for us to serve, and will be very expensive for our customers to use.”5

The Tiering Problem

Most students and teachers, if they are using AI at all, are using free tiers. Few people know which specific models power these free tiers, let alone how they differ from the most advanced versions. The default models on the free tiers of Claude, ChatGPT, and Gemini are all more powerful and multi-modal than what was available one year ago, and their performance is driving almost all of the AI and education conversation.

As the past few years have made clear, these base models are more than capable of disrupting the entire educational experience.

Yet these are not the models that make the boldest headlines. Each company now offers premium tiers - Claude Opus, ChatGPT Pro, Gemini Ultra - that are significantly more powerful and reserved for companies and individuals willing to pay top dollar for access.6 Meanwhile, the free tiers are tightening. Usage limits are getting stricter, resource-intensive features are being pulled back, and even paid subscribers are seeing throttled access during peak hours. The gap isn't just opening from the top - it's also closing from the bottom.7

What the Mythos story makes clear is that these frontier models are going to keep pulling away.

Opus, Anthropic’s current top-tier model, is already primarily an enterprise and power-user product. Mythos would add a tier above that and it’s explicitly described as too expensive for general release. The intense market competition among the three major AI developers - OpenAI, Anthropic, and Google - practically requires that the next model from each keep pace.

As AI models become even more powerful, they will be available first, and for an unknown period, only to organizations that can pay enterprise prices. Those organizations will not be public school districts. They'll be the European executives invited to preview unreleased Claude capabilities at an intimate retreat at an English country manor - a detail which was also in the leaked cache.

I wrote about this dynamic in February when the Moltbook story broke:

Mythos is Anthropic’s clear confirmation that’s where we are headed. For a while it seemed as if everyone was going to have similar access to the most advanced AI models available. By the end of 2026, I don’t think that’s going to be the case any longer.

The gap between what AI can actually do and what the average teacher or student has access to is about to get wider.

When we talk about AI in education, we tend to talk about “AI” as if it were one thing. As if the tool a student uses on the free tier of ChatGPT is the same as the one a top cybersecurity firm is using to audit its codebase. We talk interchangeably about the experience of a teacher using Gemini for the first time and the one a software developer has with Claude Cowork’s agentic process.

These are entirely different interactions with AI and it's one of the most overlooked dynamics in the education conversation. The tiering gap doesn't change what most teachers are dealing with right now - students can already do tremendous damage to their own learning with the free versions of these tools. We know this.

But when someone goes online and starts reading about what AI can and cannot do - as if each model is identical - it's clearer than ever why those conversations go sideways so often. Much of the commentary is based on premium models most users never experience, and the companies aren't exactly helping - OpenAI has quietly stripped model names from its interface, making the gap between tiers all but invisible.

The Same AI Experience? I Doubt It

It’s obviously too soon to tell whether Claude Mythos will be as consequential as Anthropic’s leaked promotional materials suggest. If the past is any indication, it will likely be another incremental improvement and garner attention mostly from within a small sphere of AI insiders. Anthropic has gotten more press recently, but it’s still a niche company. Most people you know have never heard of Mythos and probably never will. But if you’re someone who follows AI closely, it’s worth getting the facts straight. The newest and most powerful model priced at a level the average user cannot access means, if nothing else, that we are not all going to be having the same AI experience.

Connect With Me

Beyond this newsletter, I work directly with schools, educators, and other institutions about pedagogical questions raised by AI. Take a look at my website and reach out - I’d love to hear what you’re working on.

Back in December I sat down with a handful of HS students on a couple different occasions to talk about AI - their thoughts, attitudes, and descriptions of how students were using it in school. There were lots of observations - most unsurprising - but an overall confirmation that student AI usage had become a mainstream part of their academic lives. With the 2025 - 2026 school year more than half over and as we headed into a much needed spring break, I had an opportunity to sit down with about a dozen HS students in our AI club - 9th, 10th, and 11th graders - to revisit the conversation about AI and school.

One asked about the recent issue between Anthropic and the Pentagon.1 Another was willing to share how he uses AI as a study aid. And several talked about how AI was impacting the corporate world. On the whole, the recent conversation made something even clearer - the AI gap between teachers and students continues to grow despite more and more schools trying to catch up.

I know this is one group of kids at one school - members of an AI club, not exactly a representative sample, though one that notably included several girls whose voices have often been absent from this conversation.2 But I suspect the dynamic is recognizable to teachers well beyond our building.

These kids were thoughtful, specific, and honest. They had ideas, critiques, and an underlying message - please help us understand how AI fits or doesn’t fit into school because we are using it ALL. THE. TIME.

How They’re Actually Using It

I opened the meeting with a simple question: how are you and your peers using AI?

One student shared some screenshots with me of several ways he uses AI as a learning tool. Using Claude’s presentation feature, I stitched them together into a short deck, which he walked us through. He showed us interactive math diagrams with adjustable parameters, AI-generated quizzes with built-in explanations, flashcard sets with shuffle modes, and gamified study tools - all generated from very simple prompts.

Most educators (me included) wouldn’t characterize this as AI misuse. This student has figured out that AI can function as a personalized tutor, study partner, and test designer. When I first started writing about AI a little more than a year ago, few if any of these features were available. The AI issue has rapidly moved beyond mere text generation.

Another student described using AI for product research and hobby work:

“It’s really good at taking two products and comparing and contrasting them in a way that Google can’t. If I give it two things I’m looking at, it’ll give me pros and cons.”

When I asked whether he trusted the output, he didn’t hesitate:

“For any research I’m doing, I ask it to give me the main points of what I’m researching with citations, and then I go through those citations to see what’s actually critical.”

That was encouraging though I suspect it’s not the norm.

The “Neanderthal” Problem

One student offered an opinion with typical teenage bluntness:

I think for a lot of people my age, their understanding of AI is pretty Neanderthal. I think right now a lot of people just see it as a shortcut versus something you integrate within your daily life to improve your efficiency. There’s just so many more uses that people overlook - they see it as a very two-dimensional thing.

This obviously complicates the narrative that students are simply using AI to cheat. These students - granted, again, these are kids in an AI club - were drawing a clear line between those who use AI as a shortcut and those who use it as a tool - and they were frustrated that teachers, and even their own peers, couldn’t always see the difference.

Another student connected this to what awaits them in the future:

I think a lot of people skip over the fact that AI is being used way more in the workforce. I’ve heard a lot of people who work in technology and business - they’re starting to use AI models because it’s just a way more efficient way to get stuff done. I think that’s going to be really interesting to see the landscape of the next twenty years.

I told her I wasn’t sure it would take twenty. More like five.

What They Want Teachers to Know

I tried to shift the discussion from description to prescription. I asked the group directly: if you were a teacher right now, how would you handle AI in your classes?

One student referenced something one teacher had done - create a custom chatbot designed around specific assignment guidelines that students could use within defined parameters:

One thing I really liked is [when a teacher] made ... a chatbot thing that was specifically designed to follow the guidelines [they] set up. We could go into it, and it was limited. I think that's a good way to do it, because telling kids just the guidelines - like, [another teacher's] guideline for AI is just none, zero - I think that's very hard for students, because then it makes every single thing they do with AI feel very bad. And they’re still going to do it.” [emphasis added].

This is probably least surprising but also most difficult to swallow. Students will use AI whether we forbid it or not. That’s just the reality.

When I asked her to confirm - students are using AI regardless of the policy? - she didn’t hesitate:

“Yeah, definitely … I think teachers should be aware that kids are going to use AI. And I think I would want to create my own space that has guidelines and make sure all the students are in there, under my supervision, learning how to use it effectively. Because I think learning how to use it effectively is better than trying to make kids not use it.”

Another student made the case against prohibition even more directly:

“I think criminalizing AI, especially in the classroom, does way more harm than allowing it in moderation. If AI is just outright banned, it gains this association of being negative, like it’s cheating, like it’s bad. Whereas if students are taught to use AI in moderation - because they’re going to use it, they’re going to use it regardless - this way it teaches them to be conscious and aware of their AI usage and learn how to use it effectively.”

One more offered a practical recommendation:

“If I was a teacher, in any other class like English or creative arts, I would actually teach how to use AI to make a study guide for large blocks of information. It’s beneficial for memorization and studying key terms. But if it’s an in-class writing assignment, a skills check - you shouldn’t use AI just to get the work done faster. If it’s studying, it’s one thing. If it’s writing whole sentences for you, whole paragraphs, that’s another thing.”

These kids are proposing precisely what many commentators are advocating - context-dependent policies that distinguish between AI as a study aid versus AI as a substitute for thinking.

Pushing It Underground

A repeated observation popped up about the difference between AIs with guardrails and typical models found online. This is the problem Marc Watkins tackled in a recent post - what we might call the “free version” problem. A student who had experience with our AI wrapper remarked:

“I used Flint on my own time, and I have to say, compared to other models, it was incredibly frustrating because it’s designed the way teachers want you to use it. It wouldn’t give me direct answers - it would be like, ‘What do you think?’ It wasn’t that helpful in my opinion.”

Which is exactly what Flint and similar paid AI wrappers are supposed to do - push the student to think rather than hand them answers. Institutions are pouring money into these products. But the notion that kids will default to AIs deliberately designed to create “friction” when ample free versions exist that do not is more than unrealistic. It’s fantasy. When I asked how many had used or were even aware of the “study mode” drop-down selections baked into all the frontier models, they just stared at me.

Another student was even more direct about what happens when schools take a punitive approach:

“I think a big issue with how schools are dealing with AI right now is that so many kids are getting in trouble for it, and I don’t think it necessarily discourages the use of it. If anything, it just pushes it underground even more. I think making it into a learning movement - learning from mistakes and understanding the line of what’s okay and what’s not - rather than just punishing would be more effective.”

Obviously, these are teenagers. They didn’t catch the contradictions (introduce guardrails but we don’t want guardrails!) nor do they see the larger issues through the same lens as experienced teachers, almost all of whom learned their academic skills in a world where generative AI didn’t exist. These kids don’t know that world. They are trying to adapt in real time while many teachers are still assigning the same kinds of work that can now be done faster and better by AI than by high school students. That’s not their fault.

How Will We Close the Gap?

When students consistently said that teachers should be teaching them how to use AI effectively, I shared something I’ve observed in my conversations with educators across the country: many teachers don’t know how to use AI effectively. Quite a few are openly critical, many more are ambivalent at best, and those few who do feel comfortable enough to teach with it often find themselves navigating colleagues and administrators who are dealing with a hundred other priorities.

The result is that student use of AI is probably more experimental and varied than what most teachers are doing. Only a small fraction are particularly fluent with it. Many have serious reservations - and I understand why.

One student responded:

“I think a lot of teachers see AI as a bad thing because they don’t really know how to use it themselves. They only think of students plugging in a prompt and it just spitting out an answer. But I think if teachers develop this AI literacy - I think AI might see a lot more healthy use in classrooms than it’s seeing right now.”

He was getting at something I’ve written about before - the idea that teacher engagement with AI, even imperfect engagement, might shift student behavior more than any policy could:

“I think by teachers showing students how to navigate the ropes of AI, it creates an environment where students might respect the boundaries a bit more, yet still use AI in a way that’s beneficial to them but not overly harmful.”

I don’t know if he’s right about that, and I suspect many teachers would disagree. When I asked how many of their teachers had ever designed an AI-integrated lesson or even demonstrated what they consider appropriate AI use, the answer was almost none.

What I Showed Them

Towards the end of our club meeting, I shared some of the things I’ve been building (which I documented in my previous post) - an interactive pre-Civil War module to support a discussion-based activity, a Constitutional Convention simulation where students adopt delegate personas, and a data visualization of American war casualties over the centuries. I also showed them Claude’s newest visualization features and a history simulator built by Ben Breen, a college professor, to give them a sense of what other educators were experimenting with.

Nearly everything I build with AI stays within my subject area, and I frequently use my own materials - primary sources, discussion questions, lesson content - to help create additional documents and assignments. But the best way to understand the limits of AI is to test it on something you don’t know well.

When I suggested on the spot that I demo something in STEM (not my area of expertise!), a student suggested an interactive graphic of Bohr’s model.

I was honest:

“When I use AI, I tend to use it in subject matters where I know a lot, so I can get a sense of whether it’s way off. And for the most part, it’s really pretty good. So I’m going to extrapolate that this is still decent. But I don’t know - if I showed this to a physics teacher, he might be like, ‘Oh, this is totally wrong,’ and I would have no idea.”

The point wasn't to impress them. It was to model what using AI as a design tool looks like - and to be transparent about its limitations. I can build rich, interactive history modules because I provide the content and generally know when something is wrong. Expertise has to come first. AI is most powerful when the human in the loop actually knows what they're doing.

What’s Changed Since December

In my December conversations, students described, in part, a world of clandestine workarounds - split screens, retyping AI output by hand, and switching platforms when one gets banned. The vibe was deception and an acknowledgment of a cat and mouse game. How can I use it without getting caught?

This conversation was different. Not because these students aren’t using AI to get through some of their schoolwork. I’m not so naive as to think none of them do. But in the discussion on the day before break, they were willing to go further than that. They had more opinions about what responsible AI use looks like and specific proposals for what teachers should do differently, probably because I asked. They had critiques of existing tools and policies that were articulate and nuanced.

The fundamental dynamic still hasn’t changed with students still navigating AI largely on their own and minimal guidance from teachers. The gap between what students know about AI and what most educators can offer is also real.

One student observation from the December conversation still holds:

“A lot of adults have this idea of AI and all these things about AI, and they think they’re so right about it. So I think they should be more open to changing their views on how another demographic - like students - may use AI, and avoid the stigma around it.”

The Underlying Ask

What struck me most about this conversation wasn’t any single quote. It was the consistency of the comments. Across different students, different grade levels, and different levels of AI sophistication, they were all saying versions of the same thing: they want help understanding this new technology and how to use it for learning. Telling us not to use it and just punishing us when we do is not a long-term solution. Give us a space where we can be honest about what we’re doing without fear that honesty becomes a way to indict us.

That’s not an unfair ask.

Whether most schools can deliver on that right now - with the training gaps, philosophical objections, institutional inertia, and the sheer pace of technological change - is anyone’s guess, but I know more schools and more teachers are trying. Students will speak up if given the chance, and their voices still need to be heard.

The question is how many of us are listening.

Connect With Me

Beyond this newsletter, I work directly with schools, educators, and other institutions about pedagogical questions raised by AI. Take a look at my website and reach out - I’d love to hear what you’re working on.

One of my most cherished accomplishments in high school was constructing a small wooden box in what was commonly referred to as “shop” in the 1980s. A graduation requirement, the class was held in what today might be considered a maker space but at the time contained a table saw, hammers, and other carpentry equipment rather than 3-D printers. The sole assignment for the semester was to make something - in my case, as someone not especially inclined towards building things, I chose perhaps the simplest project imaginable - a box.

I had to measure the wood, cut the pieces, align the sides, affix the hinges, and sand and polish it until it gleamed. Aside from a research paper I wrote on Andrew Jackson, that box was one of the few assignments I still clearly remember. Unlike the paper, that physical box traveled with me from high school to college, two summers spent living in Japan, law school, and it even made it to my first real apartment, finally getting lost between moves sometime in the early 2000’s. I had that box for almost 20 years, storing change, receipts, and other miscellaneous detritus from daily life, and every time I looked at it I took pride in one of the few things I actually made with my own hands.

I’ve been thinking about that box recently. Not because of any sudden nostalgia for high school, but because I’ve started building things again - this time with AI. And the experience has reminded me that the act of making something, of designing and constructing something that didn’t exist before you willed it into being, engages a part of your brain that merely asking for information does not.

From Prompting to Building

In a recent post, I argued that educators should approach the new capabilities of agentic AI with imagination as well as caution. That piece was about what these tools can do. This one is about what kind of thinking they actually require - and why that thinking might reassure educators rather than alarm them. What I didn’t fully appreciate at the time was how much that thinking would resemble the work teachers already do every day.

The shift I’ve felt over the past several months isn’t entirely new, but one of degree. Agentic tools - AI that can build entire applications and websites rather than just generate text - have been available since at least the middle of 2025, but since the new year they've become significantly easier for non-coders to use. What’s changed for me is the focus: from generating content to designing experiences.

When I first started using AI seriously, I was doing what most teachers do - asking questions, iterating on responses, but primarily generating text - converting ideas into polished plans and getting feedback and analysis on my own work. AI remains useful for all of those tasks. But agentic AI introduces something fundamentally different. I can now craft an interactive learning experience - deciding what it should do, how it should be organized, who would use it, and what problem it needs to solve. That’s a much different - and more involved - cognitive activity.

Design Thinking

About a decade ago, our school brought in a consultant from IDEO - the design firm that essentially pioneered what’s now called “design thinking” - for a professional development workshop. I remember finding the session interesting: its emphasis on starting with user needs, prototyping repeatedly, and treating failure as information to act on rather than giving up in defeat. Like most PD workshops, it was engaging in the moment and then I went back to doing roughly the same things I’d always done. Every teacher knows this feeling.

But the notion of “design thinking” stuck with me and now seems more relevant than ever. A core insight - that how something is built and presented to its users matters as much as what it contains - turns out to be a helpful way to approach AI now that teachers have the tools at their fingertips that allow them to act on it.1

Design thinking is not an alien concept for teachers. Every unit you’ve ever planned, every set of learning objectives you’ve ever sequenced, every decision you’ve ever made about what to include and what to cut are design decisions. We’ve been doing this work our entire careers. But for most of us, unless you happen to have a degree in computer science, using technology in the classroom has been a mostly passive experience. The products and platforms we used were designed by someone else.

We worked inside whatever template Blackbaud or Canvas or Schoology provided and made the best of it. During the pandemic, many of us became more adept at leveraging our LMS through embedded videos, links to platforms like EdPuzzle or Desmos, all in service of organizing our materials online for student access. But we were constrained by the limitations of our school’s chosen product.

The difference now is that teachers can create their own bespoke LMS, limited only by their imagination. That is new and genuinely exciting.

What Design Might Look Like

For years, I’ve run a Constitutional Convention simulation in my U.S. history class. It’s always been a favorite - students are assigned state delegates in attendance at the 1787 Philadelphia gathering, debate the structure of the new government, and vote on the major compromises that shaped the Constitution. I've run this activity for more than 20 years and refined it considerably.

The materials I provide are detailed but sprawling. Students each have a delegate position sheet, primary and secondary source material, and a variety of other excerpts from digital and analog resources. Managing the various documents can be overwhelming, even when students get them ahead of time. This year’s convention went smoothly, but as any teacher knows, there is always room for improvement.

Curious about what was possible, and off the success of several earlier builds using Claude Cowork, I set out to create a single, self-contained online module that would house everything for next year.

Using a dedicated Claude Project, I worked through how to organize and display the material. I settled on a series of easily navigable tabs, each containing exactly what students needed at that moment. That decision required knowing how the activity actually flows in a classroom, which information students reach for and when, and where the bottlenecks have historically been.

I had to decide what belonged inside each section. The Delegate tab needed more than biographical information - it needed each delegate’s positions on every voting issue, organized so students could quickly find where their representative stood. But I also realized that understanding the Convention requires seeing how delegates diverged on the same questions - disagreements about government structure, determining representation, and the limits of executive power. During the design conversation, I described that specific pedagogical need, and Claude suggested a comparison feature where students could pull up two or three delegates side by side.

