Key points
- AI use can raise scores in the short term but worsen kids’ thinking abilities in the long term, studies find.
- Confidence and comfort with AI predict heavier use, and heavier use predicts the sharpest declines.
- Children pay twice: They lose the thinking, then cannot audit the tool they depend on.
Something strange is happening to students who use AI over long stretches of time. Their assigned work gets better, and their thinking gets worse
The clearest picture comes from a 30-month study (currently a preprint) of nearly 27,000 secondary students, which the psychologist Soren Kaplan recently wrote about. The students using AI saw their homework scores rise about 18 percent and their completion time drop by nearly a third. Better work, done faster
Then, within six months, the same students’ closed-book exam scores fell 20 percent. Their entrance exam scores fell between 18 and 24 percent, with the sharpest declines surfacing after roughly two years of use. The gain was immediate and visible on every assignment. The loss was delayed and invisible until the moment the tool couldn’t be used
A separate preprint study of 299 STEM students across multiple universities showed who these tools eroded, and it was quite surprising. You might suspect it was students who struggled, but it was actually the opposite. The students with the strongest interest in technology, the most comfort with it, and the highest certainty that they had mastered the tool were the ones who used it most routinely and declined the most. Trust and confidence in one’s own capabilities was an accelerant. And dangerously, the steepest decline was in reflection—you know, the important habit of monitoring your own work, catching your own errors, and noticing the gaps in your own understanding.
Educators are used to believing the most capable students have less risk, yet this may no longer hold water as capability with AI and capability without it have begun to come apart. A student can grow fluent at instructing AI and, in the very same motion, grow less able to do the thing the AI is doing for them. Grades still measure a product. Nothing measures whether a student can sit, unassisted, through an exam, interview, or problem without AI assistance
Children Who Use AI Might Pay for It Twice
A child pays for this shortcut twice. The first payment is the thinking the assignment was meant to build. When a student gets an answer without building the thinking, the grades may be high yet the capability never forms. This bill comes due later, because nothing ever signaled that anything was wrong
The second is harder to see. To use AI well, a person has to be able to audit it (catch sycophancy, smoothed edges, or something plausible and hollow). Auditing requires domain area knowledge, which a student won’t build if they offload that work habitually. How can a child who skipped the building check the accuracy of a tool they now depend on? They can’t: which deepens the dependence, which forecloses the capacity further. The first payment is a loss. The second continues tightening every use.
For an adult, the first payment is most of the story, and it is recoverable. An adult who leans on AI loses fluency in something they already built, and can rebuild it. That is atrophy. A child is not losing a capacity but failing to build one during the years it is supposed to form. That is foreclosure, and you cannot rehabilitate what was never there. The same shortcut that costs an adult a skill costs a child the architecture the skill would develop through
I don’t want to play the blame game where we place blame on careless people. The economists Michael Caosun and Sinan Aral describe anaugmentation trap: whoever decides to adopt the tool captures the immediate gain and does not carry the later cost of the capacity eroding underneath it. So apply this to the classroom. The visible, completed work, the easier week, all lead to the higher marks. The cost is paid later, when a student that feels successful can’t actually do the work he’s assigned to do without direct AI intervention.
What Can Actually Protect Kids in the Age of AI
I’ve heard a widening chorus of parents and teacher who want to ban AI from schools, but in my view, this only pushes the behavior out of sight. The more durable approach might be to change what the work asks, so that finishing it requires the thinking AI cannot supply. Build the friction back in on purpose. Alternate work done with AI and work done without it. Ask students why the AI got something wrong, a question no one can answer without understanding why themselves. Reward reasoning, not the output.
Artificial Intelligence Essential Reads
The Horrors of the Empty Office Building in ‘Backrooms’
AI, the Chip and the Good-Enough Clinical Question
The uncomfortable part in all this is that the children paying the most might be the ones who appear, at first glance, to be the most successful. They are fast, fluent, confident, and already generally ahead
The bill for that fluency won’t arrive until later. The work of school, now more than ever, is to make sure that when it comes, there is something underneath the fluency to pay it with
Kaplan, S. (2026, June 25). A study of 26,000 students shows the AI learning trap. Psychology Today. https://www.psychologytoday.com/us/blog/the-power-of-experience/202606/…
Caosun, Michael, and Sinan Aral. “The Augmentation Trap: AI Productivity and the Cost of Cognitive Offloading.” Preprint, submitted 2026. https://doi.org/10.48550/arXiv.2604.03501
Choudhuri, Rudrajit, Christopher Sanchez, Margaret Burnett, and Anita Sarma. “Thinking Less, Trusting More: GenAI’s Impacts on Students’ Cognitive Habits.” Preprint, submitted 2026. https://doi.org/10.48550/arXiv.2601.22430
