Ten minutes. That is all it took.
In a series of randomized controlled trials with 1,222 participants, researchers found that after roughly ten minutes of AI-assisted problem-solving, people who then lost access to the tool performed worse than those who had never used it at all — and gave up more readily. Not worse while using it; worse afterward, on their own. And more likely, when the work got hard, to simply stop.
The authors name the mechanism plainly: AI "conditions people to expect immediate answers, thereby denying them the experience of working through challenges on their own." Their conclusion is the sentence that should reorganize this entire debate: "persistence is foundational to skill acquisition and is one of the strongest predictors of long-term learning."
AI conditions people to expect immediate answers, denying them the experience of working through challenges on their own — and persistence is foundational to skill acquisition.
Ten minutes. Not a semester of cheating. Ten minutes of help.
Last week, in surveying the wave of books arriving this August — Horvath's Digital Delusion, the critical-pedagogy manifestos — the through-line was that screens degrade attention. This week the claim sharpens and specializes: a cross-disciplinary consensus may, per these sources, be starting to form that AI does something more surgical than distract. It removes the friction where thinking is built.
Start with a classroom, because that is where the symptom first shows.
The student is named Luke Johnson, and what he notices is silence. His peers have gone quiet in class discussion — not hostile, not bored, just absent from the conversation that liberal-arts education is supposed to be.
The professors have noticed too, and their responses tell you how deep the alarm runs. One swaps Aristotle for Arendt, choosing accessibility over difficulty. Another reaches for newspaper articles and comic books, anything to keep students reading at all.
The obvious diagnoses are all available. Blame No Child Left Behind, as Ryan D'Souza reportedly does, for teaching reading as answer-hunting. Blame the opportunity gap in preparation. Blame shyness.
But those explanations were available before, and the silence is new. What is new is that the space between a question and its answer — the space where a student would once have had to sit, uncertain, forming a thought — can now be filled instantly by a machine. A student who has never had to hold that discomfort has nothing to say when you ask them what they think.
This is the turn the cheating discourse has entirely missed. Cheating is legible: someone submitted work they did not do, and you can catch it, prove it, punish it. That legibility is exactly why we have fixated on it, and exactly why it is the wrong thing to fixate on.
Because the deskilling happens before any dishonesty question arises. The honest student who "only" uses AI to brainstorm — who never submits a fabricated word — is being hollowed out by the same mechanism as the one who buys an essay.
To see why, you have to understand curiosity as a physiological event, not a mood.
The dominant psychological account holds that we are wired to chase "knowledge gaps" — the distance between what we know and what we want to know — and that this chase runs directly into the brain's reward circuitry. Neuroscience studies show that waiting for sought-out information activates that circuitry and, crucially, enhances the formation of memory.
Note the word waiting. The reward, and the encoding, live in the interval.
There is a further, stranger detail. A small "priming dose" of information doesn't satisfy curiosity — it inflames it. As Kang and colleagues put it, "the aspired-to level of knowledge increases sharply with a small increase in knowledge, so that the information gap grows with initial learning." Learn a little, and you want much more. Learn enough, and the hunger subsides.
Which means learning is self-propelling — but only if the interval stays open long enough for the first dose to land.
This is what the neuroscientist Anne-Laure Le Cunff means when she writes that "curiosity opens a window, and while the window is open, learning deepens across the board." The window is the gap. The gap is the point.
"When an AI summary answers your search query in three seconds," she continues, "the window closes before curiosity can deepen. You get what you came for, but you also lose what would have turned curiosity into learning."
You lose the adjacent article you might have read. The tangent you might have followed. The connection between two ideas with no obvious relationship.
Her sharpest line names the whole design philosophy of the industry: our technology "is increasingly treating the territory between the query and the answer as dead space to be eliminated, when that territory is where most of the learning actually happens."
Dead space to be eliminated. That is the product roadmap of every AI assistant on the market — optimized for immediacy and completeness, sold as virtues, functioning as a scalpel to the exact tissue where cognition forms.
Now hold the RCT and the curiosity research side by side, and — by this author's inference, since neither source connects the two — the two halves of the mechanism appear to lock together. The three-second answer closes the window; the ten-minute habit erodes the persistence you would need to pry it back open. The tool trains you to expect no gap, then punishes you when a gap appears.
Here is the paradox that makes this worse than ordinary laziness: the better the tool writes, the more completely the thinking is bypassed.
A CEO writing under the banner "Fake Thinking," recounting a point he credits to Dan Koe, puts it as a scene. A teacher assigns an essay — should smartphones be allowed in schools? The old process: the student wrestles, changes their mind halfway through, arrives at a position they can defend. The new process: three prompts, thirty seconds, a perfectly structured essay, "probably better than what the student would have written themselves."
At what point, he asks, did the student actually think?
His metaphor, drawn from Koe, is the one to keep. It's like going to the gym and letting a robot do your squats and deadlifts. An hour later the robot is "drenched in sweat" and announces: "Done. All sets completed." You say, great, I worked out today.
Did you? Your muscles never felt a second of tension. The work got done, but the training never happened.
This is why the quality of the output is a trap. We assumed the essay was a record of thinking that happened earlier — so a good essay meant good thinking. But the essay is not the record of the thought. The essay is the thought, performed. Outsource the prose and you have not outsourced a transcript; you have outsourced the cognition itself.
The stronger the AI's prose, the more perfectly it counterfeits a process that never occurred. Which produces, in the writer's chilling phrase — one he attributes to Koe's argument — "sophists with perfect grammar" — students who can argue any side of any question, brilliantly, without ever holding a conviction of their own.
The corroboration reportedly arriving from several disciplines points in a similar direction, and that apparent convergence is what draws notice.
Microsoft's survey of 319 knowledge workers found that higher confidence in the AI was, per the study, associated with less critical thinking — a self-reported, correlational pattern suggesting that trusting the tool may accompany less checking of one's own reasoning. A Frontiers taxonomy distinguishes tools that scaffold cognition from those that "circumvent as opposed to augment" it, warning that generative systems operate as "cognitive surrogates rather than as scaffolds." And a workshop paper coins "cognitive debt" for the neurophysiological finding of "persistent underengagement after AI use" — a bill that comes due later, quietly.
Neuroscientists, psychologists, pedagogues: they appear, per these sources, to be pointing in a similar direction — not, as far as the evidence shows, by coordination, but perhaps because they are examining related phenomena.
The policy stakes follow directly, and they are structural, not disciplinary. If an institution defines learning as output-production — the essay, the correct answer, the completed set — then it will rationally optimize away everything that slows output down. It will optimize away the friction. It will optimize away the learner.
The RCT authors offer the only coherent design principle: AI must "prioritize scaffolding long-term competence alongside immediate task completion." A good mentor, they note, sometimes offers no help at all — knows when not to help. No shipping AI product does this, because refusal does not sell.
Which leaves the hardest question, the one this thread has been building toward for weeks. Can persistence-eroding friction ever be engineered back into a tool that students freely choose? A product that says no, that withholds the answer, that keeps the window open on purpose — against every incentive of engagement and market share?
Perhaps not. But the alternative is to keep mistaking the drenched robot for a workout, while the habits of mind now forming — habits the researchers warn "may prove difficult to reverse" — quietly harden.
The scandal was never the stolen work. It is the unbuilt mind.