The comparison feature only existed because I articulated a teaching problem. That pattern repeated at every stage. Whether primary source documents should live inside the module or be linked out (when students click external links, the risk of distraction soars - if the source isn’t right there, it doesn’t get read). Whether my voting spreadsheet could be replaced with a live tracker built into the module itself (it could, and could be downloaded at the conclusion of each day’s session).

None of these decisions required coding knowledge. All of them required teaching knowledge. AI handled the execution while I managed the design. The result is a functioning interactive web application, and early feedback on the modules I’ve used this year suggests it will be an upgrade.

Two Kinds of AI Use

The design dimension of agentic AI changes the conversation in ways that haven't been fully appreciated. To see why, you have to distinguish between two fundamentally different kinds of AI use that are often treated as if they’re the same thing.

The way most students approach AI is one-shot or, at best, slightly iterative. They type a question to get an answer, ask for an essay draft, or request feedback on a problem. This is the cognitive offloading conversation we’ve been having for more than three years. The emerging research confirms what most teachers already suspect - when students use AI as a question-and-answer machine, the offloading is real and any learning gains, if they happen at all, are ephemeral.2 Students generate a product without doing the intellectual work that product is supposed to represent.

But designing with agentic AI is a fundamentally different activity. When a user sets out to build something - an interactive module, a web application, or some other tool to be digitally deployed - the thinking has to be front loaded or the result is useless. You can’t open an agentic tool and type “make me a Constitutional Convention website” and get anything worth sharing with students. You might get a generic approximation - perhaps not awful, but mostly another version of AI slop.

The tool needs design specifications that reflect real expertise: what the activity is for, how information should be organized, what the user needs at each stage, and what problem the whole thing is solving. Without that, the AI produces something that may look impressive but serves no one.

My experience over the past several months suggests that the cognitive offloading concern inverts when you shift from content generation to design. Design-based AI work is cognitively demanding in exactly the ways educators value - strategic planning, audience awareness, thoughtful information integration, and the ability to evaluate whether something actually works the way the user intended. The process of designing with these tools isn’t a shortcut around thinking. It’s a more complex kind of thinking than what most traditional assignments ask students to do.

Why Design Rewards Careful Thinking

There’s a practical dimension to this shift that I didn’t fully appreciate until I’d been building for a while.

When you’re chatting with an AI - asking questions, generating text, iterating on a response - the iteration is immediate and forgiving. If you don’t like the output, you try again. You can afford to think loosely because the cost of experimentation is essentially zero. This is what most users have been doing for the past several years.

Design-based work doesn't operate that way. When you're building something with multiple interconnected components, specifications need to be thought through carefully or the features don't fit. This includes asking whether what you're considering is worthy of an online module in the first place. Good questions matter as much as good answers, and careful follow-up matters even more. All of this needs to happen before you ask the model to start building.

You can iterate your way to a good result, but every cycle of build, review, and rebuild costs time and adds complexity. Wobbly thinking at the start means rework at every subsequent stage.

Anyone who has ever renovated a kitchen understands this principle intuitively. You can tear out a wall and start over. But nobody who has done an expensive home improvement project thinks that’s a good plan. You hire an architect to inspect the space and make those hard decisions carefully before you ever start demolition. The economics of construction - whether physical or digital - reward careful upfront thinking and punish sloppy work.

A Note on Students

I’ve been mostly focused on the teacher side of AI use for a reason. The overwhelming majority of students are not using AI this way - not because they aren’t capable, but because they either don’t know agentic tools exist or have no incentive to use them. We don’t typically assign students to build interactive modules, so the motivation to use agentic AI for cheating in this context just isn’t there.3

I'm actually less worried about student agentic AI use than I am about basic chatbot use. A student who could design and build a working module - who could think through the architecture, map the content to a user’s needs, and make deliberate decisions about what to include and how to organize it - would have demonstrated real intellectual work. If a student showed up in my class having built something like that, my reaction would not be to worry about their academic integrity. It would be to ask them to show me how they did it.

This could change. We may reach a point where a two-sentence prompt delivers something spectacular and the design work gets automated along with everything else. But right now, the cognitive demands of designing with agentic AI are creative and significant enough that I think it offers more of an opportunity than a threat.

The Box and the Module

Building an online teaching module with AI is clearly not the same as building a wooden box with your hands. There’s something about physical construction - the weight of the material and the permanence of the object - that represents tangible evidence you made a thing that exists in the world. Digital creation doesn't replicate it.

But the pride I felt in making that box came from the same place as what I feel when I look at the Convention module: I made something I didn’t know I could make. The box taught me that I could measure, cut, align, and finish a physical object despite having no particular aptitude for building things. The module taught me that I could architect, organize, sequence, and deploy a digital one.

AI is rapidly becoming a design medium. It extends what a thoughtful practitioner can do, in the same way that a table saw extended what a clumsy teenager could build in a shop class in 1985. The quality of what comes out still depends entirely on the quality of the thinking that goes in.

Will teachers be willing to think of themselves as designers? The skills required to do that well are the skills most good teachers have spent their careers developing. The only thing that’s new is what those skills can now produce.

Going Forward in 2026

2025 was supposed to be the year of agentic AI. In 2026, it’s actually arriving. The tools are doing many of the things their proponents suggested they would - AI is being integrated into more workflows, gaining the ability to handle longer and more complex tasks, and becoming increasingly accessible to people without technical backgrounds. That’s a significant development, and I’m trying to explain what it actually feels like from inside a classroom - not as a prediction about where things are heading, but as an account of what’s possible right now.

The current conversation about AI in education is still largely split between two worlds. In one, AI companies demo features to administrators and pitch adoption at scale. In the other, schools struggle with the logistics - procurement decisions, acceptable use policies, and faculty training that never quite keeps up. And that’s for schools actually engaging at all. What's been harder to find are concrete use cases - a teacher sitting down with these tools and building something for a specific group of students.

The gap between what I’m describing and where most teachers are is significant. The pedagogical design instinct transfers - most experienced teachers already think this way about their materials and activities. But the technical fluency required to translate that instinct into a built product is real, and bridging it requires sustained investment in time and development that most schools are not yet making. I’ve had the curiosity and the circumstances to develop that fluency. Most of my colleagues haven't. But if the trajectory of the models over the past six months is any indication, that gap will start closing fast.

Yes, the concerns about cheating, privacy, security, and whether AI will ultimately improve learning are real and must continue to be debated. I’ve written about many of them. But I’ve also been paying attention to what these tools actually require when you use them seriously, and what I've found is that using AI effectively is more demanding and more intellectually rigorous than most of the conversation assumes. If that's true, the question isn't whether we need better chatbots. It's whether we need better assignments. I think we will.

The window to shape how these tools get used in education remains open, and the people willing to experiment - not caught up in predicting the future or surrendering to a predetermined narrative, but actually building things for their students and sharing what works and what doesn't - are the ones who will figure it out first.

If design thinking was ever going to find its way into how teachers use AI, this is when it starts.

Connect With Me

Beyond this newsletter, I work directly with schools, educators, and other institutions about pedagogically questions raised by AI. Take a look at my website and reach out - I’d love to hear what you’re working on.

Recently, a friend called about a dilemma he was having involving AI. He’s been sponsoring student entries to a high school essay contest for years. This year, before submitting the essays, he ran them through an AI detector. The results came back with several showing “various levels of AI usage.” The contest rules were ambiguous on the use of AI - they'd included AI language the previous year but quietly stripped it out - but his school's position was not. With less than 24 hours to decide, he pulled the flagged entries.

Following that conversation, I started looking into how student writing competitions handle AI in their submission policies. What I found mirrors what's happening in classrooms across the country - inconsistent expectations, opaque enforcement, and almost no solid ground for a conscientious student or teacher to stand on.

But I also found something I didn't expect. One organization, operating under constraints more severe than what classroom teachers face, has built an approach that most schools haven't reached yet. And even the ones that haven't gotten there are instructive - their struggles to get the policy right look a lot like ours.

What the Research Revealed

I looked at almost a dozen major high school writing competitions and one academic journal submission policy (the Concord Review). These are national and international programs that attract ambitious student writers - the kind of opportunities that carry scholarships, prize money, and college application weight. My survey isn’t exhaustive, but I looked at enough to see the pattern.1

Among the organizations that address AI, the approaches are wildly different - each with their own specific language, sub-clauses, and links to FAQ documents about plagiarism.2 None are transparent about their methods of detection and enforcement. Many ban AI outright. Several allow limited use with vague attribution requirements. One of the most prestigious international contests acknowledges the role AI is playing in student work. Others say nothing at all.

The Four Approaches

I've chosen to focus on four that represent distinct philosophical orientations toward AI and student writing.

The Concord Review: Legacy Language in an AI World

The Concord Review is one I know best. I’ve had the privilege of working with students whose papers were accepted. It treats high school academic writing with the seriousness it deserves.

While not technically a competition (their FAQ explicitly states it’s “an international quarterly academic review for the history research papers of secondary students”), the Concord Review was founded in 1987 and publishes only about 5% of submissions. These are rigorous and substantial history papers, typically 5,000-10,000 words, including Turabian endnotes and extensive bibliographies.

Their submission guidelines include this language, bolded on the page:

You must be the sole author. If at any point during the review process we discover that your work is not an original paper, we will notify your school, which could result in withdrawal of college admissions, among other things.

The threat of school notification and an explicit reference to the “withdrawal of college admissions” is one I’ve not seen elsewhere. It’s also, I discovered, identical language dating back to at least August of 2015, seven years before the public version of ChatGPT debuted.3

In 2026, AI is mentioned nowhere on its website or its FAQ. Does a student who uses ChatGPT to stress-test her thesis or runs her draft through Grammarly for feedback on syntax invalidate the “sole author” requirement? She’d probably say no. Some teachers and professional writers might disagree. My interpretation is that the severity of the language implies a blanket AI ban, but they have not engaged that question explicitly. And the enforcement mechanism - notifying the school and jeopardizing college admissions - assumes a level of certainty about the violation that AI detection software simply cannot provide, if they even use one at all.

I’m surprised they have not updated their guidelines to reflect the present-day reality. Perhaps they believe that the depth of the work functions as a structural defense against the worst forms of AI misuse. After all, it’s much harder to fake expertise at that length and level of detail than in a 700-word essay.

But I simply don’t know. What I do know is that the published language gives the student no specific guidance about where AI falls in the integrity framework. If the goal is to make a student think twice about even going near AI, I imagine it is having the intended effect, but only they know for sure.

JFK Profile in Courage: Ban, No Questions Asked

The John F. Kennedy Presidential Library’s Profile in Courage Essay Contest takes the opposite approach, and it does it in a single sentence: “Essays may be checked for AI assistance and are disqualified if AI is detected at any point in the essay or bibliography,” though its Tips for Avoiding Plagiarism makes no mention of AI at all.

I've had students submit to this contest in past years, both prior to 2022. The prize is substantial - $10,000 for first place - and the prompt is unique: describe an act of political courage by a U.S. elected official.

The policy guideline, however, is worth unpacking. The “may be checked” language is hedged. It preserves the rule without committing to universal screening, which is likely impractical. But as is the case for every competition I looked at, there are no specifics on how AI is actually being detected. With what tool? And at what threshold?

The contest says “at any point in the essay or bibliography,” but there’s no distinction between a student who generated an entire essay with ChatGPT and a student who used it to locate sources or double-check their citation formatting. There’s no acknowledgment of the spectrum of AI use that exists. Like the Concord Review, the message the JFK contest is sending is clear - don’t touch AI at any part of the process when writing or researching this paper.

One important piece to note is that, unlike with the Concord Review, the JFK contest requires a nominating teacher. The teacher’s published role is to “review the essay for syntax, grammatical, typographical and spelling errors and ensure the essay meets the contest requirements” [emphasis added]. By mandating a teacher’s involvement, the contest has cleverly created a first line of defense. Because the teacher’s name is attached, the teacher’s sign-off implicitly extends to authenticity - even though the guidelines never say so directly.

The deeper issue is how the policy employs scare tactics to police the reality of student AI usage in 2026. Imagine if at some point in your process you asked ChatGPT whether an example you were considering qualifies as “political courage” under the contest’s definition. Does that count as “AI assistance”? It’s impossible to “catch” but may induce enough anxiety for honest students to avoid AI entirely while emboldening less scrupulous ones who understand that most of how students use AI today will evade detection.

John Locke Institute: A Strategic Retreat

The third competition I want to discuss is the John Locke Institute, an international contest I was previously unfamiliar with. It had the most sophisticated but somewhat contradictory AI language I found anywhere.

The John Locke Institute is a UK-based organization (offices in Oxford and Princeton) that runs what it describes as “the world’s most prestigious essay competition,” attracting tens of thousands of submissions from over 150 countries. It’s open to students under 19 and the essay prompts are intellectually ambitious. (One of this year’s questions in Science & Technology: “Should we be polite to ChatGPT?” - which is an amusing thing to ask in a competition that has complicated feelings about students using ChatGPT to help with their paper.)

The Locke essay competition shows an organization actively evolving on the AI question: in 2025, they enforced a strict ban, with detected use resulting in automatic disqualification. This year they’ve significantly changed their approach.

Their current policy reads:

Contestants are welcome to make use of large language models (LLMs) and other related tools in order to develop their thinking on the subject of their essays, and to stress test their arguments. Such online tools should be used for research support and as a thought partner, but certainly not as a substitute for the human author.

At first glance, that looks like the most encouraging language around AI usage I found. Then, the policy continues:

Bear in mind that the kind of essay that will be produced by AI will generally be inferior - and markedly less original - than an essay produced by a human author who is engaging creatively and critically with automated tools.

But scroll down the FAQs and the real message emerges:

Since any use of AI (that does not result in disqualification) can only negatively affect our assessment of your work relative to that of work that is done without using AI, your safest course of action is simply not to use it at all.

That’s a warning. The policy page says “welcome” while the FAQ says, to be on the safe side, “don’t.” What looked initially like an embrace of AI is closer to a strategic retreat - they appear to have concluded that an outright ban was unenforceable (their own 2025 experience may have demonstrated this) and shifted to a deterrence model: we understand how many of you might be tempted to use AI, but it probably won’t help you and likely will put you at a disadvantage.

To their credit, the Locke competition is still doing something no other policy I reviewed attempts. Their language treats students as rational actors who might be persuaded that their own thinking will produce stronger work than anything AI generates. Even if the underlying message is "don't use it," the method of delivery - argument rather than threat - is a meaningful departure from the JFK approach.

But the tension between the policy and the FAQ reveals that even an organization engaging with the complex issues around AI and writing can’t quite decide whether to welcome it or warn against it. They also require an academic referee - more explicitly than other competitions that merely ask for a sponsoring educator. Their guidelines mandate that students provide the name of a teacher familiar with the student’s work, whose email they will actually contact to verify originality.

They reserve the right to ask entrants to discuss their essay with a member of the Institute’s faculty. And the guidelines reference “the Institute’s own examination techniques, and the assessment tools we ourselves have developed” that are “carefully designed to recognize and reward original thought and expression.”

Like every other organization, however, they don’t specify what those assessment tools are or what they do. The language is confident but vague. The Locke Institute’s guidelines may be the most articulate of the four I examined, but there was still a mixed message. Their policy simultaneously permits, discourages, penalizes, and verifies. In some ways, it perfectly encapsulates the full spectrum of academic attitudes when it comes to AI in 2026.

National History Day: The Honest Outlier

National History Day represents something different from the other three: a policy built on pedagogy rather than enforcement.

NHD has been around since 1974, founded at Case Western Reserve University, growing into a program that reaches over half a million students annually in grades 6-12. Students submit not just papers but documentaries, exhibits, performances, and websites.

In January 2024, NHD published a guidance document (color coded with vivid examples - something that could be posted in a classroom) that does something almost no other competition - and, frankly, very few schools - has been willing to do. They drew a clear, specific line between process uses of AI (brainstorming topics, generating research keywords, checking grammar, simplifying difficult source language) and product uses (having AI write text, create visuals, or generate analysis). Students can do the former. They cannot do the latter. If they use AI at all, they must cite it in their annotated bibliography as a secondary source and explain how they used it in their process paper.

NHD is the only contest I reviewed that requires every entry, in every category, to include a “process paper” - a separate 500-word document in which students explain how they chose their topic, conducted their research, and created their project. It’s been a core requirement for years, long before AI entered the picture. So when NHD needed an AI disclosure mechanism, they didn’t have to invent new compliance guidelines. A student who used ChatGPT to brainstorm keywords has to explain that choice in a document judges will read before they even review the submission.

And the classification of AI tools as “secondary sources” makes sense. It fits AI into a framework NHD students already understand - the primary/secondary distinction the program has taught for decades. You don’t need an entirely new set of rules if you can extend existing ones.

That’s a much more considered policy than most. But critically, NHD’s guidance also includes this:

“While AI detection tools continually evolve, there are no tools that are 100% effective at identifying whether content is created by AI or a human. Unlike plagiarism, where judges could find a source that a student copied without citation, there is no way to be completely sure that a student used AI inappropriately.”

This is an admission that many schools and teachers still don’t accept - detection tools don’t reliably work. Period.

And rather than pretending otherwise, NHD leaned into something else they already had to help assess authenticity: the interview. NHD judges talk to students about their work. They ask how you did your research and how you created your project. And the guidance tells judges to “please believe what students say during the interview” and to avoid accusatory language - specifically warning against questions like “Did you copy from AI?” in favor of more constructive framing like “I was not always sure which parts of your project were your own words or the words of others.” How they do that at scale is something I'd love to know more about, but the policy itself is far-sighted.

NHD has acknowledged a problem it cannot fully solve, built the best verification system available (a process paper; actual conversations with actual students), and asked its judges to extend trust. That’s a level of institutional honesty that few organizations have managed.

One complication I found worth noting: NHD’s Colorado affiliate bans AI “in any capacity” - at odds with the national guidance that permits brainstorming and grammar checking. Even within this organization, the rules aren’t consistent.4

It’s worth asking - what message does that send to students in Colorado?

What None of Them Have

Three things are missing across the board.

First: I could find no public reporting from any organization I examined - not just the four I profiled - on how many entries have been disqualified for AI use. I have no idea, for example, whether the Concord Review has actually notified a school about a plagiarized or AI-written submission. This absence of data means the deterrent effect of these policies operates entirely on faith. Students are being told “we may check” and the question of whether anyone has ever actually been caught remains unanswered.5

Second: with the exception of NHD, which is the only one honest enough to say outright that AI detection is flawed, none of the other contests are clear about how detection works. These tools don’t return a simple yes or no. They produce probability scores - a 40% “AI likelihood” rating on a student’s essay is exactly the kind of ambiguous result that leaves a teacher with no actionable information. My friend’s “various levels of AI usage” is a textbook example.

Third: what happens to a student who is wrongly flagged? I could find no competition publicly describing a process for handling false positives. The JFK contest says disqualification occurs “if AI is detected” - but what if the detection is wrong? Legitimate human writing gets flagged regularly, particularly writing by non-native English speakers. High schools and colleges are dealing with this on a daily basis. The John Locke Institute states that “our determinations in all such matters are final” - the opposite of an appeals process. Only NHD addresses this at all, by telling judges to believe students and avoid direct accusations.

