The marshmallow test was devised in 1970 to better understand when children developed the ability to “delay gratification.”1 The original experiment was straightforward - children were given the option of an immediate reward (typically a marshmallow) or two smaller rewards if they waited for a period of time and did not eat the treat. A researcher would leave the child alone for about 15 minutes and if they came back and he or she had not eaten it, they would be given their additional rewards. I remember watching videos of children squirming in their seat eyeing the marshmallow and desperately trying not to eat it when I learned about this test in a psychology class in college.
Though the experiment has since failed to be replicated and its conclusions have been challenged in other ways, for years it was used to buttress the claim that children who were able to wait longer for additional rewards actually had better life outcomes than those who opted for instant consumption. In other words, those children who showed greater self-restraint and “delayed gratification” were better equipped to navigate a world in which planning for the future was more productive than securing immediate gains.
One can see the intuitive appeal of such a theory.
I often think of the marshmallow test when considering how AI has found its way into high school classrooms. Teachers are rightfully concerned that students will simply not be able to help themselves when confronted with challenging work - whether in the form of science or math problems, or as in my discipline, expository writing - and will instinctively reach for chatbots to either guide them through or completely finish an assignment.
This “cognitive off-loading” is not just a theoretical concern but a real threat. With dozens of free AI LLMs available and a slew of AI writing products marketed towards students, all with no training or guardrails, many are turning to AI before they even begin to engage with or understand what they are being asked to do. I can easily imagine how I would have reacted as a 14 or 15-year old adolescent, still trying to master more advanced reading and writing assignments, if I were given a tool that could complete all my homework and essays instantly.
Intentionally introducing the power of these tools - and they are incredibly powerful - into a classroom setting is fraught with risks.
The AI Balancing Act
I’ve tentatively experimented with AI in a variety of ways in my classes. I’ve allowed students to use it for brainstorming, to conduct research, to analyze texts (both secondary and primary sources) and images, and to produce outlines, study guides, and other supplementary materials. Recently, I’ve been demonstrating the newer Deep Research models2 both as a way to illustrate what these newer tools are capable of as well as to serve as a cautionary tale about why having AI generate an essay and submitting it is the most blatant example of academic dishonesty one could imagine. How AI fits - or does not fit - into the writing process of novice writers is going to be the trillion dollar educational question of our time given that these tools are only going to get more and more advanced.
To deny that AI tools might be beneficial for feedback on one’s writing is simply ignoring the reality. But, while there may be ways to utilize AI productively while still maintaining control over your own writing process, it takes significant cognitive effort, skilled deployment of surgical AI prompts, and a willingness to roll up your sleeves and stay focused and on task. I’m not convinced that the majority of teenage brains are capable of using AI without resorting to serious cognitive shortcuts.
The Evolution of My Own AI Use
My AI use has evolved considerably since I’ve incorporated it into my work routine. It has undeniably had a positive impact on my ability to accomplish both academic and personal projects in novel, creative, and productive ways. It has empowered my writing, improved my output, and freed me from past inhibitions - primarily by serving as both an interrogator of my ideas and a final editor for my work. The ability to have a perpetual, patient, and almost infinitely knowledgeable writing partner willing to answer any questions, review rough passages, and offer suggestions on complex syntax and grammar is something unprecedented in my lifetime.
But I have seen in my own case the importance of doing the hard work first.
The Case for Cognitive Struggle
What do I mean by the “hard work”? This is the heart of the AI conundrum - the ease with which AI can generate readable prose in response to even a vague or poorly written prompt makes the temptation to offload the demands of the initial “thinking” process incredibly high. As a result, I have learned to leverage AI to assist with the creation of more and more detailed and finely tuned prompts that allow me to pinpoint precisely how and in what ways I would like AI to give feedback on my writing, but only after I have done the initial heavy lifting. It means I have to write the rough draft. I have to work through my ideas and try to get them on the screen. I need to struggle.
