Hollywood is losing writers to AI. That is a labor-market story. It will reorganize an industry, and the industry will survive. Set that story next to a quieter one unfolding in classrooms at the same time: a generation of students is outsourcing to AI the one thing education was supposed to build in them, the ability to think without help. The two stories look like the same story. They are not. One is a sector reorganizing. The other is a generation hollowing out.
In June 2025, researchers at the MIT Media Lab wired 54 college students to EEG caps and asked them to write essays over four months in three ways: with ChatGPT, with a search engine, and with nothing but their own head. The scanners told a clear story.
A signal students once calibrated against is quietly going dead.
Data: Harvard registrar figures, as reported by NPR, November 2025.
The ChatGPT group showed the weakest connectivity across the brain networks that handle memory, attention, and executive function. They had the hardest time remembering what they had just written. More importantly, when they later tried to write without the tool, the disengagement did not immediately reverse. The authors gave the state a name: cognitive debt.
The word "debt" is the most honest word in the whole study. Debt means the bill is not now. It is later.
A separate study, published in Frontiers in Psychology in July 2026 by Cui and Jalal, found a psychological mechanism underneath that debt. Students who leaned on AI to produce their coursework gradually came to attribute their success to the tool rather than to themselves. The researchers named this the illusion of competence. In their model, that illusion sits upstream of something older that psychology has studied since 1967: learned helplessness.
Martin Seligman named learned helplessness after an experiment with dogs who, having learned that nothing they did stopped a painful shock, stopped trying even when escape was later possible. The mechanism in the AI classroom is softer, but the shape is the same. If a student does not experience themselves as the agent of their own success, they stop trying to be. Luminis would add one line here: the illusion of competence is more dangerous than a bad grade. A bad grade is a signal; a student can read it and ask for help. The illusion of competence is a silent credit card, and the bill often does not arrive until the first job interview, or the first open-ended problem no AI has a shape for.
AI dependency erodes intrinsic motivation by way of learned helplessness. AI literacy is a moderator that weakens the first arrow, which is the empirical case for "teach literacy before you give access."
Model: Cui and Jalal, Frontiers in Psychology, July 2026.
Grades themselves are a signal, and the signal is quietly going dead. At Harvard, A-range grades now account for 85% of all letter grades, up from 62% twelve years ago. Yale, Emory, and most of the elite tier show the same arc. Nationally, A grades make up over 60% of all college grades, a 25% rise in twenty years. The point is not that schools have become too soft. The point is that the signal students once used to calibrate against, "I got an A on this; I actually learned something," has decoupled from what the student can actually do.
Pair that with the Cui and Jalal mechanism and the trap closes. The output looks fine. The student does not know what their effort was actually worth. There is nothing left to calibrate against.
A third strand, from a 2025 Microsoft Research and Carnegie Mellon study of 319 knowledge workers across 936 real AI uses, maps the same pattern into adult work. The researchers found an inversion worth sitting with: the more confidence a worker had in generative AI, the less critical thinking they applied; the more confidence a worker had in themselves, the more critical thinking they applied. The observed shift is from "doing tasks" to "supervising tasks," what the education literature calls cognitive offloading.
Supervision without the underlying work is brittle. A teacher who has never done long division cannot catch a student's long-division mistake. An engineer who has never debugged a race condition cannot catch an AI's confident, wrong answer about one. The ability to supervise is downstream of the ability to do. If the second erodes, the first becomes unreliable long before anyone notices.
Return to Hollywood. The disruption there is real. Writers' rooms are shrinking, visual-effects pipelines are changing, people have lost real jobs, and they deserve real support. But markets reorganize around disrupted industries over a decade or two. New forms grow in the cracks. Barriers to entry drop. A different kind of storyteller emerges. It hurts; it is not permanent.
Markets do not reorganize around a generation of young people who no longer have the cognitive habits that let them learn in the first place. That is not a sector reorganizing. That is every sector reorganizing, simultaneously, around people quietly deskilled before they ever entered. The damage is cumulative, cross-domain, and multi-decade. There is no next cohort of adults waiting to replace this one.
Luminis has been circling this for a year. The Gates trilogy names the structural choices society faces, from building new institutions to taxing labor and capital honestly. The productive struggle concept is the exact opposite of cognitive debt. Even the Fraga piece written for parents of toddlers carries the same bone, at a smaller scale: real cognitive work has to happen inside the child, not around the child. A tool that completes the work for the child has not taught the child.
Three concrete things, one per audience.
If you are a student (or you parent one): write the first draft before you touch the AI. Even when it feels slow, even when the result is worse than what the AI would have given you. The MIT finding is specific: students who wrote first and only later used AI retained the strongest brain engagement across the three groups. Order matters. The AI is better as an editor than as a ghostwriter.
If you are a teacher: grade the process as much as the output. A ten-minute oral follow-up on a student's own paper, where they walk you through how they arrived at the argument, where they got stuck, what they threw out, is the simplest, cheapest, most honest assessment that exists. Students who used AI to skip the process will hand in beautiful essays and have no idea how they got there. Students who did the work can tell you.
If you are a school or district: teach AI literacy before you give AI access. The Cui and Jalal moderation finding is specific: AI literacy weakens the dependency-to-helplessness pathway. Order matters here too. A student who understands what the tool is for is safer with it than a student who has only ever used it to produce an answer.
The question was never whether to use AI. The question is whether the generation now in school will still have the cognitive habits to use it well, twenty years from now, or whether we are quietly handing them a tool and simultaneously taking back the ability to use it. The second outcome is not a sector disruption. It is a civilizational slow leak.
Questions to sit with.
1. When was the last time you let yourself struggle with something uncomfortable, with no AI help available, and finished the struggle on your own? What did that leave with you that the finished product did not?
2. If a child in your life came home with three perfect essays this month, how would you know whether they wrote them? What would you ask?
3. What is one skill you want to pass to the next generation that no AI tool can hand them? What would it cost you to teach it the slow way?
Research cited in this piece: Kosmyna et al., Your Brain on ChatGPT (MIT Media Lab, 2025); Cui and Jalal, AI dependency and academic motivation (Frontiers in Psychology, 2026); Lee et al., The Impact of Generative AI on Critical Thinking (Microsoft Research and Carnegie Mellon, CHI 2025); NPR coverage of Harvard grade data (November 2025). Classical ground: Seligman and Maier's learned-helplessness experiments at the University of Pennsylvania (1967).