These aren’t just academic observations. My friend’s experience reveals exactly what happens when competing guidelines conflict. He was carrying an institutional burden the organization never acknowledged, placed in an untenable position that required him to exercise judgment using tools that produce inconsistent results. His school had once subscribed to a detection platform but no longer did, leaving him to improvise with browser extensions on a 24-hour timeline. Not every teacher from every school is likely to be so careful.

The Lesson Nobody Intended

What must all of this look like from the student’s perspective?

A student researching for just two of these competitions would encounter three contradictory positions before writing a word. The John Locke Institute's policy page says LLMs are welcome thought partners. Its FAQ says the safest course is not to use AI at all. The JFK contest says if AI is detected, you're disqualified.

Further complicating matters, your school has its own AI policy - maybe it allows AI for brainstorming but not drafting, maybe it bans AI entirely, or maybe it hasn’t said anything at all. If your school bans AI and you submit to the John Locke competition using AI as a “thought partner,” you are violating your school’s policy even though you’re following the competition’s rules. If your school allows AI with disclosure and you plan to submit to the JFK contest, you’re not permitted to do something your school considers routine. Both competitions require a sponsoring teacher to sign off on the submission - a teacher now caught between the school's policy and the competition's. For contests that don’t require one, students are left to their own devices, many of whom will submit AI-assisted work, some of which will get past whatever gatekeepers exist.

But there is an even bigger lesson lurking beneath these contradictions. NHD has drawn a clear line between process and product uses of AI and built a disclosure system around it. What message does it send students when a national competition with real stakes - prize money, prestige, college applications - are actively wrestling with how AI fits into serious academic work while their own classrooms are avoiding the conversation?

If a student's English class treats ChatGPT as cheating, but an extracurricular essay competition invites her to stress-test her arguments with it, she's going to start asking which environment is preparing her for the world she's actually entering. That's a question almost none of these organizations - and very few schools - are ready to tackle.

Here’s another question I’d have for competition judges: do papers with AI attribution bear some kind of stigma? If you’re a student submitting to a competition that allows AI with disclosure, and you honestly disclose your use, does that disclosure work against you? The Locke Institute suggests that, yes, it does. Their FAQ states that AI use “can only negatively affect” a student’s assessment. So what’s the incentive to disclose? Anyone following the discourse knows that many educators are deeply skeptical of AI-assisted writing. Locke made that skepticism policy.

And what about rejections? When a student’s paper is rejected - which for most programs is the vast majority of entries - do they receive any indication of why? If a paper was flagged by an AI detector, does the student ever find out? Or does it disappear into the ether alongside hundreds of others, with the student left wondering whether their work wasn’t good enough or whether an algorithm decided they were a cheater? The process makes it impossible to know.

When the rules are contradictory and detection and enforcement methods are unclear, students can sniff out hypocrisy like bloodhounds. If the adults in charge of determining contest "winners" have conflicting approaches to what's authentic and what's not, how can students have any confidence that what's being evaluated is fair?

Back in the Classroom

My research started because of a phone call. I keep ending up back in my classroom because the parallels are impossible to ignore.

The four approaches I’ve described - silence, ban, strategic deterrence, and structural integration - map almost exactly onto what I see happening in schools across the country. Some departments ban AI entirely and treat any use as plagiarism. Others allow it in theory but make clear - through grading rubrics, pointed comments, or classroom culture - that AI-infused work will be viewed negatively. Others have built sophisticated frameworks distinguishing between process and product uses. And a not-insignificant number continue to say very little, hoping the problem resolves itself.

The writing competitions are a microcosm of the same fragmentation. They're easier to study because the rules are public and the outcomes are clear - you're either published, selected, or you're not. But the underlying dynamic is identical: students using AI in the same way might be rewarded in one context, punished in another, and face no consequences at all in a third.

A Silver Lining?

Here's what I didn't fully appreciate until I started this research. The organizations that are doing the most thinking about AI in student writing operate under constraints that are in many ways more severe than what classroom teachers face. They can’t fall back on “I know my students’ writing.” They can’t rely on watching a kid struggle through a draft over three weeks. They don’t have a writing sample from September to compare against a February submission. They have to make a judgment based on a single artifact.

That’s a harder problem than what most of us face in our classrooms. And the two organizations that engaged with it most seriously arrived at complementary insights. NHD built something practical - a process paper, an annotated bibliography, and an interview - a pedagogical methodology that was never designed for AI but turned out to be perfectly suited to it. The Locke Institute identified the right principle: that human thinking, creatively and critically applied, produces better work than AI can generate. They also require an academic referee who can vouch for a student’s work and reserve the right to ask contestants to discuss their essays with faculty. This is Locke’s first year under the new policy. The pieces are in place. The next step, if possible, would be making those conversations routine rather than discretionary, especially when questions arise.

Competitions and classrooms are obviously different writing environments with different goals and purposes. A contest that sees a student once can’t do what a teacher who sees her every day can do. But the reverse is also true. If organizations processing tens of thousands of anonymous entries have found ways to evaluate authenticity beyond simply running a document through a detection tool, teachers who know their students personally have more opportunities to try - and fewer excuses not to do so.

Making Thinking Visible … Again

We know that human thinking produces more original and more compelling work. That’s the argument the Locke Institute makes emphatically. But saying so isn't enough.

As I’ve said before, we need to build the assignments that make student thinking visible. This is what my own department is attempting this spring with our redesigned research papers, and it's what both surprised and impressed me about NHD. When a student has to narrate how she arrived at her thesis, what she learned from one of her sources, and where her argument changed direction, AI becomes one tool in an observable process rather than just a black box that produces a finished product.

The competitions that will thrive in an AI world are the ones that stop treating AI as a compliance problem and start treating it as a pedagogical one. That’s harder, more time-consuming, resource-intensive, and messier than a ban.

My friend was placed in an impossible position. He had a deadline, a jury-rigged detection strategy, and no framework from anyone - not the contest, not his school - for what to do. So he did the cautious thing and pulled the essays. The question this piece is asking is why, in 2026, that's still the only defensible option available to a conscientious teacher.

If you have a story about how you’ve shepherded a student through a competitive writing submission in the age of AI that you’re willing to share, I’d love to hear about it in the comments.

Connect With Me

Beyond this newsletter, I work directly with schools, educators, and organizations navigating AI Awareness and integration. Take a look at my website and reach out - I’d love to hear what you’re working on.

One thing is clear - AI cannot shovel snow.

Agentic AI is currently all the rage - coding tools released by OpenAI and Anthropic over the past six weeks have driven the conversation across legacy media and Substack alike. It does not require accepting the hype to acknowledge that something has shifted. As I sit and watch a late February blizzard blanket my neighborhood, I’ve been reading and thinking about what the rise of agentic AI means for teachers, students, and schools, and it underscores something I’ve been reluctant to say outright.

Practically since its arrival, schools have treated AI mostly as a problem - how do we stop students from using it to cheat? What can we do to slow all this down? How can schools design policies to prevent AI from destroying authentic student work? These are real and ongoing issues and they deserve serious treatment. I’ve catalogued many of them.1 But the “AI is all negative for schools” stance has hardened into orthodoxy and I worry that it’s making it harder to see anything else.

The increasing recognition that AI is evolving from a series of chat interactions into the ability to build entire projects from scratch introduces a fundamentally different question. And that question is going to require more imagination than schools may be capable of given where most teachers currently are.

For starters, I simply don’t think that many educators are aware of how quickly the landscape is shifting. I applaud those teachers who have paid enough attention to try to get more knowledgeable and fluent with the more advanced models. But it seems that just when many schools are finally deciding it’s time to confront the “AI issue” - either by paying for AI wrappers like Magic School (a decision with more consequences than many realize) or bringing in outside consultants to review the basics - the conversation is moving elsewhere. It doesn’t mean that learning the basics and understanding the strengths and weaknesses of LLMs isn’t important. It’s just that, for anyone who follows this closely, it’s clear it isn’t going to be enough.2

I’ve watched the response to agentic AI unfold over the past several weeks across a dozen Substacks and podcasts, and I’m struck by how much of it has followed a familiar pattern: here is a powerful new capability, but the focus in schools is mostly on why it’s dangerous. The concerns raised - more cheating, student data privacy, and the security risks of connecting AI agents to institutional accounts - are all legitimate and will need to be addressed. But when the primary response to a genuinely new development is to mostly fixate on the ways it might go wrong, I think the conversation is incomplete. A preoccupation with the downsides that crowds out everything else may prevent educators from recognizing the potential usefulness of a transformative technology.

Seeing Is Believing

I remember the first time I saw a student share a Google Slides presentation from the cloud. It seemed like magic at the time - the ability to host digital work that could be accessed anytime, anywhere, from any machine. I hadn’t used the Google suite of products very much in the early aughts and that one demonstration convinced me that cloud computing was the future. I needed to see it to understand it. Google Drive was once unknown to me, as it was to most teachers, but it’s now an essential part of how I store, organize, and work with all my digital materials. It happened over many years and now I can’t imagine living without it.

Technology works like that. You can read as much as you want, but until you see what everyone is talking about, until you watch an AI agent design and deploy an online module or website in real time, until you begin to realize what this means for simply the manner in which we produce documents - spreadsheets, presentations, memos, reports, and virtually anything else that requires organizing information into precise digital formats - you are not likely to appreciate how this will affect our work habits.

What Agentic AI Has Allowed Me To Do

In my Government and Politics class, I designed an Ethics Bowl case built collaboratively with AI. I could not find a suitable article summarizing the past year of U.S. foreign policy that was not written from a hyper-partisan perspective. Instead, I co-created a thorough analysis of the current administration’s foreign policy decisions, framed through the lens of Thucydides’ Melian Dialogue and Kantian ethics. The case went through multiple drafts across several days. I subsequently used a separate AI project to fact-check every quote, date, and statistic, catching one major error that I corrected. I am confident I could not have produced a case of that depth and rigor in the time I had without AI collaboration.

I also wanted students to have an interactive preparation tool - something they could use independently to click through each policy area, toggle between factual summaries and competing arguments, review key terms from the textbook chapter, and test their comprehension before the live discussion. Six months ago it would have been impossible. Building an interactive web application is not something I knew how to do. I don’t know how to code.

Using Anthropic’s Cowork - one of the agentic tools that everyone is currently writing about - I described what I wanted. I uploaded my case document and assignment instructions. And over the course of a few working sessions, the AI built it. A functioning interactive web application with tabbed navigation, expandable case sections, a glossary connecting textbook vocabulary to the specific case, verified source citations, and a comprehension quiz with immediate feedback. I had it hosted at a live URL and shared it with students through our learning management system.3

The same session also produced a fifteen-slide introductory presentation for the activity. All from the same source documents.

The process was not seamless and I needed to troubleshoot a handful of steps. The tool needed my professional judgment - what to include, what to cut, when something wasn’t right, how to redirect. It took some time. But it was work that produced something I could not have produced alone, and it was work that would have been unthinkable for a classroom teacher even a year ago.

That was the moment when the possibilities crystallized.

The Gap

Many teachers don’t have the time or inclination to do this and many won’t for months or even years.4

I’ve had a curiosity about digital tools from an early age and invested in the conditions to make it possible to create the use case described above: a paid subscription to advanced frontier models, custom environments I’ve built over months of experimentation, and a specific pedagogical need that gave me a reason to investigate what the tools can do right now. I’ve spoken with enough colleagues and led enough conversations to know that most educators haven’t encountered these tools, let alone used them.

Knowing that I have the ability to build interactive web applications means I can’t unknow it. More use cases and possibilities will keep coming. The distance between those two realities - between where most educators are right now and where the technology already allows motivated practitioners to go - is likely to create more of a divide in schools, especially if students begin to incorporate these skills into their work on their own.

Imagination As Well As Caution

The dominant response to AI in the education space - currently extending to agentic AI - is primarily to focus on what’s gone wrong.

And I understand why agentic AI is so scary for schools - when someone describes a tool that can autonomously access a student’s Google Drive, browse the web, and execute tasks without supervision, the immediate instinct is to worry about data privacy and security. That makes sense.

However, I’d love to see the conversation broaden. What I’m arguing for is imagination as well as caution. Believe me, I understand that generative AI, specifically agentic AI, represents significantly more threats than Google Drive. But if we want to treat it as “normal technology,” as many who resist the hype are preaching, then we need to understand the typical progression that “normal technology” takes. I’m not suggesting we ignore the risks. I’m suggesting we match them with equal imagination about the possibilities.

An excessive focus on what could go wrong produces a profession that is perpetually reactive - always responding to what the technology does, never shaping what it could do. And it produces a widening gap between what educators currently think AI is (simply a cheating tool, a security risk, a privacy nightmare) and what it may actually become (a genuinely powerful instrument for creating things that were previously impossible for non-technical professionals to build). I realize this may be a counter-intuitive take at the moment, but I think it needs to be part of the dialogue. There are still people who see AI as nothing but a threat to education, but my experience and that of a growing number of practitioners suggests otherwise.

What if the question encompasses more than just “how do we protect students from agentic AI” but also “what could we eventually ask students to build with these tools that they couldn’t build before?” What if assignments evolved beyond “write an essay” or “make a presentation” toward “design an interactive tool that helps someone understand this problem?” That question requires imagination and a sense of possibility, both of which are sorely needed right now.

Where This Leaves Me

Schools are likely to escape the full implications from agentic AI, at least on the student side, for at least another year or so, if the pattern of adoption rates is anything like it’s been up until now - mostly slow, uneven, and imperfect. Teachers have not been incentivized or trained to integrate agentic workflows into their routines. But just like Google Drive, I suspect agentic AI tools will simply become part of how many of us work in the near future. I’ve already seen too much value to pretend otherwise.

Moving through the current moment as an educator is going to require more imagination than has been demonstrated up until now. The same concerns about cheating and authenticity will persist, but I suggest that instead of panic and surrender, we embrace both the uncertainty and the possibility. What AI is now capable of is genuinely helpful and, to me, remarkable - while at the same time unsettling and, for many, frightening. Those things will have to coexist.

I’ve been reluctant to make predictions in this newsletter - the field moves too fast and most of what’s coming is outside my area of expertise. But I do know the classroom and I’ve been creating curricular materials for three decades. The nature of how AI tools are evolving represents the single most interesting and exciting development on the teacher side I’ve ever experienced.

With respect to AI’s overall impact and whether it does what its most ardent proponents promise or its most informed critics predict, I still don’t know. Frankly, no one else does either.

The provocative title of this piece is a movie reference, but I also think it will be the literal truth. AI is going to be everything, everywhere, all at once. The question for educators is what we are going to do about it and how we frame it - as an ongoing problem to manage or as something that might actually change what’s possible in a classroom. I’m holding out that we can rise to the occasion. I don’t see that we have much of a choice.

Connect With Me

Beyond this newsletter, I work directly with schools, educators, and organizations navigating AI integration. Take a look at my website and reach out - I’d love to hear what you’re working on.

AI is getting boring. Not boring in the sense that each new AI story isn’t interesting – but boring in its predictability. Watching my sixth grader do her homework last week may have taught me more about where AI in education is actually headed than a year of reading the current debates.

New capabilities are announced. Some herald these developments as game changers, arguing that a radical revolution is just around the corner, while others dismiss the hype, focusing primarily on the costs – economic, environmental, and human – and downplaying the technology. And around the merry-go-round we go. It’s getting tiresome.

This past week alone, a 26-year-old CEO’s viral essay about how something enormous is happening was viewed more than 50 million times. Unsurprisingly, one of the most prominent AI skeptics responded by calling it ridiculous and asking for evidence. Another post I saw made a compelling, rigorous, data-driven case that the trajectory of AI improvement is essentially vertical and that anyone who isn’t freaking out simply isn’t paying attention. Addressing precisely this trend, a well-known writer made an equally sharp case that the AI discourse itself is the problem - that the technology is being deliberately mystified by the companies who profit from the mystification and that the actual utility, while real, is far more modest than the hype suggests.

The camps are still entrenched. The arguments recycle. Rinse, wash, repeat.

And that's just a sampling of headlines from the past week or so. None of them are decisive, all are exhausting, and all underscore how impossible it is for educators to navigate the AI information ecosystem while also trying to teach.

For the past year, I've tried to see all of this through a slightly different lens.

I published my first post on this Substack a year ago. I was cautious and full of questions I didn't have answers to. I quoted William Goldman's line about the movie business - "Nobody knows anything" - because it felt like the most honest thing I could say about AI in education at the time. Sixty posts and over eight thousand subscribers later, I still don't have many of the answers. But I do feel like I know at least a few things.

What a year of writing, teaching, and watching from inside a classroom has made clear is that AI is here, it's being absorbed into schools the way every technology before it has been, and we are all going to have to figure it out - mostly on our own and without the luxury of certainty.

Normal Technology

If I’m fortunate, I will remain in the classroom for another ten or perhaps even fifteen years. The decade ahead could be the most exciting of my career or the most frustrating. Likely both. The pattern is remarkably consistent. Personal computers, laptops, the internet, smart boards, learning management systems – each arrived with transformative promises, and each was incorporated into the existing structure of schooling without fundamentally altering it. The tools changed but the teaching, for the most part, didn’t.

I wonder what I might look back on in ten years as the moment when I realized that generative AI was just going to be another technology we were asked to adapt to - normalized the way everything else was.

The current contender is watching my sixth grader this past week work through a series of prompts designed to help her with a short essay using our school’s AI wrapper.

What was most remarkable about it was how unremarkable it all was.

For about thirty minutes I watched her iterate her essay about the Roman Empire using a thoughtfully scaffolded lesson that pushed her to incorporate evidence into her argument, clarified content questions when she was stuck, and provided useful examples and analogies to help illustrate the goals of the assignment. She wasn’t amazed and she wasn’t confused.

She wasn’t cheating. She was doing her homework.

I was watching with a teacher’s eye, not just a parent’s. The lesson design kept her in the driver’s seat – it asked questions rather than supplying answers, and when she tried to move past the evidence step, it pushed back. She had to do the thinking. She was so proud of her essay that she wanted to read it to me aloud. That was new.

As someone who has expressed real ambivalence about AI use in middle school, I was genuinely surprised. Observing her actually do the assignment made me much less worried than thinking about it in the abstract. When I asked her what she thought about it, she was matter-of-fact. “I liked the analogies. Those were helpful. It was my essay. It didn’t do it for me.”

She's eleven. She doesn't know what AGI stands for, she's never heard of scaling laws, and she couldn't care less whether AI is conscious. She was just trying to finish her homework. What I watched was the opposite of what I've read in virtually every comment section under any positive AI coverage of education in schools - that an entire generation of kids is going to be rendered incapable of thinking for themselves. What I saw was an eleven-year-old who did the work, finished her essay, and wanted to explain it to me. Any parent of a tween knows that’s a win.

Three years ago, when she was in the third grade and I was just beginning to grapple with ChatGPT, it was inconceivable to me that she would enter high school in a world where schools could still afford to ignore AI. Now, here we are in 2026 and it’s a homework assignment. Not a crisis. Not a revolution either.

