A 26,000-student study shows AI's hidden learning cost takes two full years to surface

Self-reported AI usage rose from near zero to about 80 percent over the study period, with a big jump coinciding with the releases of DeepSeek V2.5 in September 2024 and DeepSeek R1 in January 2025. The most popular tools were Doubao, DeepSeek, ChatGLM, Ernie Bot, and Qwen.

The study takes advantage of the fact that students discovered AI on their own at different times. The authors use a difference-in-differences design, a method that measures the change in outcomes for a treated group before and after an intervention and subtracts the change over the same period for an untreated comparison group.

Here, they track how each student's performance shifted before and after they started using AI, then contrast that trend with students who weren't using AI yet. The timing of first use comes from self-reported data, and the causal claim assumes both groups would have developed similarly without AI.

Better homework, worse test scores

Six months after first using AI, homework scores rose by 18 percent while average time per assignment fell from 64 to 45 minutes. At the same time, scores on monthly closed-book exams dropped by 20 percent.

The effect on high-stakes entrance exams was just as large but built up more slowly. Regular exam performance fell off within half a year, but the full impact on entrance exams took about two years to appear, ranging from an 18 to 24 percent decline. Short-term studies therefore miss the long-term cost to learning, according to the researchers.

Four out of five long-term users show signs of outsourcing

After more than five months of AI use, about 81 percent of students finished their homework in under 50 minutes, faster than even the quickest non-users. They got high homework grades but bombed exams. The combination of short completion times, high homework grades, and low exam scores suggests these students were outsourcing their work to AI, the authors write.

AI users who spent a similar amount of time on homework as their non-AI classmates, on the other hand, scored just as well on exams while also earning better homework grades. This group showed no sign of positive selection based on prior performance, meaning they weren't simply better students to begin with, and AI isn't harmful by default. It causes damage mainly when it replaces independent thinking.

Social sciences take the biggest hit

Social science subjects like politics and geography saw an average decline of 27 percent, STEM subjects 22 percent, English 17 percent, and Chinese 9 percent. That matters because most previous experiments have focused on math, programming, and foreign languages.

The effects also varied sharply across student groups. Younger students in lower secondary school lost more than older ones (24 versus 17 percent), and boys were hit harder than girls (21.6 versus 18.4 percent), which the study attributes mainly to heavier AI use among boys.

Top performers suffered the most, with the top third seeing a minus 24 percent effect compared to minus 16 percent in the bottom third. A dose-response pattern showed up as well. Students using AI for up to one hour per week lost about 5 percent, while those using it five hours or more lost 30 percent.

Why almost no one is pushing back

The estimated learning penalty fell from about 25 percent in early 2023 to 16 percent by June 2025. The decline also showed up in a fixed group of early adopters, suggesting some degree of adaptation by students and teachers, but the losses haven't gone away.

The study explains why the reaction has been muted. Teachers typically see students in only one subject, where a 20 percent grade drop isn't unusual on its own. The aggregate effect on the county average didn't reach about minus 10 percent until June 2025 because few students had been using AI long enough for the damage to accumulate. Students themselves often don't connect the dots, mistaking the mental effort of independent learning for a sign that they're learning poorly.

As countermeasures, the study suggests giving students credible information about the long-term costs of outsourcing, putting more weight on in-person exams, and tracking completion time instead of homework grades. AI erodes the value of homework as a signal, and among AI users with above-average homework scores, higher homework grades actually predict worse exam results.

Anthropic researcher Andrej Karpathy has argued that schools should stop trying to police AI-generated homework and instead shift the majority of grading to in-class work. His reasoning aligns with what this study found. When students know they'll be tested without AI, they stay motivated to actually learn the material.

The pattern lines up with recent findings from other settings. An Anthropic study recently showed that participants who learned new programming skills with AI help scored 17 percent worse on follow-up knowledge tests than the control group, without saving any real time. The results depended on how people used the tool. Those who simply copied AI answers performed worse, while those who used AI to better understand the tasks didn't see the same decline.

A study by the Swiss Business School found a negative link between AI use and critical thinking. A separate study by researchers at several American and British universities showed that people who treat AI mainly as an answer machine lose cognitive skills the fastest.

A UC Berkeley study analyzing more than 500,000 grades also showed that the share of top A grades in writing- and programming-heavy courses has risen by 13 percentage points since ChatGPT launched. There, too, the effect was concentrated on unsupervised homework, while proctored exams showed no comparable gains.