While this has been obvious to most AI critics from the beginning, the core insight I’ve reached, especially over the past 6 months as the quality of AI models has gotten better and better, is how essential it is to write independently before bringing in AI.3 In other words, using AI up front, at least with respect to producing my own written work, is liable to derail the process in significant ways. Half of the battle with creating a piece of writing is muddling through the dead ends, false starts, and incomplete thoughts. The temptation to reach out to AI, however, just like the temptation to go for that marshmallow, can be so overwhelming that I don’t know how many students are able to resist it.4
The Inevitable Reality
Students are going to use AI no matter how many bans we put in place or how much we threaten them with violations of academic integrity. This is our new reality and denying it won’t make it go away. AI is here to stay.
So how do we impress upon them the importance of “doing the hard work” first, especially when AI offers a tantalizing promise to save time, get stuff done, and alleviate the anxiety of the blank page? Even if students convince themselves that whatever they produce with AI will be reviewed, edited, or ultimately discarded, the reality is that as soon as AI text begins to fill the page, students have a plausible piece of writing from which to begin.
Many educators and writing coaches advocate having AI generate rough drafts as a starting point. But this approach undermines the very cognitive processes that develop strong writing skills. When a student begins with AI-generated text, they're starting with someone else's thinking structure rather than tapping into their own. The initial challenge of the blank page isn't just a writer’s myth - it's a necessary developmental stage that strengthens the neural pathways essential for critical thinking and original expression. Sifting through word choices, weighing competing ideas, and ultimately deciding what to put on the page literally is writing.
The problem with many AI first-draft exercises, as I’ve seen them described, is that the punch line is usually how “wooden” and “soulless” the quality of the AI writing appears. The class feels superior as AI text is held up for all to mock, the emperor with no clothes. I fear this will only be the case for so long - what happens when a “first draft” of AI writing is as good as or better than the average person and can maintain a voice and a tone and a point of view? Many would argue that is already the case. But an AI first draft, no matter how well prompted, will not involve the necessary friction required to fully reflect the intention of the writer, regardless of how well it’s “written." It cannot and should not be a starting point for novice writers.
The Future of Writing
Despite my concerns about students, I’ve also arrived at a contradictory conclusion: for professional writing, we will soon reach a stage where the difference between human writing and AI writing will not only be undetectable, but ultimately irrelevant. If paid writers want to generate first drafts with AI, as many are currently doing, they have earned the right to their own process. Written communication thrives or dies based on readership and consumption. If the public wants to read AI-generated fiction or embraces human-AI hybrid writing with their wallets, the market will support it, regardless of the fury and gnashing of teeth from AI skeptics. In many sectors, we’ve already crossed this threshold.
The Takeaway for Students and Teachers
But all bets are off for beginning writers. They need to struggle with that blank page. The cognitive development that comes from wrestling with one's own thoughts cannot be bypassed or outsourced without developmental consequences. Let the professionals or, at the very least, those who feel some level of confidence in their writing ability experiment (or not) with AI. There is a lot of writing in many different fields where the goal is to transmit information accurately and clearly to a variety of audiences - AI excels at these tasks.
But to students at almost every level, I urge caution: you need to write the rough draft. Polish it until you think it's perfect. Proofread and polish it again. Read it out loud. Finalize the project. Only then, perhaps consider whether AI might offer valuable feedback or help you refine specific sections.
Do the hard work first.
For educators, our challenge is to design assignments that make the hard work both necessary and meaningful. We need to create contexts where students can see the value of their cognitive struggle, where the process of thinking through writing is as rewarding as the finished product. This might mean more in-class writing, a more scaffolded drafting process, or an assessment model that values originality and personal voice over a perfect paper.
Maybe there is a place for AI after that. We'll see.
I never liked marshmallows anyway.
I am not as adamant with this rule for tasks that may involve open-ended discussion, brainstorming, or research. In those cases, I frequently turn to AI for a variety of reasons in the beginning, middle, and end stages. In this post, I am really drilling down on the writing process itself.
We know from recent studies that more and more them are failing the test.