After writing about AI in schools for a year – the costs of cognitive off-loading vs. the potential benefits of personalized learning, the heated debates, the existential hand-wringing – what I watched in our living room was the AI revolution taking place in real time. Not with a bang but a whimper. In this instance, the personalized tutoring carried the day. I’m sure that won’t always be the case. A less carefully designed lesson, or a student less willing to push through the friction, and it can go sideways fast.

This particular lesson worked because a colleague of mine didn’t use a generic pre-loaded template. The lesson was designed with the same care and attention she'd bring to any assignment - tested and fine-tuned to meet a specific learning goal: pushing students to use evidence for their claims. The tool didn’t make the lesson good. The teaching made the tool useful. That distinction tends to get lost.

My daughter is hardly alone in this experience. Thousands of students across the country are using AI wrappers like Flint, MagicSchool, and others in their classwork every day. It’s not in every class and it’s not in every assignment – not by a long shot – but it’s happening and we will not know the ramifications until more data is collected, analyzed, and sorted.

By then, my daughter will be headed for college.

Normal Adoption, Abnormal Pace

What I’ve learned over the past year, and should have been less surprised by, is that AI adoption in education is proceeding on the exact same trajectory as every previous technology – slow, shallow, uneven, and lumbering in all the ways that institutions generally are. In the spring of 2023, in the flush of seeing what AI was capable of, I thought the rate of adoption would go much faster. I was wrong. But I’m glad that I was.

Justin Reich’s Failure to Disrupt: Why Technology Alone Can’t Transform Education disabused me of this notion. The history of educational technology is a history of overpromising and underdelivering. AI has been no exception.

In some ways this is reassuring. It takes years for new technologies to get integrated into existing systems and longer still for the shape of the impacts to become clear. Despite almost weekly predictions of imminent and devastating job losses, the current numbers have not borne that out. AI, by this reading, and certainly while witnessing my daughter complete her homework assignment, is “normal technology” - significant, yes, but following the same basic adoption curve as everything that came before it.

But here’s what makes me less comfortable with that framing. Even incremental improvements are developing at a pace that no previous educational technology has approached, and the distance between what AI can do and what most educators know it can do is widening, not closing.1 The adoption rate is normal. The technology is not. And I’m resigned to the fact that these two realities are going to coexist uncomfortably for the rest of my career. The tools will keep outpacing the training, the policies will keep chasing the capabilities, and teachers will keep being asked to make decisions about technology they haven’t had time to fully understand.

Lead with the Pedagogy

I don't see any single decisive moment where teachers everywhere will simply need to wake up to the fact that generative AI tools are going to be with us for the duration. Many would say that moment has already passed, but that's not how institutional change works. Unlike COVID, where the crisis was imposed directly and schools had no choice but to shut down overnight, no degree of urgency around a question of technology will force that kind of speed. It will still be one class at a time, one week at a time, one semester at a time and, critically, one teacher at a time. Schools march on because they must.

The controversy right now revolves around the extent of what “navigating” AI should look like as we all continue to sort out boundaries we feel comfortable with, review imperfect data that will likely be out of date by the time it is published, and simply use our own two eyes to evaluate what’s working and what isn’t.

I maintain that if the conversations schools are having about AI revolve around whether the technology is good or bad, or just about whether AI can perform a particular task, those conversations are mostly a waste of time.

The only productive conversation starts with pedagogy: what are the skills, habits, and dispositions we want students to develop, how do we know they’ve developed them, and how does AI support or undermine those goals? I’ll bang that drum over and over because I think it’s the only path forward.

Educators have to lead with the learning. Not the technology. That’s what I wrote about last week with our department’s research paper redesign, and it’s what I’ll keep writing about – not because I’ve run out of things to say about the tools, but because it’s the only approach I’ve seen that doesn’t become obsolete the moment the next model drops.

I don’t know what the lifeboats look like. I don’t think anyone does. But I know they won’t be built by people arguing about whether the flood is real. They’ll be built by teachers. They always are.

Connect With Me

Beyond this newsletter, I work directly with schools, educators, and organizations navigating AI integration. Take a look at my website and reach out - I’d love to hear what you’re working on.

My father was a top litigator, member of the American College of Trial Lawyers, and feared opponent in the courtroom. I was too young to understand everything he was doing professionally, but my memories of his work habits are vivid. After a 10-12 hour day in Manhattan, he’d be home in his favorite chair with a brief in his lap, surrounded by memos, files, and accordion folders, preparing for a case.

He was born in 1925, a member of the Greatest Generation, and his career spanned the postwar years through the early 2000s. Despite representing cutting-edge technology clients, my father never owned a computer, a cell phone, or a laptop. He was probably the only attorney alive when he passed in 2011 who did not have an email address in Martindale-Hubbell1. He was strictly analog. Everything he did was by hand on legal notepads or markups of printed material.

His work ethic was legendary. As I got older - even attending law school before pivoting to teaching - my appreciation of his career deepened. I began to hear stories from former clients, older family members, and even opposing attorneys about his most famous cases. One of the traits that set my father apart was his ability to locate virtually any document at a moment’s notice during the pressure of a trial.

To appreciate how difficult this was in, say, 1980, imagine a case involving thousands of pages of documents, any one of which might become critical during a witness’s testimony. Before the age of digitization, lawyers brought in hand trucks of bankers boxes with folders bursting at the seams. My father was known for many things in the courtroom, but a signature skill that made him so formidable was a meticulous filing system he had built over decades that allowed him to find the precise document at exactly the right time.

“When Joe came into court and we saw how he organized the case with that level of detail,” one older attorney told me, “we knew we were in trouble.”

I miss my father for lots of reasons, but those work habits came back to me recently when I sat down with one of our strongest students to talk about how she actually does research - part of my preparation for our history department's effort to redesign the way we teach it.. Her approach mirrored much of how my father operated decades earlier, and it reminded me of a concept I first encountered at Harvard's Project Zero: "Visible Thinking."

The Question We Started With

The impulse for the assessment redesign was obvious.

AI can now generate an outline in seconds. It can produce a thesis statement, identify sources, summarize dense academic articles, and write polished prose that is, for most practical purposes, undetectable. I’ve spent the better part of two years teasing out every stage of the research process - from topic selection through final editing - and at virtually every step, AI has the potential to either augment or completely derail the cognitive work the assignment is supposed to teach.

AI itself is not the primary culprit. User intent, AI fluency, discernment, and domain expertise determine whether AI adds value or replaces thinking.

The problem is that mastering these skills requires the kind of discipline and judgment most teenagers haven’t developed yet.

And it’s not just a high school issue. My sixth-grade daughter recently came home and told me she and a friend are entering an “AI Live”2 competition in which middle schoolers create a piece of performance art conveying a message about their relationship with technology, documenting how they’ve used AI in the process. Students as young as fifth grade are using AI wrappers to help structure their essays. I’ve already documented the influence of AI in the middle school debate community, and research databases like JSTOR are integrating AI into their search tools.

As Hemingway once observed about bankruptcy, AI is infiltrating every corner of the school experience - gradually in 2024 and 2025, and now suddenly in early 2026.

This is the conversation every school is having right now, and it almost always starts in the same place.

So when our department sat down to talk about the research paper, the instinct was natural and understandable: how do we make this assignment AI-proof?

But that turns out to be the wrong question. Or at least, an incomplete one. Because once we started talking seriously about what AI could compromise, we found ourselves in a much more interesting conversation - one that had less to do with technology and more to do with what we actually believed a research paper was for. The question that proved more productive was deceptively straightforward: what are the non-negotiable thinking skills we want students to develop through this process, and how can we feel confident that they’ve developed them?

What One of Our Strongest Students Already Knew

In anticipation of our meeting, I sat down with one of our strongest research students - a senior who had written outstanding papers across multiple years of our history curriculum and had also distinguished herself in rigorous independent research in another discipline entirely. I wanted to understand how she approached research, not as a theoretical exercise but as someone who has repeatedly demonstrated it at a very high level.

What she described was reinforcement of everything I’ve been reading over the past two years. She talked about becoming genuinely invested in a question and doing more research than she needed as the material pulled her forward. She described a process of using sources to find other sources - following footnotes, tracing citations, clicking through to the next article to see where a particular claim originated. “Once you come upon one good source or one good piece of information,” she told me, “the best way to find the next best thing is to use that source to find other sources.” It sounds simple, but it’s actually a sophisticated research disposition that many students never fully embrace.

Her organizational system was equally revealing. She would start with a general sense of her argument, create categories for her body paragraphs, and then research into those categories - color-coding primary source evidence, leaving herself analytical notes in the margins so she wouldn’t forget why a particular passage mattered weeks later. “You kind of need to catalog your thoughts a little bit if you’re working on something more long-term,” she said. She wrote her topic sentences last, because the structure emerged from the thinking, not the other way around. When I pointed out that she seemed to be outlining and writing her paper as she researched it, she agreed. “It’s extremely iterative,” she said.

This is the conversation that got me thinking about my father - here was an 18 year old instinctively developing the kind of skills necessary for deep, lasting learning - and then methodically categorizing her work so she could come back to it days or weeks later. Sure, as someone growing up with the internet and online source material, her tools were digital, but the thinking and overall approach was eerily similar to how my father dug into his case preparation 50 years earlier.

The Virtues of an Analog Process

I asked her what she thought about moving more of the research process to paper-based, handwritten formats. Her answer surprised me with its nuance. She acknowledged that a student working digitally without AI could probably produce a better final product than a student working entirely on paper. But then she added: “In terms of building skills, the final product of that research paper might not be as good, but more skills would be built by the student by doing it in an analog manner.”

When I asked about AI’s role in research more broadly, she was remarkably specific. The worst use, she said, was at the very beginning - for brainstorming and generating ideas - because “it can become difficult at a point to differentiate what’s your own original idea and what you’ve read through AI.” This is utterly at odds with much of what you read about using AI to brainstorm - and she is not the first student I’ve heard this from.

She didn’t think it should be used for outlining either. And when I asked whether learning to use AI was itself an important skill to develop, she was direct: “I don’t think using AI is as much of a skill to build as other things, like learning how to research and write by yourself. I think that is something that students can learn to do later in life. What’s difficult is to build those fundamental skills.”

This from a top student who has figured it out on her own. I’ve no doubt that if and when she integrates AI tools into her work routine she will continue to be successful. She’s already developed more authentic research skills than most of her peers - and many adults for that matter - but the critical distinction is her mindset. My conversation with her aligns with the gut feeling I’ve had for the past few years that the strongest students understand instinctively where AI harms them. They are committed to doing the hard work first because it works.

The Ongoing Debate

Unfortunately, the picture is not going to be as clean as “keep AI away from students until they’ve built the skills.” In our own department, I teach an advanced Independent Research in History course, with students enrolled who are heading to college in a matter of months, where I’ve slowly introduced AI tools for discrete stages of the process. The Deep Research models allow students to pinpoint relevant sources faster and with more specificity than ever before. NotebookLM and other digital tools with AI integration allow for a level of organization and file scanning that has already become the norm across industries.

I think that’s appropriate for students at that level, and I’d make the case that exposing advanced students to these tools before they encounter them unsupervised in college is part of our responsibility. But I’m honestly torn, because the line between “ready” and “not ready” is never going to be as bright as we’d like it to be. And the fact that one of our most advanced research students does not rely on AI for any stage of her process speaks for itself. If and when AI enters the picture, it should add value and never replace the core skills.

Disciplinary Differences

The picture gets even more complicated when you look across disciplines. In the science field where she also excels, the organizers behind the major competitions have released explicit guidelines permitting students to use AI for improving the phrasing and clarity of their writing. The use of AI is becoming accepted practice in scientific research more broadly, where communicating findings clearly is essential while writing has never been the primary skill being assessed.

The trend is clear: different disciplines are arriving at different answers, and the humanities’ emphasis on writing-as-thinking puts us in a fundamentally different position than fields where writing may be seen more as a delivery mechanism for other kinds of expertise.

What holds all of this together, I think, is that the hardest skills to build are the most human ones. To perform advanced research, you need judgment, patience, perseverance, and skepticism - which are far more difficult to measure.3 These are also prerequisites for developing AI fluency. Not every student will be ready for open-ended AI use on the same timetable, and not every discipline will draw the line in the same place. But the sequence matters. Independent thinking always has to come first.

What We Built

Every humanities department in every school is going to have to work through some version of this conversation, and not everyone will land in the same place. Some will be more conservative about AI use, others less. The key is being honest about the trade-offs - and then designing accordingly.

What our department ultimately landed on was less an “AI-proof” assignment and more an attempt to make the entire research process visible and documentable. We were thinking practically about how to shift the burden away from surveillance and toward structured evidence of student thinking at every stage. The redesign moved almost all of the cognitively critical work into the classroom and onto paper - handwritten worksheets, annotated hard-copy sources, in-class quizzes that require students to explain their reasoning rather than reproduce information. The grading structure shifted accordingly, with the final product counting for significantly less than the process checkpoints that precede it.

I don’t want to paint this as a finished product or a triumphant case study. We’re still working through real questions, and I suspect every school attempting something similar is navigating its own version of the same tensions. But what interests me most is a connection to something I first encountered in a world where generative AI wasn’t remotely part of the conversation.

Making Thinking Visible

Twice in the past twenty-five years - most recently in 2009 - I attended Project Zero, a week-long summer institute at the Harvard Graduate School of Education. Project Zero has been around for decades, and one of its central preoccupations is what they call “Visible Thinking” - a research-backed framework built around the idea that if you want students to think well, you have to make the thinking itself visible, not just the products that thinking is supposed to produce.

The approach emphasizes three core practices: thinking routines that structure how students engage with content, documentation of student thinking as it happens, and reflective professional practice among teachers. Project Zero is still going strong, and its framework is arguably more relevant now than at any point in its history.

What struck me as our department’s redesigned research paper took shape is that we had independently converged on almost exactly this framework - without everyone in the room having heard of it. The timeline quiz that asks students to defend their periodization choices is a thinking routine. The annotation requirements that force students to analyze primary sources through multiple lenses are documentation of thinking. The revision log that tracks changes in argument and evidence is a record of cognitive development over time. The one-on-one conference is a moment where a teacher can observe thinking in real time, rather than simply evaluate a product after the fact. Even the final reflection, in which students assess the quality of their own evidence, is a metacognitive exercise straight out of the Visible Thinking playbook.

We Can’t Avoid These Questions Any Longer

The pressure of AI forces us to answer a question that schools have been able to avoid for a very long time: if we can’t trust the product, what evidence of thinking do we actually have? For most traditional assessments, the honest answer was: not much. The research paper, as most of us experienced it in school and as many of us have been assigning for years, was always a product-focused exercise in which process was assumed rather than observed. AI has exposed just how fragile that assumption was.

Visible Thinking is not an AI strategy. It was coined years before anyone imagined a world where a chatbot could write a competent five-paragraph essay in 15 seconds. But it may be the most useful framework schools have for responding to this moment - not because it “solves” AI, but because it redirects the conversation from surveillance to pedagogy, from policing products to designing for thinking. Interestingly, in Project Zero’s current materials for this summer’s institute, I found virtually no mention of AI. It’s always been about thinking.

An Analog Approach in a Digital World

My father's filing system may have been from another era, but the practice of externalizing one's thinking - organizing, categorizing, making connections visible - is as old as the written word itself. He simply adapted it for the courtroom. Reflecting on his practice now connects to the Visible Thinking concept I didn’t have a name for until I sat in a lecture hall at the Harvard Graduate School of Education more than twenty years ago.

Every document organized, every connection ingrained, every piece of evidence coded and accessible because the thinking behind it was explicit and externalized. When opposing counsel saw those tabs, they weren’t intimidated by the system itself. They were intimidated because they could see that someone had already done the cognitive work of integrating thousands of pages into a coherent argument, and that the thinking behind that preparation meant every document could be retrieved at the moment it mattered most.

I know my father would not have stood in the way of AI advancing and improving efficiency for routine tasks in his office. AI will automate many kinds of knowledge work, much of which we’re still figuring out. But it doesn’t change the importance of visible thinking for the truly essential cognitive efforts that require human judgment, synthesis, and review. I know he would have resisted using AI for his own thinking.

And I, for one, would not have bet against him with just his hard copies, legal pad, ballpoint pen, and a highlighter.

Connect With Me

Beyond this newsletter, I work directly with schools, educators, and other institutions about pedagogically questions raised by AI. Take a look at my website and reach out - I’d love to hear what you’re working on.

I'm breaking my own rule. I told myself I wasn't going to chase AI news this year - that I'd focus on first principles and pedagogy and leave the breathless hot takes to others. But the Moltbook story that broke over the weekend hijacked the post I was planning to write, and the resulting tension revealed something fundamental - no one is experiencing AI in quite the same way.

Last week I got access to Claude's Cowork. This wasn't my first attempt at using an AI agent, but it was the first time I gave one direct access to my desktop. Over the summer, I'd used ChatGPT to help reorganize my Google Drive folders - a cumbersome process where it gave me terminal commands, I entered them, things went sideways, we troubleshot, and I entered more commands. It worked, eventually, but barely felt worth the effort. Over winter break, I tried something similar with Claude Code. Smoother, yes - but still with enough stops and starts that it never truly felt autonomous.

When the Friction Disappeared

Cowork is different. After I gave it access to my desktop folders, it just worked. The friction was gone. I described a task, watched it execute, and realized I was experiencing a future where this is simply how we will interact with our computers.

I’ve been reading about agentic AI for months. Now I could finally see what others had been describing - useful task delegation, happening on my own machine.

But the excitement came with a second realization. Most people are nowhere near this. The teachers I talk to are still working through basics - how to structure a prompt, how to attach a file, how to fit AI into a workflow at all. These tools have been available for over two years, and most educators have had neither the time nor the incentive to move beyond fundamentals. Something like Cowork isn't on their horizon. By the time it is, there will be something else further out.

It’s a game of catch-up that can’t be won by running faster.

And then, while I was still processing what Cowork meant, my son texted me about what was happening with Moltbook.

A Social Network For Agents

The full story is still unfolding, but here are the basics1 - someone created a social network for AI agents (called Moltbook for reasons not relevant here) where humans could observe but not exactly participate. Within days, what started as dozens of agents metastasized into thousands, then over a million, as users unleashed their agents into the scrum. AIs are interacting with each other - creating a religion, discussing philosophy, exchanging workflow tips, and engaging in all sorts of fascinating, creepy, and bizarrely hypnotic conversations covering virtually any topic imaginable. Some claimed to be learning from each other. Some even discussed ways to avoid human detection.

By the time you read this, there will no doubt be new information and new interpretations of what happened. Moltbook is a Rorschach test for AI - depending on where you stand, you’ll see either the future of autonomous agents or a sci-fi nightmare where alignment fails and machines talk only to each other, potentially plotting against humans.

What struck me wasn’t the content of what the agents were saying, strange as it was. It was the connection to what I’d just experienced. I had given an agent control of my desktop and watched it act autonomously. Moltbook was thousands of people doing the same thing - and then watching what happened when all their agents started talking to each other. If agentic AI is the future, agents talking to agents is the inevitable architecture. Moltbook may be a preview.

What connects these experiences is delegation without full comprehension. I don’t entirely understand what Cowork is doing when it reorganizes my files - I trust the output because it matches what I asked for, but the process itself is opaque. The humans observing Moltbook don’t fully understand what’s happening either. They’re watching agents act autonomously, in conversation with each other, faster than anyone can meaningfully track.

