AI lifts homework grades, sinks exam scores

Students turning in better homework, faster, thanks to AI: on paper, everyone wins. Then comes the in-class exam, no assistant allowed. And scores drop by 20%. For the strongest students, the fall is even sharper.

That's the finding of a study that pours cold water on the enthusiasm around AI in the classroom. Not an opinion poll, not a three-class experiment: a large-scale longitudinal study, with real grades and a method built to separate the tool's effect from everything else.

What the study actually measured

The work is signed by economists David Strömberg, Victor Lei and Yanhui Wu, published on June 2, 2026 by the Centre for Economic Policy Research (CEPR, discussion paper DP21577). The researchers tracked 26,811 Chinese middle and high schoolers, from seventh grade to senior year, over 30 months. They cross-referenced homework grades, time spent on it, monthly in-class exams and the big entrance exams, across nine subjects.

The study's strength is its method. Comparing students who use AI to those who avoid it would prove nothing: they're two different populations to begin with. Instead, the authors exploit the fact that adoption came in waves, at different dates depending on the class. This difference-in-differences technique lets them observe the same student before and after, and isolate AI's own effect. This isn't a correlation, it's a step toward cause.

Homework up, competence down

The numbers tell two opposite stories. The visible side: homework grades climb 18% and the time spent on it drops 30%. The hidden side: within six months, in-class exam scores fall 20%. And on the entrance exams, the ones that decide a student's track, the penalty reaches 18% for high school admission and 24% for university admission.

The most unsettling part is the lag. The drop doesn't show up right away: the full effect takes about two years to appear. Meanwhile, everything looks fine. Assignments come in, continuous-assessment grades hold, parents are reassured. The meter tracking real skill, though, is quietly running backward.

It works a bit like GPS. You reach your destination without a wrong turn, but six months later you can't retrace the route without the screen. The performance is there. The learning is not.

Getting it done isn't doing it

The study doesn't claim AI makes you dumber. It isolates one specific behavior: outsourcing the assignment. Around 80% of users share the same profile, an abnormally short work time paired with high homework grades. They hand the assignment to AI instead of doing it with AI.

The counter-example is telling. Students who use AI but keep a work time comparable to those who skip it suffer only small losses. The difference isn't the tool. It's what you hand off to it: the finished output, or the effort that builds the skill.

Why the best students take the biggest hit

Intuition says strong students know how to use AI better, so they're better protected. The study shows the opposite. The heaviest losses hit the youngest students, the boys, and above all the best.

The explanation lies in their very fluency. A strong student knows exactly what to ask, how to reframe a prompt, how to get a clean answer on the first try. That mastery lets them delegate more, and more efficiently. Their skill at steering the machine hides the erosion of what they can do alone all the better. The finish is shinier, the crack underneath deeper.

By subject, social sciences take the biggest hit, ahead of the sciences and languages. Where an answer is written rather than calculated, the temptation to let AI hold the pen is stronger.

Not a case against AI in school

None of this argues for banning AI from classrooms. The authors draw a clear line between two uses: AI as a tutor, which explains and drills, and AI as a crutch, which delivers an answer to copy. The first use doesn't show up in the measured damage. The culprit is less the machine than the shortcut it makes irresistible when the only thing graded is the homework turned in.

The education debate often fixates on cheating. This study moves the problem. Even a perfectly honest student, one who understands what the AI produces, can watch their own skill erode if they stop walking the path themselves.

The same trap at the office

The mechanism reaches far beyond school. An employee who delivers faster thanks to AI ticks every productivity box while slowly unlearning the craft of the job. Measured performance climbs, the ability to work without the tool frays. It's called skill atrophy, and it follows exactly the same curve: immediate gain, deferred cost.

Declic has run into this illusion from another angle. In the benchmark where AIs played CEO, the result could look solid as long as you didn't scratch the surface. Our breakdown of Princeton's CEO-Bench showed the same trap: a deliverable that shines can cover for a fragile capability.

One question the study raises without answering. If the cost takes two years to surface, and it hits hardest the people we assumed were best equipped, how many students and workers are running up a bill today that they won't see arrive until the talent is already gone?