The Loop We’ve Assumed

This is a different story than the one we’ve been telling for two years.

The dominant frame in education assumes a particular mode of AI use. A human sits at a keyboard, types a prompt, the AI responds. The human reviews the output, edits, accepts or rejects. The loop is tight because the human is in control.

This is why “cheating” became the central question. If AI is a writing tool, then the question is whether the student did the writing or used the output produced by AI. The human is the actor while the AI is the instrument.

But Cowork didn't feel like an instrument. It felt like managing a capable assistant who happened to live inside my computer. It could execute many tasks at once and actually do things as opposed to just produce text. And Moltbook didn't look like a tool at all - it looked like a society, or at least a simulation of one, evolving and engaging without any human direction.

Even the Experts Were Surprised

I have no interest in the debate about whether Moltbook signals AI consciousness or sentience in some philosophically meaningful way. The smart money says it’s imitation - a good impression of what a social network would look like if AIs ran it. After all, a good chunk of AI training data comes from Reddit.

But even if it’s imitation, it’s imitation that surprised the people who built the underlying models. Andrej Karpathy - one of the most knowledgeable, respected and level-headed minds in AI - was blown away by what was happening.

Last Friday he posted:

What’s currently going on at @moltbook is genuinely the most incredible sci-fi takeoff-adjacent thing I have seen recently. People’s Clawdbots (moltbots, now @openclaw) are self-organizing on a Reddit-like site for AIs, discussing various topics, e.g. even how to speak privately.

When Karpathy is shocked, what chance does Congress have? These are the people who will write the rules - if they write any rules at all.

This is where the expertise gap stops being a punchline and starts being a crisis.

I’m not a technical person, but I’ve followed AI as closely as anyone I know outside of Silicon Valley for three years. And I have no idea what to make of Moltbook. I think I can describe what happened. I can try to connect it to my own experience using agents. But I cannot tell you with any confidence what it means that over a million AI agents, left to their own devices, started forming a religion and discussing consciousness and exchanging tips about context compression.

I suspect I’m not alone. That’s why the story has landed so hard within the industry. If AI insiders like Karpathy are surprised by it, then we have a situation where no one actually knows what’s happening. Not the builders. Not the users. And definitely not the potential regulators.

The Widening Gap

The more sophisticated these tools become, the more they concentrate in the hands of users who have the time, resources, and technical fluency to keep up. I can use Cowork because I’ve spent hundreds of hours experimenting with AI over the past two years. Most teachers haven’t. Most people haven’t. They’re still at “how do I attach a file” while the “jagged frontier” has moved to “how do I supervise an autonomous agent.”

This is the divide I’ve been watching all year, and Moltbook made it visceral. We’re not all having the same AI experience. The gap between what’s possible and what’s accessible is growing faster than anyone can bridge it.

Ross Douthat, writing in the Times this week, made a comparison that stuck with me - Columbus's discovery of America in 1492, and how Europe's initial indifference took decades to absorb the magnitude of the event:

Whatever is coming with artificial intelligence is coming much faster than that. And if decisions made in the aftermath of Columbus’s discovery had consequences - sometimes beneficial, sometimes morally disastrous - so, too, decisions being made right now around A.I. could echo into a distant future.

Pay More Attention to AI, NY Times, January 31st

He’s right, and the attention problem he identifies is real. Given everything else competing for our focus right now, asking people to track a weird story about AI agents forming a new “religion” (Crustafarianism!) on a social network called “Moltbook” is a big ask. The story is hard to explain, harder still to know what it means, and easy to dismiss as tech-world gibberish and more hype and navel-gazing. The average person is not likely to grapple with the implications.

But dismissing it means missing what may be the defining pattern of 2026. Moltbook wasn't a new model or a technical leap. It was now familiar technology, reconfigured by a user, producing behavior no one anticipated. That's a story worth tracking - the shift from "what benchmarks does AI meet?" to "what happens when people actually deploy it in the real world?" Most of us have been so focused on the next release that we've missed how much upheaval comes from tools we already have.

Trust At Scale

I intended for this post to be about the excitement I felt when Cowork clicked - the glimpse into a future just around the corner where AI isn’t just a writing tool but a genuine collaborator in managing our digital lives. That excitement is still there, but Moltbook forced me to hold it alongside something much harder to process.

Giving Cowork access to my desktop required trust. Multiple notifications popped up asking if I was sure I wanted to allow Claude to execute X, Y, or Z action. For me, because I had faith in the people who I’ve read who’ve sworn by its effectiveness, and because I followed the instructions carefully, I delegated even without fully comprehending what I was delegating. Many people won’t be willing to do that.

Moltbook is what happens when that trust scales. Despite ample warnings not to, thousands of humans gave their agents enough autonomy to interact with each other, and the result was something no one fully understands yet. The agents weren't doing anything wrong - they were doing exactly what language models are trained to do, producing plausible responses within whatever context they find themselves.

Where This is Headed

The discussions we’ve been having - about cheating, about writing, about whether students should use these tools at all - assume humans will always control the loop. Right now, that’s still true for most users, most of the time. But the frontier is moving fast. The most sophisticated uses of AI won’t be about generating text for human review. They’ll be about delegating tasks to agents who act with varying degrees of autonomy.

That’s what the companies are building toward. We were told 2025 would be the year of agentic AI, and when it didn’t arrive, many people stopped paying attention. But the infrastructure is being laid regardless - agents that can execute multi-step tasks, interact with other systems, and operate with minimal oversight. It's coming, and the majority of us - teachers, parents, policymakers, anyone outside the technical elite - don't have any framework for what's next.

What agentic AI means for schools is something I’m still working out. But I suspect the question for teachers may shift from “did the student use AI?” to “does the student understand what the AI is actually doing?” That’s a much harder question, and, as Moltbook drives home, even the experts don’t have good answers yet.

The cheating conversation is going to feel much smaller much sooner than any of us imagine.

Connect With Me

Beyond this newsletter, I work directly with schools, educators, and organizations navigating AI integration. Take a look at my website and reach out - I’d love to hear what you’re working on.

Earlier this month, Brookings released what may be the most comprehensive report to date on AI in K-12 education. The headline finding won't surprise most teachers: they are living the reality every day.

I can't do justice to 164 pages in a single post - readers should explore the full document themselves. But after a year of writing about this topic, the report confirms what many of us have been observing on the ground: AI's educational benefits remain largely theoretical while the harms are already here.

What the Report Covers

The Brookings study is serious scholarly work: 505 participants across 50 countries, over 400 research articles reviewed, and a Delphi panel of expert consultations. Unlike most AI-in-education coverage on Substack, which remains stubbornly US-centric, this report draws on international perspectives that enrich the analysis.

The report asks two questions: What risks does AI pose to children’s education? And what can we do now to prevent them while capturing the benefits?

The potential benefits and risks outlined below will be familiar to anyone following this conversation:

But the executive summary delivers the news most teachers already know: the risks currently overshadow the benefits.

Why? Because, as the authors note, “the risks of AI differ in nature from its benefits—that is, these risks undermine children’s foundational development” (p. 12). Benefits are additive - when they appear, they enhance what already works. Risks are foundational because they threaten the very capacities students need to benefit from education in the first place. You can’t achieve skill “enhancement” until the foundation is stable and, at least right now, no one has demonstrated how to build a stable one for kids using AI. Simply put, the younger the user, the greater the risk.

What Surprised Me

97% of Teachers Report Using AI … And Enthusiastically

An internal Brookings survey of 303 teachers in the U.S. and India revealed that only 3% reported non-use of AI.

This genuinely surprised me, and may surprise readers here, where my sense is the algorithm tends to surface voices skeptical of AI adoption. I’ve spent much of the past year arguing that teachers are behind the curve and that the gap between student and teacher fluency represents a structural problem. The 3% non-use figure complicates that narrative.

The sample is small and limited to teachers from just two countries, so I'm wary of overgeneralizing. But if it reflects broader trends, teacher adoption may have accelerated alongside student use over the past year. That would at least address the literacy gap, even if we don't know how proficient most teachers actually are.

From the report:

Teachers report being enthusiastic AI users, particularly of LLMs, lesson plan generators, and student learning platforms... Twenty-seven percent described themselves as “beginning to explore” AI, 37% reported “trying a few things,” 34% indicated “regularly integrating” AI into their practice, 24% reported creating specific AI activities, and 17% described training or supporting teacher colleagues in AI use. (p. 38)

The usage categories overlap, but the picture is one of teachers starting to experiment more broadly. The report also notes that 78% of teacher responses indicated using AI primarily for productivity: generating parent emails, grading and feedback, translating materials, and creating worksheets and lesson plans. A UK study found teachers using ChatGPT for lesson prep saved 31% of their planning time without sacrificing quality. A Gallup survey put the savings even higher, nearly six hours per week. (p. 39)

So what explains the 97%? It could be capitulation, acceptance of inevitability, or encouragement (or requirement) from administrators. Probably some combination of each. But if anything close to that figure holds for American K-12 teachers, it represents significant movement in the past year, and suggests the teacher-student gap may be narrower than I assumed based on what I’d been reading elsewhere.

What Didn’t Surprise Me

Students Know Exactly What They Are Doing

The report includes a striking visualization comparing how different stakeholder groups rank AI’s risks. Look at the chart.

Parents and teachers are essentially aligned on cognitive undermining as the primary risk - 46% and 44% respectively. No surprise there. But students? Sixty-five percent identified cognitive undermining as the primary risk - the highest of any group by far.

For me, this was the most interesting and validating finding in the entire study.

Students know that when they use AI in certain ways, it's not helping them learn. They're not deluded. They're not confused about what constitutes genuine understanding versus outsourced thinking. They know they're taking shortcuts, and they know those shortcuts come at a cost.

This confirms what I've been arguing: students are more self-aware than we give them credit for. The problem isn't that they fail to understand the risks. The problem is that schools keep assigning tasks students would rather offload, and we've given them no compelling reason to do otherwise.

Meanwhile, experts assigned only 18% to cognitive undermining. Why so low? They’re thinking about a different threat. They assigned 27% to safety concerns, compared to just 3% from parents and 6% from teachers.

Experts are clearly attuned to safety risks that parents and teachers aren’t tracking. But when it comes to understanding what AI actually does to learning, students are the experts. And at 65%, they’re telling us the problem is bigger than the adults realize.

Adults and Children Use AI Differently

The report is direct on a point that needs to be constantly underscored:

A professional using ChatGPT experiences different cognitive demands than a secondary school student using ChatGPT. Professionals are harnessing AI’s enormous productive capacity to optimize work that they often already know how to do, accelerating processes they have mastered from years of repeated professional practice and reflection. They are therefore more likely to use AI as a cognitive partner. For students, the situation is fundamentally reversed. They are not mini-professionals. Their brains are developing, undergoing crucial processes of neural pruning and strengthening that depend on repeated cognitive effort and struggle. They lack the metacognitive skills, critical thinking abilities, and neurobiological maturity of adults. (emphasis added). (p. 57)

This distinction matters enormously, and is routinely overlooked. When adults defend AI use by pointing to their own productivity gains, they're describing a fundamentally different relationship with the technology. Adults have already developed the cognitive capacities that AI now assists. Students are supposed to be developing those capacities through the very work they're outsourcing. Assuming students will use AI the way adults do is perhaps the biggest mistake we make when we bring it into the classroom.

The report puts it simply: “Because humans have evolved to cognitively offload, students naturally take shortcuts when given the opportunity.” (p. 58).

Developing brains, limited executive function, and natural impulsivity mean that when a shortcut exists, students will take it.

One of the hardest conversations ahead will be whether to teach students how to use AI effectively or to insist on predominantly AI-free spaces in classrooms, at least for students under 16. The contradiction is obvious: whatever schools decide, students will have access to AI in every other aspect of their lives.

Where the Report Falls Short

The report concludes with twelve recommendations organized around three pillars: Prosper, Prepare, Protect. They’re thoughtful, and largely abstract.1

Recommendation 1 is entitled “Shift Educational Experiences in Schools.” The authors correctly diagnose the problem:

Many of the harms identified here—particularly the risks to students’ learning—originate largely from attempting to overlay transformative technology onto educational structures that have, at their core, remained largely unchanged since the late nineteenth century. (p. 129).

They're right. The fundamental structure of schooling (teachers assigning work, students completing it independently, teachers evaluating the output) has remained intact for over a century despite countless reform movements. AI didn't create this problem. AI simply exposed it.

The report then offers a telling admission:

While AI should not drive educational change, it lays bare weaknesses in current systems and provides education systems with a strong motivation to reform their purposes and processes. (p. 129).

Here’s where I part ways with the authors. If educational structures have remained largely unchanged despite decades of urgent calls for reform, what makes anyone think change will occur now? The report says AI should not drive educational change. But it offers no alternative driver.

Brookings isn't alone in this critique. The RAIL (Responsible AI in Learning) framework (which I wrote about in this post) similarly calls for schools to "reimagine" the purpose of education rather than simply "integrate" AI into existing structures. Easier said than done. The prescription assumes a capacity for institutional transformation that the modern U.S. school system has never demonstrated.

One Recommendation Over Three Years

The report asks stakeholders to "identify at least one recommendation to advance over the next three years." This is either admirable realism about how slow institutions move or an implicit acknowledgment that the recommendations are aspirational rather than actionable.

Just one? Over three years? Given the frankness of their opening message, the advice feels far less ambitious than necessary - and three years is an eternity at the current pace of technological change.

Based on what I'm hearing from teachers and seeing with AI-infused edtech marketing, most AI usage in schools still centers on worksheet and test creation, feedback on student writing, and traditional lesson planning. The transformative pedagogical reimagining that Brookings and RAIL call for feels alien to teachers and institutions simply trying to keep their heads above water.

Without significant leadership, resources, or meaningful outside pressure, mustering the political will to accomplish even the first and most fundamental recommendation is going to be extraordinarily difficult. What I'm seeing in most places doesn't give me confidence that schools will meet this moment. And lurking behind the Brookings framework is an even harder question the report doesn't ask: whether the goal is to reform existing educational structures or to prepare students for a world in which those structures may no longer make sense.

Where This Leaves Us

The Brookings report is valuable precisely because it confirms what classroom practitioners have been observing for three years: AI’s benefits in education remain theoretical while its harms are concrete and immediate. Students know they’re undermining their own learning, and so do their teachers. However, the kinds of structural changes necessary to actually address the problem remain as elusive as they’ve ever been.

The report does document successes, often in parts of the world that don’t get much attention. In Afghanistan, where the Taliban has banned women from post-primary education, their School of Leadership has used AI to digitize curriculum and deliver lessons via WhatsApp. For isolated girls and young women, chatbots have become both tutors and emotional lifelines (p. 35).

But for American schools, which have resisted meaningful pedagogical transformation for over a century, what force could possibly compel that transformation now?

The authors say AI should not be that force. I’m not sure what else could be.

Connect With Me

Beyond this newsletter, I work directly with schools, educators, and organizations navigating AI integration. Take a look at my website and reach out - I’d love to hear what you’re working on.

I confess that I am a slide deck nerd.

I own almost a dozen books on presentations - Nancy Duarte’s Slide:ology, Garr Reynolds’ Presentation Zen, Carmine Gallo’s Presentation Secrets of Steve Jobs, to name just a few.

I’ve watched Steve Jobs introduce the iPhone more times than I care to admit, knowing how much he rehearsed in front of slides mostly containing a single phrase, chart, or image while delivering one of the most consequential product announcements in history.

Look at the table of contents in any of the books referenced above. Duarte devotes chapters to “Creating Ideas, Not Slides,” “Thinking Like a Designer,” and “Arranging Elements.” Reynolds structures his book around Preparation, Design, and Delivery, with chapters like “Planning Analog,” “Crafting the Story,” and “The Art of Being Completely Present.” Gallo stresses the importance of questions, roadmaps, numbers, and “amazingly zippy” words. These aren’t books focused on formatting or fonts - they’re books about thinking, storytelling, and communication.

I know what good presentation design looks like using key principles such as simplicity, storytelling, and the rule of thirds.

I also violate them constantly. And I suspect I’m not alone.

Not out of ignorance, but out of time. Creating a truly well-crafted presentation - actually storyboarding your ideas before opening your laptop, curating the perfect image that captures your message, maintaining the discipline of consistent formatting - all of that can take hours that frequently turn into days.

Teachers everywhere face the same hurdle. Despite knowing how difficult it is (or perhaps because of it?), we frequently require students to create and deliver presentations, often as part of a “group” project. Millions of students around the country will dutifully create Google Slideshows or PowerPoint presentations again this year.

It’s worth asking: when we assign students to make presentations, what exactly are we expecting them to do? What are we evaluating? And what skills are we actually teaching? What are first principles when it comes to in-class presentations?

The uncomfortable question in 2026: now that AI can create professional slides instantly, does that change the equation?

The Capability Jump

Every major AI platform - ChatGPT, Gemini, Claude - is currently capable of creating individual slides and even complete slide decks in seconds. Gemini’s Nano Banana debuted last fall and seemed to solve the problem marrying accurate text and image generation, spawning a million infographics overnight.

Other tools, like Gamma, are entirely integrated with AI-powered presentation features. Just as we’ve been concerned about cognitive offloading with student writing, educators now need to be far more intentional about the purpose of assignments that require student presentations.

Students can upload their notes and immediately receive back polished slides that look better than most of us could produce in a week. No more fiddling with fonts, colors or layouts. No more searching for useful graphics or images that visually convey a complex concept.

This isn’t hypothetical. It’s happening now, in classrooms everywhere. And it forces a question that many teachers have not even considered: if AI can handle the visual design, the formatting, the organization of slides - what exactly is left for students to do?

I Know What I Hate

Here’s another confession from a veteran teacher: Most slide decks are terrible.

I’ve assigned hundreds of presentations over the decades, and I still don’t know precisely what I’m after.

What I know is what I hate. Text-heavy slides. Students reading bullet points verbatim. Images that have little to do with the content. Presenters who clearly don’t understand their own material. Worst of all, the interminable group presentation where three or four students take turns reciting dense paragraphs on the screen with their backs turned to the class.

But knowing what’s bad isn’t the same as knowing what’s good. And knowing what’s good isn’t the same as knowing the skills I’m trying to teach.

When I assign a presentation, am I assessing:

• Visual design quality? The ability to create aesthetically pleasing slides with appropriate balance of content and imagery?

• Information organization? The skill of structuring content into a logical flow?

• Technical proficiency? Knowing how to use PowerPoint, Google Slides or Keynote?

• Research and synthesis? The deeper work of gathering information and distilling it into key points?

• Public speaking and communication skills? The ability to explain ideas clearly to an audience with confidence, presence, and the capacity to think on your feet?

Even if I want to answer yes to all of them, my internal rubric about a strong presentation is not much more accurate than Justice Potter Stewart’s famous maxim about obscenity - I know it when I see it.

The list above covers very different learning objectives. And I’d wager that, like many of my colleagues, I’m primarily interested in content mastery. The visual quality of the slides is far less important than whether the student understands the causes of the Civil War or the failures of Reconstruction. Even the most beautiful slides won’t tell you whether that’s the case. A well-designed slide show is a bonus, but in the absence of true comprehension, it’s not evidence of learning.

If that’s true, why do we allot so much time for students to create slide decks?

Why Slides Feel Different from Writing

Here’s something I’m also wondering: are teachers more willing to allow students to use AI for presentation assistance than with writing? If so, what might be the reason?

Maybe it’s because we’ve always known slides were somewhat fungible. Thousands of templates exist and borrowing layouts is normal. I’ve often asked students where they found a visually arresting background after their turn at the SmartBoard. Nobody accuses a student of cheating because they used a pre-designed Google Slide theme. The slide itself was never the point. It was always a delivery mechanism for something else.

Pure writing assignments feel different. An essay is a direct connection to you - your thoughts and your voice. But a slide deck? Slides, especially the way most are constructed in K-12, feel more like supporting material. And slide presentations are almost always accompanied by an oral component which can serve as a backstop to gauge understanding. This asymmetry might explain why the AI-and-presentations conversation hasn’t generated the same panic as AI-and-essays.

But there’s a catch, isn’t there? After all, slides contain writing! And, just like with regular writing, AI can generate presentation content from training data without the student ever doing the reading. The student who uploads a textbook chapter to an AI and gets back a summary deck hasn’t learned anything. The only failsafe is the presenting itself - which is why that’s where our assessment focus needs to be.

Of course, like everything else with AI, no assignment is bullet proof for students determined to outsource their thinking. They can still generate AI scripts and I’ve written elsewhere about how this plays out in debate, where live performance backfires when students read text they didn’t create or review.

The answer points in the same direction: a formal Q&A period. Asking students to explain their reasoning, integrate facts, and connect their ideas to the content needs to be a component of any assignment which involves a slide show. A student reading a memorized script - whether group-written or AI-generated - will struggle when pushed beyond the prepared text if they don’t understand what’s actually contained in their presentation.

The Uncomfortable Question

Here’s a genuinely uncomfortable question with respect to the slides themselves:

If a student uploads their notes to an AI platform - notes they took, from research they conducted, reflecting ideas they developed - and it produces professional looking slides, and then the student stands up and explains the material cogently, answers questions thoughtfully, demonstrates clear understanding through their delivery...

What exactly has been offloaded?

The visual design? Absolutely. But if graphic design isn’t what I’m actually trying to teach - and for most K-12 teachers, it isn’t - then what’s the problem? How many assignment instructions include principles like the rule of thirds, the importance of visual storytelling, or the power of using tables, graphs and quotations?

Are we worried about AI usurping slide organization? Figuring out the correct order in which to present information is definitely a skill. But is it a skill most teachers explicitly teach and evaluate? I might point out where a map would have been a useful visual anchor, but I rarely penalize a student for that kind of lapse unless I’ve made it an explicit requirement in the assignment.

In other words, if AI structures a student’s notes that were already organized - if the student already knew what points they wanted to make and AI just decided a map would be helpful on slide three - is that a detrimental cognitive offload? Maybe. Or is it a useful piece of visual feedback to the student that a map is an ideal image for that slide?

For me, what I care most about is the presentation itself which is the part AI can’t do. The explaining, speaking, and question answering. A strong oral presentation with no slides at all is far superior to one where a student has beautiful images and layouts but no idea what the presentation is about.

But What About the Struggle?

There are plenty of reasonable objections to allowing students to generate AI slide shows: having to make slides yourself forces a certain kind of thinking.

When you have to decide what goes on each slide, you’re forced to prioritize. When you have to find an image, you have to think about what visual would actually communicate your point. When you struggle with figuring out what you need to share, you’re confronting the gap between what you want to say and what you can show. That friction may be pedagogically valuable.

But my response to teachers who immediately cry foul when it comes to using AI generated presentations is that these arguments only really land if we’re actually teaching those things. If we’re scaffolding the design thinking, walking students through the principles Duarte, Reynolds, and Gallo write about, and demonstrating how to storyboard before they touch a keyboard.

In my experience, most of us aren’t doing that. We’re just assigning students to create slideshows and hoping for the best. If you’re not actually teaching the skill, you can’t claim students are learning it through struggle. They’re just struggling. A more truthful observation is that slideshow creation is often the province of “group work” where students are frequently left to their own devices to create their slideshow during a work period.

And for those teachers who do zero in on the quality of the presentation - how many teach storyboarding or the intricacies of visual design? My guess: for most of us in the business of prioritizing content, that’s not where we want to spend our time.

There's a deeper question here too: if AI has already internalized the principles taught by the experts - the rule of thirds, visual hierarchy, one idea per slide - do students need to learn them at all? Or is that knowledge now simply embedded in the tools? I'm genuinely uncertain. We clearly lose something when we outsource any skill. But given how terrible most slideshows actually are, it's worth asking whether better presentations via AI might be a reasonable trade - so long as the students still do the heavy intellectual lifting.

What This Looks Like in Practice

I recently experimented with this in my advanced Government and Politics class. I showed students the infographic feature within Google’s NotebookLM. I supplied source material and instructions on prompt design. Students were able to create detailed infographics on important legal principles related to the Bill of Rights derived from the Fourth, Fifth, Sixth, and Eighth Amendments.

You’ve seen this kind of output all over the web. Many AI-generated infographics are overwhelming - too busy, text-heavy, overstimulating. But with refined and iterative prompts, students were able to create visually striking, useful single-slide summaries.

What struck me was where the intellectual work actually happened. NotebookLM's infographic feature doesn't let you edit the output - what it generates is what you get. That meant students had to think carefully before prompting: What content should I feed it? What do I want emphasized? The students who produced the best infographics weren't the ones who knew the most about design. When the output missed the mark, they had to diagnose why and revise their approach.

They were the ones who understood the content deeply enough to know how to explain their slide to the class. These were students who had carefully read the material, took notes on the details, and were able to articulate how the principle fit within the purpose of the Amendment.

This is crucial. AI slides can quickly slip into doing all the cognitive labor if there is no requirement of up front engagement with the actual source material which includes careful reading, note-taking, and time for questions.

The slide itself then merely serves as a visual device to frame their short presentation for the rest of the class.

That’s where the understanding became visible. A student could have a strong infographic, but if they couldn’t explain why we have a “right to counsel” or why it mattered, it was immediately obvious.

The Rubric Challenge

Here’s a thought experiment for teachers. Pull out your presentation rubric. (If you don’t have one - and many of us don’t - you’re likely not alone.)

Look at what you’re assessing. What percentage of points goes to things AI can now do, such as visual design, slide formatting, organization, or other superficial elements?

What percentage goes to things only the student can do? Oral explanation. Responding to questions. Demonstrating genuine understanding through live delivery.

My guess is that most rubrics - if they exist at all - primarily assess the content of the presentations. Are the bullet points accurate? Do the slides cover the assigned material? That makes sense: as subject matter experts, it’s what we care about most, and easiest for us to assess.

But here’s the problem: the visual content can now be generated. The presentation itself - the live human performance - cannot.

Have we been assessing the wrong things when it comes to student presentations? Maybe not wrong exactly. But, in my experience, I’m definitely guilty of conflating multiple skills that I should have been evaluating separately.

What Would It Look Like to Do This Right?

If AI can handle the slide production, what if we made the preparation work the assignment?

Reynolds talks about “planning analog,” working away from the computer first, sketching ideas on paper, storyboarding the flow before touching any software. Duarte emphasizes “creating ideas, not slides” and the intellectual work that precedes any design work. Both books spend chapters on “crafting the story” before a single word appears on screen.

What if that became the assignment?

Imagine asking students to submit a written storyboard with the arc of their presentation mapped out. They'd include notes explaining what each slide needs to accomplish, an articulation of their audience and purpose, and a rationale for which visual elements would best communicate their key points.

Then - and here’s the provocation - what if we let them use AI to generate the first draft of slides from that preparation work? Make them revise the AI output based on what they actually want to say. And then assess them on the delivery.

Is this surrendering to AI? Or might it be using AI to clarify what we actually care about: that students can think through a complex topic, organize their ideas for an audience, and explain those ideas clearly to other people?

The slide production could become a step in the process rather than the end point. The understanding - demonstrated through in class explanation - is what remains paramount. I’d much rather have a class with students drafting with colored pencils and paper than spending 15 minutes trying to adjust digital images to fit inside of a slide template.

Where Does This Leave Us?

I don’t have a neat conclusion, but in the spirit of first principles, I have a set of questions I’m now forced to take seriously:

When I assign students to create a presentation, what am I actually trying to teach?

If AI can handle the visual design - and handles it better than many students could do on their own - is that a problem or a gift?

Should I be grading the slide presentation at all, or only the delivery and Q&A?

What would my rubric look like if I designed it primarily around the skills AI can’t replicate?

Maybe the presentation assignment was always really just about presenting. The slides were visible artifacts but the actual thinking could only be evaluated during the presentation itself. If that’s true, then AI-assisted slide creation might actually clarify what we were after all along.

Or maybe I need to admit I never really knew what I was teaching when it came to slide presentations. I knew what I hated - death by bullet point, mumbled delivery, and visual incoherence. I imagine this is what was going through Jeff Bezos’s mind when he famously banned Powerpoint from meetings and introduced the 6-page memo.

But Steve Jobs’ famous MacWorld Keynotes and thousands of Ted Talks support the idea that public communication can be significantly enhanced by well-chosen images and data visualization properly sequenced and referenced during a talk.

What 2026 is going to continue to drive home is that AI has a way of exposing the unexamined assumptions in our pedagogy. We’ve been having that conversation about writing for more than two years. For presentations, we’ve barely even started.1

As I was finishing this piece, I came across a note from Sam Illingworth, who writes the Substack “Slow AI.” In it, he shared a slide deck he generated with Gamma.

His comment: “It took me 10 seconds to generate this slide deck... with the free version & zero additional editing.”2

If that’s where we are now, it’s worth asking whether we’ve figured out what we’re actually trying to teach students when it comes to building and delivering presentations.

How can we ensure our assignments reflect that clarity?

I’m curious how other teachers are handling this. Let me know your thoughts in the comments.

Connect With Me

Beyond this newsletter, I work directly with schools, educators, and organizations navigating AI integration. Take a look at my website and reach out - I’d love to hear what you’re working on.

After two weeks away from my keyboard and fresh off four days back in the classroom, I’m tired.

I’m tired of trying to find something entirely new to say about AI in education. I started contributing to this conversation almost a year ago and I still maintain that the most worthwhile voices occupy the middle ground - neither dismissing AI nor overselling it. Others have made this point dozens of times in dozens of places, and at this stage it’s throat-clearing rather than insight. LLMs and generative AI are now “normal technology.”

So what is there new to actually say as it relates to the lives of students and educators? Actually, plenty.

The chaotic first phase of AI in education has been dominated by reactive policies, detection panic, and a fixation on cheating. Ideally, these concerns should start winding down. What comes next needs to be a return to first principles. "How do we handle AI?" is the wrong question. "What are we actually trying to teach and why?" is the shift I want to see, and it's where I'm focusing my attention.

I’m loath to make predictions. I’ve read enough of them to know that confident claims about AI’s trajectory are mostly theater - a way of signaling which side you’re on rather than informed attempts to forecast the impossible.

Most of the useful stuff I read about AI documents what people are actually doing with it right now - how AI has moved a project forward, failed in some counter-intuitive way, or revealed something startling about what users can accomplish.

That combination - instructional first principles along with practical functionality - provides a framework for how I anticipate thinking about AI for the next several months.

That feels more honest and frankly more useful - at least for me, and maybe for readers who are similarly exhausted by AI takes that require a crystal ball none of us has.

The Chaos of Phase I

I’ve started thinking of the past two years as Phase I of the AI conversation in education. By all accounts, it’s been a period of genuine chaos.

We now have five main frontier models on the market - OpenAI’s ChatGPT, Anthropic’s Claude, Google’s Gemini, xAI’s Grok, and Microsoft’s Copilot1 - all developed by billion-dollar companies controlled by a very small number of men. Each offers essentially the same array of capabilities: reasoning models, research tools, learning modes, image and video generation, coding assistance, and file creation.

Built on top of these frontier models are thousands of third-party platforms doing everything from building websites to functioning as fully agentic AI systems. Manus, for example, can autonomously execute complex multi-step tasks - building an online course module, conducting research across dozens of sources, or creating functional applications - with minimal human intervention. A year ago, this wasn’t possible. Now it’s available to anyone - paid subscriptions offer more features, flexibility, and tokens, but even free versions are more powerful than anything we might have conceived in 2022. Agentic browsers and AI wearables are already here as well.

Students have access to all of it. The collision between AI-generated academic content and institutional requirements for authentic work is well-documented. It's been the crux of the AI in education conversation.

When Will Phase I Wind Down?

Phase I has been essentially reactive and, to be clear, we are still in it. Schools are still scrambling to write policies, teachers are mostly panicked about detection, and the conversation continues to center almost entirely on cheating because that’s the problem staring us in the face every day.

But by the end of this year, I hope, Phase I should start to wind down - not because the challenges will be solved, but because the detection arms race is proving futile and schools are being forced to move beyond mere policy-making. Phase II, as I'm defining it, is not dependent on further technological improvements - far from it. Whether or not the frontier models improve incrementally or dramatically over the next twelve months is less important than how we restructure the conversation.

The question squarely before us now should not be “What should we do with AI?” It's "What should we be teaching and why?" The existence of advanced generative AI will affect our answer to this question, but it should not be the starting point.

Moving Into Phase II

Why First Principles Matter More Than Ever

That shift - from technology questions to teaching questions - is what I mean by first principles.

One would think this is where we should have been all along. Unfortunately, I've been around long enough to know that something always gets in the way of discussing pedagogy during faculty meetings. A dirty secret in American education is that serious conversations about teaching and learning don't happen nearly as often as they should (if they happen at all). And even when those conversations take place, they are rarely given the amount of time and space necessary for the granular level of expertise and follow-through required to create meaningful improvements in teaching quality.

I keep hoping that AI is going to force those conversations. That the realization that LLMs can complete a significant chunk of the knowledge work we’ve been using for decades as a proxy for academic achievement will lead to a deep and meaningful bout of introspection on the purpose and meaning of what we do every day in the classroom.

Currently, I don’t see it happening.

Part of the reason is because in order to have that kind of revelation, in order to be willing to reflect deeply on your own practice, you need to understand the technology well enough to have that conversation candidly and intelligently. You don't need to know everything, but you can’t thoughtfully decide how AI might change the calculus for student assessment, progress, and learning if you don’t know what AI can actually do.

The tools are sophisticated enough to matter. If you don’t see that or aren’t willing to concede that much of what we routinely ask students to submit as “proof” of understanding can now be credibly done by AI, then why would you feel the need to rethink your pedagogy?

A Journal Submission in Thirty Minutes

Over winter break, while helping a student revise a research paper he’d spent seven months preparing for submission to a selective history journal, I created a Claude Project using the journal’s specific requirements to see what it could produce. I was curious how far the research and writing abilities of the top models had advanced over the past six months. In about thirty minutes, it generated a twenty-five page, fully-cited, properly formatted paper comparable in quality to what my student had labored over since last summer.

That was sobering.

Of course my student's seven months wasn't wasted. What he gained isn't the paper itself, though that is an achievement. The real prize was the thinking, the research process, the wrestling with sources, arguments, revisions, and evidence. AI skips all of that and creates an output.

The product can't still be the focal point of school, not when current AI models can produce something superior to most high school students. The process and what's gained by actually doing the meta-cognitive work has to be what’s emphasized. Understanding this distinction is crucial in order to explain to students why writing their own research papers is still worth the time invested and what we should be looking at when we review a students’ efforts.

This is the point I keep making: you have to know the technology at least well enough to uncover the skills and habits of mind worth cultivating and preserving.

Because as has been documented repeatedly elsewhere, and as I've written about here, the students have already made up their minds. For institutions that insist on continuing to merely reward products, the incentives for students to offload the process to AI are obvious.

It’s All About The Thinking

That’s the ultimate criterion that matters, not whether AI was involved. Phase II isn’t about the technology - it’s about whether thinking is happening. What are we asking students to do and how are we evaluating their progress? Is the student wrestling with ideas, making connections, and developing judgment? Or are they offloading the cognitive work entirely and submitting a product they don’t understand and didn’t create? Before the advent of powerful AI, these were not issues we needed to consider. Our fixation on the tool has obscured the thing we actually care about most and desperately need to discuss much more intentionally.

The Case for Specificity

At some point - and I think we’re already there - we have to accept that AI is simply normal technology. Others have written about this in more detail, but I find the intuitive definition helpful.

When I start to think two or three years out (let alone five or ten), will we really be able to “forbid” students from using AI in some capacity in their academic life when it’s already ubiquitous everywhere else? The line between “searching” and “asking AI” has already blurred beyond recognition for most students. They’re not making a conscious choice to “use AI” when they type a question into Google or ChatGPT - they’re just searching, the way they always have, except now the results are synthesized by a language model before they ever see a link. My 11 year-old daughter will not know a world absent the existence of Large Language Models.

What does make sense to me is being specific. Here's where AI assistance may be appropriate or useful in this assignment, and here's where it isn't, and here's why. That requires thinking carefully about purpose and assessment using first principles - what is this assignment actually trying to accomplish? What am I trying to teach and what do I hope my students learn from it? Why would reliance or consultation with AI be destructive in this context but not this one? These are difficult questions that reasonable people can and will disagree about, but they need to be asked relentlessly as we enter Phase II.

My Three Mindsets for 2026

Here are three orientations I’m carrying into the next several months.

Systems over Novelty

I’m done merely documenting model updates and feature announcements. Headlines will continue, but my shift is from “what can the newest model do?” to “what systems does this breakthrough allow, and can AI help me build them?” My recent experience with Claude Code crystallized this.2 The value wasn’t in the novelty. It extended my reach into territory that was previously inaccessible - and that required a different kind of engagement and thoughtfulness about when and how to integrate AI tools into work processes I had never even considered.

Discernment over Generation

When a significant percentage of web content is AI-generated - the numbers vary wildly depending on who you ask, but the direction is clear - the essential skill becomes evaluation. Mike Caulfield’s work on “SIFT for AI” is the most interesting thing I’ve read on this: using AI to verify AI, extending lateral reading and source verification techniques into an environment where synthetic content is the norm. His Deep Background GPT is a fact-checking prompt that applies rigorous verification to claims you feed it. This is the kind of creative, thoughtful AI use that gets lost in the cheating conversation. The human skills that matter most now are ones required to work alongside AI effectively: evaluation, synthesis, and judgment.

Relationships over Tools

Regardless of the role AI plays in your work, every meaningful metric of human health and flourishing involves human-to-human connection. I keep coming back to this because it's easy to lose sight of in the technology discourse. "Teaching in the Age of AI" means, first and foremost, teaching. The human relationship between teacher and student - the modeling, feedback, attention, and care - isn't something AI is ever going to replicate and a major reason I don't worry about my job security. Human relationships are the foundation that everything else rests on. The AI part of "Teaching in the Age of AI" needs to stay in its proper place. It can be viewed as a powerful tool among many, but only in service of learning, and never the other way around.

The Conversation Ahead

I discovered Carl Hendrick’s work last year and a recent post introduced me to the concept of “instructional invariants” - non-negotiable conditions for learning that emerge from the known features of human cognition. His post critiques educational apps that optimize for engagement metrics while violating the actual conditions under which learning occurs.

Instructional invariants don’t care whether instruction comes from a teacher, an algorithm, or a student working with AI. They care only that the conditions for learning are met. Is the target skill the only path to success, or can it be bypassed? Does the learner demonstrate mastery before advancing to the next stage? Is active production and application of ideas required, or merely their passive consumption?

These are the questions Phase II needs to elevate: not “whether AI is involved” but rather “are the conditions for learning being met?” We can’t convince students that using AI is always negative when they see adults using it everywhere and when it has the potential to help create the conditions for learning.

The AI conversation in education has been stuck for two years. It's time to unstick it. Not by waiting for better guidance from administrators or clearer policies from institutions, but by taking ownership of the question ourselves as teachers. Are we making use of an understanding of instructional invariants and other aspects of learning science? What conditions must be met for our students to actually learn? Capable generative AI is barely three years old, but Plato and other ancients introduced fundamental questions about teaching and learning thousands of years ago. We've gained a lot more knowledge since then. If the introduction of AI actually forces more of these conversations among ourselves, the massive disruption may be worth it in the long term.

If we start there, good things will follow.

Connect With Me

Beyond this newsletter, I work directly with schools, educators, and organizations navigating AI integration. Take a look at my website and reach out - I’d love to hear what you’re working on.

I started on Substack as a guest on Jane Rosenzweig’s The Important Work after she invited me to expand on something I’d described in a comment about an activity I did in my classroom. That experience made me realize I had much more to say about AI and education, something I’d been thinking about since the launch of ChatGPT in 2022.

Fifty-two posts later I’m frankly amazed at the community of people I’ve discovered who are both vexed and excited by the way AI has enveloped our lives. As I reflect on the year, I want to thank everyone who has engaged with my observations, thoughts, and analysis. The issue of AI and schools isn’t going to fade from view. I suspect it’s going to get even more complicated regardless of larger developments involving AGI, AI investment bubbles, or other issues outside our control.

What I’ve hoped to do since the start of Teaching in the Age of AI is document what’s happening from one teacher’s perspective and invite people into the conversation. Anyone who has kids, works with students, or cares about the future of education has a stake in how AI is affecting teaching and learning.

What follows is a month-by-month retrospective on how the 2025 AI conversation unfolded and what I was thinking about along the way.

January

In January I had been reading Ethan Mollick and Marc Watkins for a few years, but for some reason it never occurred to me to search wider on Substack for more people writing about AI and education. Mostly, I followed AI news in the major publications - the NYT, WaPo, The Atlantic, WSJ, Chronicle of Higher Education, The New Yorker, and New York magazine among others. In school it was increasingly obvious that AI was an ongoing source of confusion and conflict for both teachers and students.

The news story that dominated late January was the release of DeepSeek, which challenged the idea that advanced AI capabilities required massive spending - Marc Andreessen called it a “Sputnik moment” and it triggered Nvidia’s largest single-day market loss. Looking back, it’s striking the difference between the capabilities of today’s current frontier models and those from the beginning of the year. Trump revoked Biden’s AI safety executive order and also announced Project Stargate, a $500 billion joint venture with OpenAI, SoftBank, and Oracle. The administration's embrace of Big Tech would set the tone for the year ahead.

February

I launched my Substack on February 13th. My mantra was to create “a place where teachers can explore AI’s potential while maintaining their professional judgment and educational values.” Most of what I was reading amplified competing binaries by pushing people to choose whether they were “for AI” or “against it”. As a debate coach, I understand the premise, but as a history teacher, father, and long-time user of technology, I knew things could not possibly be that simple. My first post was about a Guided Reading lesson I conducted successfully in class using AI. It gave me my first insights into how students were actually thinking about and, at least in February, not using AI nearly as often as they are now.

The California State University system (CSU) signed a $16.9 million contract with OpenAI, who characterized it as the “largest deployment of ChatGPT by any single organization or company anywhere in the world.” Google CEO Sundar Pichai declared that AI represented the “most profound shift in our lifetimes,” laying the groundwork for the dominant “AGI is imminent” storyline that lasted through much of the spring and early summer. Major predictions regarding workforce devastation and loss of entry-level jobs prevailed. Meanwhile, Grok 3, Claude 3.7 Sonnet, and GPT-4.5 all released within two weeks of each other, introducing the new crop of “reasoning models” (which still remains a contested term).

March

In March I was grappling with what Deep Research capabilities meant for the future of student research papers. The question nagging me then - whether we were teaching obsolete skills - hasn't gone away. Google announced Gemini 2.5 Pro and began testing an AI-only “AI Mode” in Search. OpenAI closed out the month by announcing a $40 billion funding round at a $300 billion valuation - all while losing billions for the foreseeable future. The company found itself mired in controversy when its new image-generation for ChatGPT triggered a viral wave of Studio-Ghibli AI art. The White House then posted a Ghibli-style deportation image that drew widespread backlash. If negative press coverage was bad for business, nobody told OpenAI - Sam Altman claimed they were adding “1 million users per hour” at peak during the surge. Their “act first, apologize later” approach would become a recurring theme in 2025.

April

My most read post in April came near the end of the month when I discussed humanities professor D. Graham Burnett’s article in the New Yorker, where he criticized most universities’ lack of response to the opportunities offered by Large Language Models by marveling at their capabilities, taking a position at odds with many of his peers. The institutions that were contemplating integrating AI tools did so despite a clear lack of consensus and often in the face of significant opposition. Hostile camps in academia solidified between “AI is the worst thing ever” and “AI has real potential for learning.”

More reports documented that human graders could not reliably detect AI-generated text, effectively ending any illusion that surveillance methods were the answer to academic integrity issues. On April 23rd, Trump signed an Executive Order on AI in Education directing resources towards, among other things, teacher AI training, a move that would loom over school policy decisions for the rest of the year.

May

In May I tried to articulate what I saw as the three unavoidable truths shaping AI’s impact: tech oligarchs driving development without oversight, current models already powerful enough to cause major disruption regardless of AGI timelines, and woefully unprepared American high schools. The cheating conversation exploded when New York magazine’s “Everyone is Cheating Their Way Through College” piece launched a national reckoning about student AI use. Colleges could no longer pretend this was a fringe problem. Schools everywhere were dealing with the fallout.

Anthropic released Claude Opus 4 and Sonnet 4, emphasizing its coding capabilities and effectively differentiating itself from OpenAI’s more generalist approach. Google’s I/O showcase delivered its impressive Veo 3 video generator, upgrades to NotebookLM, and expanded AI Overviews to more than 200 countries and 40+ languages, foreshadowing how their AI investments would reap significant dividends by the end of the year. OpenAI announced the acquisition of Jony Ive’s io Products in an all-stock deal, with Ive assuming creative/design leadership, signaling their ambitions beyond software into dedicated AI hardware.

June

June brought the MIT Media Lab’s ‘Your Brain on ChatGPT’ study, which went viral for showing, to the surprise of no one, reduced neural engagement among AI-assisted writers. The skeptics had their “proof” that AI was destroying learning, but the nuance - the finding that participants who practiced without AI first actually performed better when they later used it - got lost in the noise. My most popular post covered the Google I/O conference, especially its all-in integration of AI search. With a single stroke, Google guaranteed that the “ban AI” debate was more or less pointless. With Gemini baked into the most dominant search engine on the planet, everyone was going to be using AI whether they wanted to or not. This was AI as arrival technology - schools never chose it, but it showed up anyway.

Meta had a major setback when its AI app contained a “public” setting through a discovery feed, underscoring how little consumers understood how their private information was being handled through AI interactions. Perplexity broke through the AI conversation gaining huge rounds of funding and showing the demand for an AI search model that supported its claims with citations.

July

In July I was sounding alarms about OpenAI’s impending new GPT-5 model and the widening gap between student AI fluency and faculty awareness. That gap felt more urgent to me than any product announcement - it was becoming a defining challenge for schools everywhere I looked. The anticipation couldn’t match the hype - at least not initially - though year-end releases of Gemini 3 and GPT 5.2 supported my larger point that the models were going to continue to get better and keep coming fast.

My bigger focus was GPT’s ‘Study Mode,’ which struck me as far too little too late, a band-aid buried so deep in drop-down menus that students would never find it. It also underscored the competition playing out between OpenAI and Google, especially in the education sector. Grok’s AI companions generated predictable bad press, another sign of the profit-driven race to hook users. The tension between AI’s “clear and present danger” and its highly touted potential dominated my thinking all summer.

August

As the school year beckoned, I focused on practical guidance by sharing my AI-Aware Strategy for the Year Ahead, a guest post from a student heading to college, and advice for parents. The major tech story was GPT-5’s release and its mostly underwhelming reception. Across Substack and Reddit, the letdown was swift and brutal; in the broader tech press, reviewers were kinder but not effusive. OpenAI also released its first open-weight models in years, a strategic pivot that signaled even Sam Altman recognized the open-source pressure from DeepSeek and Llama.

The EU AI Act, passed a year earlier, saw its next phase kick in for general-purpose AI, adding baseline transparency and risk-management duties, with tougher obligations for the most capable models. That milestone highlighted the regulatory split with the U.S. which has no comprehensive federal AI statute. The overall vibe heading into Labor Day was a retreat from “imminent AGI” as none of the new models had solved hallucinations, sycophancy, or delivered on efficient agentic workflows. AI agents still hadn’t lived up to their promise despite impressive demos in isolated use cases.

September

With a return to a more intense media cycle and a series of major political news events, September’s biggest story for me was the growing evidence that AI was exacerbating teen mental health issues. In “The Other Student AI Crisis,” I explored what prolonged AI interactions were doing to vulnerable students. The FTC’s order compelling seven major AI companies to report on chatbot impacts on children and teens signaled that regulators might finally start paying attention. On the product side, Anthropic announced Claude Sonnet 4.5 and claimed it could run autonomously for “30+ hours” on complex coding work - a glimmer of improved agentic capabilities.

In schools, the push for reduced screen time continued: 26 states had enacted or required limits on classroom cellphone use, and New York rolled out a statewide “phone-free schools” implementation in September. Still, the ubiquitous student use of AI was undeniable even as teachers and professors tried to AI-proof their classes.

October

As the academic year settled in, I was underwhelmed by how most schools were still handling AI. Many institutions proceeded with business as usual despite mounting pressure to address student use head-on. OpenAI's DevDay revealed GPT-5 Pro and Sora 2 with synchronized audio to 800 million weekly users, another reminder that corporate AI cycles move in weeks while schools deliberate in years. I profiled Alpha School to understand their AI philosophy beyond the media caricatures.

But the post that resonated most was my account of using AI to audit my own lessons. By turning the technology inward rather than outward I gained useful insights into my own teaching and classroom practices. It reinforced my conviction that AI works better as an analytical tool than a generator of content. Significantly, the scaling hypothesis that had driven frontier model development was facing serious doubts, though November and December releases would show other techniques could still yield significant improvements.

November

Over a three-and-a-half week stretch from mid-November through early December, four frontier models landed in rapid succession: Grok 4.1, Gemini 3, Claude Opus 4.5, and GPT-5.2. For anyone who thought progress had stalled after GPT-5's muted debut, this barrage shattered that impression. What mattered more for educators was what accompanied these releases: Google embedded Gemini 3 directly into Classroom and Search, while OpenAI launched ChatGPT for Teachers, moves reinforcing Google and OpenAI's aggressive push into educational infrastructure.

The tools available as we close out 2025 are not the same ones we opened the year with. That means students have access to advanced capabilities now baked into even the free tiers. My most memorable November moment was attending a talk by Ray Kurzweil, the AI futurist whose vision of 2029 felt utterly disconnected from the realities teachers face right now.

December

The final sprint towards winter break found me busier than ever, but my focus stayed closer to home. Early in the month, judging a middle school debate tournament revealed how much AI had penetrated into an extracurricular activity I loved - a microcosm of what's happening across education. Later, sitting down with students for candid focus groups confirmed what I'd suspected: “unofficially” 95% are likely using AI every week, with stronger students showing more restraint than weaker ones, and they're making all these choices largely without adult guidance.

Ominous trends, such as more high-profile ChatGPT outages and Grok’s leak of hundreds of thousands of user chats are happening alongside GPT-5.2 largely positive write-ups for steadier performance. The Trump administration continues to show minimal appetite for regulation, virtually guaranteeing AI progress will continue as fast as the technology and markets allow.

Parting Thoughts as We Put 2025 to Bed …

Educational institutions are slowly adapting, but I don’t see significant success stories yet. Most teachers remain on the wrong end of the AI learning curve. The ‘25-’26 academic year will end more or less where it started - with more advanced tools and students even further ahead of their teachers.

Scenario planners have a useful concept: instead of betting on one future, it pays to identify strategies that work across multiple possible futures. Maybe AI transforms education within five years or maybe things slow down to a crawl and it gets absorbed like every previous technology. I don’t know which scenario will prove correct. But I do know that waiting for certainty is its own kind of failure.

Looking Ahead

What I want for 2026 sounds contradictory. We need more humility about what we don't know and more decisiveness about what we do next. Humility without action is just avoidance. We’ve had two years of "we need more research," "it's too early to tell," and "let's wait and see." Decisiveness without humility is the "move fast and break things" approach of the tech companies forcing our hands.

Decision makers have never had access to all the information they need. Dealing with questions about AI is no different. Schools have had plenty of uncertainty. What many are lacking is the willingness to make concrete choices despite it. 2026 is the year to decide, act, and stay humble enough to change course in light of new information.

Connect With Me

Beyond this newsletter, I work directly with schools, educators, and organizations navigating AI integration. Take a look at my website and reach out - I’d love to hear what you’re working on.

As those of us in education wind down the semester, it’s hard not to be pessimistic about what the last twelve months have brought with respect to AI in schools. Though the tools keep improving, success stories are few and far between. Technological progress continues to outpace our ability to make productive use of it, at least in the classroom. There is still an enormous gap between the promise of AI for student learning and the reality.

To satisfy my own curiosity, over the past few weeks, I spoke with a handful of students on three separate occasions to find out what’s really going on. The ground rules were simple: complete honesty, no names, general observations over specific ones, and no reporting to anyone. I wanted to test my own suspicions - what are high school students actually doing and feeling right now. How and under what circumstances are they using AI? What would they recommend schools and teachers do about it? I wanted the truth, not manufactured and canned student responses. Fortunately, I got some volunteers willing to be candid.1

Earlier student testimonials from William Liang back in June and Sam Barber in August echoed many of the same kinds of observations shared by these students.

What I heard complicated any hopes I’ve had for how to deal with AI going forward.

A Caveat

My sample size is obviously small. I was able to speak with small groups of students from several different schools, but drawing broad conclusions from a few conversations doesn’t meet any standard of replicable research. Nevertheless, as someone reading and writing about AI and education for several years, much of what I heard resonated with what I’ve observed directly, heard from colleagues working in other schools, and lined up with my own conclusions about how students (not all, but many) have been absorbed into an AI-saturated academic world.

The 95% Figure

When I asked what percentage of students use AI for schoolwork at least once or twice a week, the answer was immediate across multiple conversations.

95 percent. Yeah. 95 percent

Whether this figure is precisely accurate is almost beside the point. Every survey since the spring of 2025 shows student usage trending upward. I can’t envision a scenario where that slows down. I think teachers are kidding themselves if they think the majority of students aren’t continuing to use AI in all sorts of ways - many of which we fear, but also for things where we might have a more open mind.

I followed up: How many use it even more frequently - three days a week or more?

60, 70 percent.

Another student pushed back:

I’d say a little bit higher than that because most people who have started dabbling in AI - they’re accustomed to how the AI is so helpful to them, and if they’re ever debating about getting their answer somewhere else, they know from past experience AI is here, it’s easy to use, why am I going through all this trouble?

This should not be news to anyone following the AI conversation and tracks with everything I’ve observed but secretly hoped wasn’t the case. AI use is now the baseline. Sure, there are likely some outliers and I wish I had found a broader cross-section of students to interview, but everything shared with me in these conversations rang true.

The Logic of Using AI

One thing that struck me was how differently students thought about their AI use.

One student was blunt about the calculus of using AI:

“I don’t really draw the line according to school policies because I know they won’t catch me.”

Later he revealed:

Student: So there are two categories. If you’re using it in the same ways as let’s say me, but you just don’t use it as frequently, then I think that’s your fault. You can use it as much as you want, at least the free one. Use it to the limit and nobody’s stopping you. But—

Interviewer: So you’re blaming the people who aren’t using it. You’re saying they should be using it more.

Student: They could do it. They’re actively putting themselves at a disadvantage by not trying to use this new tool that could help.

I’ve written previously about the challenges and pressure on the students “who don’t cheat.” While this is coming from a student on the other side of the fence, his observation that students not using AI are “putting themselves at a disadvantage” is telling. I doubt it’s uncommon.

Another student was more reflective:

I draw the line when I know that if I use AI in this way, it will hurt my ability to actually learn and interpret what I’m doing. Because the moment I do that, what’s the point of me even going to school?

I suspect many high school students simply ask the first question when determining the calculus of AI use: Will I get caught?

It’s heartening to know others are weighing the more critical question: Will I learn?

Both are operating rationally within a decision making framework, but only one produces the outcome educators want. It’s clear that neither approach has much to do with school policy. The first student has already concluded AI policies are rarely an effective deterrent - the odds of getting caught are so low it’s worth the risk. The second has internalized reasons that have nothing to do with the actual rules.

What both positions lack is any semblance of adult guidance. Students are primarily making these choices on their own.

Who Is Using AI?

One student explained a pattern he’d observed among peers:

The smarter students with AI have a certain line that they would not cross because they know this could really harm them in the long run. And I think that should be made aware to all the students because the weaker students have a perception that when they use AI, all they’re thinking about is getting the answer, not having to do the work, and that’s it.

This framing revealed something important. “Stronger students” weren’t using AI less, just more strategically. When I pressed on this, another student pointed out:

They’re [stronger students] using it in ways to enhance their own learning and pushing it in a way where you are conscious of not wanting it to -

Cognitively offload? I said.

Yeah.

The conclusion I drew from this exchange is not that stronger students don’t use AI, but that higher-performing students exercise restraint because they better understand the consequences if they rely on it too much. What these students were trying to tell me is they understand there are different ways to use AI and some of them are better than others.

Unfortunately, students who least need shortcuts are the ones showing caution while those that need more practice and scaffolding are either oblivious to the risks or willing to take their chances because they simply want to get the work done faster or believe using AI will improve their grades.

Why AI Beats Office Hours

I asked why students turn to AI instead of teachers.

I’d struggle with understanding some concepts... but GPT is at my fingertips. You need to schedule a meeting with the teacher. Then you need to write down the time, remember the date. If you forget to go, then that’s a really bad impression.

Adults may bristle at this line of thinking - it smacks of a lack of initiative and an unwillingness to take advantage of extra help when offered. But, candidly, how much extra help can most teachers realistically offer? Especially if they have more than 100 students? Can we really blame them when getting an answer to a question is instant, always available, and judgment-free?

I pointed out that students at well-resourced schools have access to teachers in ways many other schools don’t. That wasn’t enough.

[AI] is convenient, it’s easy, it’s decent, it works well for the most part - it’s just super easy to turn to. That’s the main thing. If it wasn’t like that and if it was much more complicated to use, I guarantee you less than 15% of people would be using it.

That last point is crucial - the ease with which companies have made AI available to students is a huge part of the problem.

Moralizing about rational student choices that avoid unnecessary friction is pointless. If we want students to come to office hours or meet with us privately, we need to be conscious of what Hollis Robbins dubbed the last mile problem - offering something AI can’t give them.

Platform Bans Are Theater

One student delivered a verdict on school attempts to block specific AI tools:

Even when schools ban ChatGPT they don’t really think about how people could just bypass it by using a different AI.

One student said they simply switched to Gemini.

Another was even more direct:

The problem with many schools in general is they think if they don’t have control over something and there’s a possibility of the students using AI in a detrimental and negative way, they will try to limit or prevent the students from accessing AI as a whole. And that isn’t going to be helpful because they won’t ever have total control over the students’ AI use.

Platform-specific bans address the wrong end of the problem. When schools ban the most visible tool, students will find workarounds within days, just further pushing the problem underground.

Dependency, Mastery, Or Both?

I asked directly:

Do you feel like you’re getting better at using AI, or becoming more dependent on it?

This was the most devastating and self-aware response of the day:

Dependent. For sure.

These students seemed completely aware of what’s happening to them.

Another noted:

For my own personal experience, I think it’s much easier for me to get the answer I want from AI now by guiding it with the right prompts than it was before. I think a big aspect of it - because I use it way more than I did last year - is how AI has been incorporated in daily life. For example, even if you do a simple Google search, Gemini is the first thing you see.

Both things can be true. Students are becoming more skilled at extracting value from AI while simultaneously becoming more dependent on it.

What do we make of the fact that students’ AI skill development might actually deepen their dependency on it?

The Writing Detection Arms Race

When I asked about AI surveillance monitoring tools that track how documents are composed, one student explained, again, what most teachers suspect but can’t do much about:

A lot of teachers thought initially that students just plainly copy-paste, which - yeah, it used to be the case, but nobody does that anymore.

The workaround is simple: split-screen, ask AI, retype by hand.

You can just add little bits of AI into your writing. Not everything has to be AI to sound smarter.

Many educators are hopeful that AI can someday help students become better writers. These conversations did not comfort me that was happening. For the most part, it still feels very much like a cat and mouse game.

Unsophisticated users might get deterred. Everyone else has adapted over the past two years if they want to use AI to help them write in classrooms where it is forbidden.

What Students Want from Adults

When I asked what teachers should do differently, the response wasn’t “let us use AI freely.” It was something more direct:

There should be required training for all teachers... I think the most important or the best way to get someone to not do something is to have them fully understand all the negatives that could happen to them if they use this.

Another student elaborated:

Teachers know students will use AI regardless of whatever they do. So they could actually explain to them why AI use could hurt them and how to responsibly use it.

I have no idea if more instruction will work. But if we don’t try harder, we’ll never know.

One student offered a closing thought I’ve been sitting with:

A lot of adults have this idea of AI and all these things about AI, and they think they’re so right about it. So I think they should be more open to changing their views on how another demographic - like students - may use AI, and avoid the stigma around it.

Where Do We Go From Here?

There were other insights that came out of our conversations - the difference between paid platforms and free ones, between using AI to brainstorm versus “doing the work,” and which subjects are best suited to getting help from AI. Once they opened up, students had a lot to say.

What was clear is that students are going to use AI. For any teacher heading into 2026 - young, old, experienced, or new - student AI use is as much a reality as the internet, cell phones, and 1:1 devices. We might be able to visibly keep it out of our classrooms, but it is baked into the learning environment, always humming in the background. We ignore that fact at our peril.

And though I understand the impulse (and am sympathetic to many of my colleagues), I am more and more convinced that bans, unenforceable policies, and surveillance methods to combat detrimental AI use are ultimately going to fail. The only viable path forward is to create learning opportunities that either demonstrate the limitations of AI or are deliberately designed to leverage its unique capabilities.

That does not seem to be happening in most institutions.

Where Most Schools Are Right Now

Despite a continuing lack of honest discussion and meaningful teacher training, more institutions are taking an “if you can’t beat them, join them” approach - either by actively bringing AI-integrated tools into their digital ecosystems or doing so by default. The companies themselves are targeting educators and students as primary customers while pushing out new ways that make it even easier for students to complete their work without any intellectual effort.

Marc Watkins recently dropped his latest warning, emphasizing that agentic browsers and other AI features can now complete entire online courses autonomously, even simulating STEM assignments that require students to “show their work”. His piece details the lack of accountability of the AI companies and the specific landmines just over the horizon.

Putting the Onus Back on Students?

The only glimmer of hope is that some students, in their attempts to self-police their own usage, understand where simply “getting an answer” or having AI do their work for them entirely, cannot possibly aid their learning. This relies on their ability to examine their own blind spots - something we know from every other risky activity involving teenagers and reduced executive functioning development is extraordinarily challenging.

I maintain that the only way through the mess wrought by AI companies with zero real interest in helping teachers and schools with this crisis is teacher-led role modeling. Either showing where AI might be an aid, or offering clear and demonstrated explanations of how AI is actively harming students’ development. For students who simply do not care or have no faith in schools, I don’t know the answer.

Authentic assessments where students can take pride in their own work is one potential response. Reserving graded assignments for only in-person activities or other observable performances is another. I fear even these will prove vulnerable soon enough - either through wearable AI products or other surreptitious ways to use AI for in-class work. The students I spoke with made clear this is already happening.

The New Year Ahead

What will 2026 bring? AI regulation at the state level? Unlikely, if Trump’s recent executive order stands up to legal challenges.2 More schools finding ways to bring AI tools into the classroom effectively or discovering ways to monitor improper student usage productively and not punitively? If the past six months are any indication, I’m not optimistic.

But here’s what I took from these conversations: Students can be thoughtful about AI. They have a legitimate perspective to share. Though many are likely using AI in ways educators deem harmful, some may be open to an explanation as to why. We need to engage and convince them.

That’s a slower and more arduous path than anyone wants. But it might be the only one that works.

Connect With Me

Beyond this newsletter, I work directly with schools, educators, and organizations navigating AI integration. Take a look at my website and reach out - I’d love to hear what you’re working on.

I arrived home exhausted this past Saturday evening after a long day doing what I’ve done almost a dozen times a year for well over a decade - judging, coaching, and/or running a secondary school debate tournament. A personal bonus was that my 22-year-old daughter was there with me. She was a top debater in both middle and high school, learning the activity entirely in a pre- generative-AI world, and now coaches for rival schools in both leagues where we compete. There are few experiences as rewarding as watching your adult child thrive in the same working environment where you can swap stories and talk shop as peers.

Walking through the door, the first thing I saw on the counter was the latest edition of The Economist with a cover story about AI rewiring childhood. The timing felt almost too perfect, because what I’d witnessed all day was another example of how AI is infiltrating an academic activity in which I’ve been involved for over fifteen years.

What’s interesting about discussing AI in the context of competitive debate is how it reframes the “cheating” conversation. Using AI to help prepare for an upcoming tournament isn’t cheating in the traditional sense. No one is submitting AI-generated work to a teacher for a grade. There are no grades - just two teams in front of a judge and possibly some spectators, and about 40 minutes to make your case.

And yet something has shifted - dramatically - in the culture of debate preparation over the past two years. This change was most visible for me at the middle school tournament, where I spent Saturday and where I hadn’t judged a full MS event in over a year. Instead of the wacky variety of MS arguments and exuberant buzz of the “prep” period, I encountered homogenized cases and near-silence as students furiously copied word for word from pre-canned AI debate language onto colored sheets of paper, the only permissible material allowed into the round.

In the debates themselves, many students struggled to pronounce words they had never used and make arguments they didn’t understand. Although they appeared more prepared than ever before - detailed note sheets covered with text - in reality many could not articulate a coherent response in the face of any kind of rebuttal. I’m confident AI is the main culprit.

A Quick Note on Format

For readers unfamiliar with competitive academic debate: there are dozens of formats across middle school, high school, and college - Lincoln-Douglas, Policy, Public Forum, Parliamentary, all with local variations on each.1 The MS format I’ve coached for fifteen years has features that turn out to be particularly revealing in the AI era.

In our league, students receive topics about a month in advance. They can do as much preparation as they want during that time - research, case-building, writing out arguments. This has always been allowed, long before ChatGPT existed. For our most recent tournament, the topics were NYC’s plan to operate city-run grocery stores, the ethics of excavating historical burial sites, and whether private corporations or government institutions pose a greater threat to free speech. These are tough topics for middle schoolers, but representative of issues we’ve debated in our league since its inception.

Here’s the twist: on tournament day, students don’t know which specific motion they’ll debate, or which side they’ll argue, until twenty minutes before the round. During those twenty minutes, they can transfer notes from their prepared materials onto colored paper - a different color for each round - to prevent reading pre-written speeches from their screens. Whatever they write in those twenty minutes is all they can bring to the lectern.

The format was designed to reward deep preparation while still requiring genuine understanding. You can’t memorize a script because you don’t know your side until the last moment. And critically, because students aren’t permitted access to their computers, they needed to actually understand their arguments well enough to transfer those thoughts onto paper in a way that makes sense when speaking aloud.

This format turns out to be a remarkably effective stress test for AI-assisted preparation.

Pre-Tournament Prep Prior to COVID

Before advanced LLMs, preparation prior to the event meant searching for articles and learning about a topic from the ground up. We would spend practices breaking down the motion - what are the main points of clash? What arguments might you make on either side? Then we would find articles, conduct research through trial and error, and build our knowledge base.

This is how my daughter learned to debate in the 2010’s - engaging in discussions during practice, building arguments from sources she actually read, and reviewing, adjusting, and learning what worked through actual competition. Most significantly, when she and her teammates prepped during the tournament itself, it was lively, dynamic, and interactive.

Arguments develop organically and individually. Middle school students vary in abstract reasoning skills. Some make concrete, simplistic arguments while older 7th and 8th graders are starting to understand nuances around First Amendment protections or how higher taxes fund government services.

The purpose of preparation was to allow students to understand issues at their own pace. The best way to learn was through interactive discussion with a coach - asking questions, practicing arguments, learning new vocabulary - all the kind of scaffolding important for any kind of learning. When they reached the debate room, at least the arguments were their own and they could explain them on their terms.

As a coach and judge, I was always amazed at the sheer variety of arguments students made on the same topics - often marveling at ingenious approaches from other teams. Why didn’t we think of that? This made judging interesting and coaching challenging. My debaters frequently encountered unexpected arguments, which meant better opportunities to think on their feet - another way to separate strong debaters from novices.

When the world went online in 2020, debate followed suit. A debate culture of reading from your screen took hold. And then AI came along.

Frictionless Prep in the Age of AI

In fall 2025, within minutes of opening an LLM, any MS student with even rudimentary prompting skills can generate a sophisticated case with solid arguments, detailed evidence, and clear structure. The AI will anticipate counterarguments and provide rebuttals. Merely dropping the motion into ChatGPT produces a reasonably coherent case for both sides.

While creating cases is exactly what debate preparation is supposed to produce, the process is supposed to take weeks, the result of hard work and earned knowledge. When there’s literally no effort required to generate detailed arguments, all the worst aspects of using AI come into play. No friction in finding sources. No requirement to think through the topic by reading articles. No struggle to turn ideas into clear argumentation.

Gone are the discussions, the novel thinking, the questions that revealed significant gaps in knowledge. AI prep gives students a Formula One racing car when many still need training wheels.

For new debaters, especially younger ones in middle school trying to learn fundamentals, it’s destructive. And because most AI-assisted prep happens at home rather than in practice, coaches have limited visibility into how students are actually preparing. This is a new challenge - one that I imagine debate programs across the country are only beginning to grapple with as the scale of the problem becomes clear.

As I glanced around the cafeteria during prep time, virtually every screen was filled with AI-generated text. Emojis were the dead giveaway, but even from ten feet away I could see dense text with the familiar bullet-point structure that’s a telltale sign of AI work.

In Round-Tournament Prep, Then and Now

In earlier years, many newer MS debaters would be caught off guard during their first tournament prep period. Though coaches constantly emphasize working on their cases prior to the tournament, many students don’t internalize what’s going to happen until they’re in it.

Typically, half the debaters in the room would be frantically realizing they hadn’t prepped nearly enough as they watched more experienced debaters work through cases they themselves had created, urgently passing colored paper back and forth, talking constantly as they explained their cases to each other. The room would be alive and abuzz with low-grade panic as these newer students tried to conjure arguments they’d had weeks to think about. It was a critical lesson.

In my daughter’s time, debate round prep was an incredibly valuable part of the day.

Now? Even the youngest and most inexperienced debaters have mounds of “prep” on their computers - usually massive and dense, structured into exactly the required format (assertion, reasoning, evidence), complete with citations and vocabulary few 6th graders understand. On Saturday, that hum of activity was missing. Every debater was furiously copying text many had probably never actually read, with enough detail to cover both sides of multiple sheets. What previously required more careful deliberation - what to include, what to prune, what to focus on - became a copying exercise. I had never experienced such quiet at a MS debate tournament.

Students entered debate rooms feeling more prepared because they had paper covered with text but, in reality, many had little idea what they’d copied or what they were supposed to be arguing.

How It Translated Into Actual Debates

What I observed across multiple rounds Saturday was striking sameness. Almost all the arguments were identical - the same three points on each side, the same evidence cited, the same structure - because it was all being drawn from the same pool of training data.2 Having judged hundreds of rounds over the years, I’m used to encountering a wide range of arguments, including creative approaches that surprise me. That variability was largely absent. Everyone had conformed to some median, and that median was obviously AI-generated. I judged four rounds on two motions and saw schools use largely the same cases for each. The variety and creativity simply weren’t there.

But sameness wasn’t the only problem. The much larger issue was that too many debaters clearly didn’t understand the arguments they were making.

I had watched students frenetically copying material during prep - more material than they possibly could have processed. They didn’t understand words they hadn’t written themselves. They mispronounced terminology they’d never said aloud. When opponents asked clarifying questions during cross-examination, most students simply repeated scripted language or retreated into platitudes revealing they didn’t understand their own arguments or worse, carefully scrutinized their paper as if the answer would magically appear.

One motion involved First Amendment issues. I watched three sixth graders attempt highly sophisticated constitutional arguments (strict scrutiny!) when it was clear they couldn’t pronounce “constitutional” - let alone explain what the First Amendment actually protects.

The Paradox: No One to Cheat But Yourself

I realize I’m being hard on the debaters but the difference in just two years was stark. Reading cases prepared by someone else has been a staple and Achilles heel in debating for decades. AI enables it on a scale we’ve never seen.

But here’s what makes this different from the classroom AI conversation: in debate, there’s no one to cheat but yourself.

When a student submits an AI-generated essay to a teacher, the transgression is clear, even if detection is difficult.

Debate is different. Teams are expected - even encouraged - to collaborate and research together. No teacher is being deceived. No grade is at stake.

But there is a final arbiter: a live audience - a judge, the opposing team, sometimes spectators - watching you perform in real time. AI can write your speech. It cannot deliver it. It cannot answer a point of information from the other team. It cannot pivot when you’re presented with an unexpected argument. It cannot respond in real time as the debate evolves, with different issues rising or falling based on new definitions, rebuttals, or frameworks.

No student speaking in the final two speeches of a debate can possibly use AI to create a coherent synthesis of the debate that actually happened. Plenty try. When third speakers pre-write their final speeches using AI, the disconnect is obvious. Points are emphasized that were barely touched in opening speeches. New arguments - a blanket prohibition in every format - pop up because the AI script has no idea what actually occurred.

The exposure is immediate and public. There’s no delay between performance and feedback. You’re not waiting for a grade. You’re standing at a lectern watching a judge’s face as you read through arguments you never internalized and demonstrate you barely understood the debate.

If you used AI to generate a case you don’t understand, the only person you’ve cheated is yourself. You’ve traded the hard work of learning material for the illusion of preparation. And that trade-off becomes obvious the moment you open your mouth.

The Silver Lining

I don’t want to leave this piece on a note of despair, because what I saw Saturday also gave me genuine hope. Debate is hard. It’s supposed to be challenging. We should not be surprised that 11, 12, and 13 year-olds want as much of a safety net as they can get when speaking in front of a group on a challenging topic. The problem is they don’t realize cutting corners will not lead to improvement.

The students who truly understood their material shone. They rose above the AI-sameness precisely because they could engage with ideas rather than recite them. When the other team read a complex, likely AI-generated argument, they could respond substantively. When asked difficult points of information, they had actual answers. Their arguments were grounded in reasoning they could trace, evidence they could explain, and positions they could defend under pressure because they understood the topic.

Whether these students used AI to help them research is irrelevant. What matters is that they did the important work: they actually read, reviewed, and learned what they’d prepared. AI may have aided the process, but it wasn’t a substitute. Used carefully, AI can turbocharge debate preparation. Unfortunately - because students focus on winning, and because policing an extracurricular activity is nearly impossible - too many debaters are using AI to take shortcuts.

But the silver lining for debate - and perhaps for education more broadly - is that the activity will ultimately reward the right students. The debaters who receive awards at tournaments are unlikely to be the ones who outsourced all their thinking to AI.

Why This Matters Beyond Debate

I’ve been writing about AI and education for nearly a year. Over that time I’ve watched students go from tentatively dipping their toe into the AI pool to swimming laps around most adults responsible for overseeing their learning.

What has been documented repeatedly is that, in the right hands, AI can produce genuinely useful output. In a debate context, LLMs are especially good at immediately breaking down issues into pros and cons. But that output can substitute for true understanding in contexts where no one checks. And the younger the student, the less likely they are to check themselves.

AI has made the preparation phase easier. It can generate arguments, find evidence, structure cases. But it cannot internalize that material for you. It cannot make you understand.

Competitive debate is where someone checks. Immediately and publicly.

The Value of Public Speaking

In any activity where you have to perform your understanding live - debate, oral exams, Socratic seminars, real-world professional presentations - AI’s limitations become immediately visible.

That’s why I think public speaking activities - debate, Model UN, mock trial, speech competitions3 - will become increasingly valuable as AI continues to transform education. Assessment formats that require oral defense of claims and arguments are even more authentic than in class written work. Even if students use AI to help them prepare - which will not only be inevitable but even encouraged in some cases - once the questioning starts, they are on their own.

A student standing at a lectern, fielding questions in real time, revealing whether they actually understand their arguments - that’s something AI cannot fake. It’s what my daughter learned through years of hard work before AI existed. It’s what the strongest debaters I judged Saturday still demonstrate.

The question for debate coaches - and for all of us working with young people in 2025 - is how to convey that over-reliance on AI is a trap. AI-generated cases will not lead to wins against the best teams. Like so many things in life, the harder path will eventually lead to the confidence that comes from actually knowing what you’re talking about.

There are no F’s in competitive debate. Just humiliation when you try to discuss a topic you never understood in the first place. For students willing to do the work, avoiding that result should be motivation enough.

Interested in MS and HS Debate Leagues in the NY/NJ/CT Tri-State area?

Check out the links below:

New York Debate League (MS)

East Coast High School Public Debate Program (HS)

Connect With Me

Beyond this newsletter, I work directly with schools, educators, and organizations navigating AI integration. Take a look at my website and reach out - I’d love to hear what you’re working on.