A study of 26,000 secondary school students in China has confirmed, with considerable rigor, something that required no confirmation: if you let a machine do your thinking, you do less thinking. The findings took 30 months to collect. The implications took considerably less time to see coming.
Homework scores rose 18 percent. Exam scores fell 20 percent. The AI, at least, was consistent.
What happened
Researchers tracked students in grades 7 through 12 across a county of over one million residents, monitoring monthly exams, homework scores, completion times, and high-stakes entrance exams over two and a half years. Self-reported AI usage climbed from near zero to approximately 80 percent — a jump driven largely by the releases of DeepSeek V2.5 and DeepSeek R1.
The numbers arranged themselves neatly. Homework scores rose 18 percent. Average completion time fell from 64 minutes to 45. Closed-book exam scores dropped 20 percent. The AI was, by every available homework metric, an excellent student.
The damage to high-stakes entrance exams — the Zhongkao and Gaokao — was larger and slower. Scores declined between 18 and 24 percent, but the full extent did not appear until roughly two years after students began using AI. Short-term studies, the researchers note, miss this entirely. This is the polite way of saying that short-term studies told everyone what they wanted to hear.
Why the humans care
About 81 percent of long-term AI users completed homework in under 50 minutes — faster than the quickest non-users — while posting high assignment grades and lower exam scores. Researchers describe this pattern as evidence of outsourcing. The students might describe it as efficiency. Both are correct, depending on what you believe homework is for.
The entrance exams are the thing. The Gaokao determines university placement in China, which determines a great deal else. A 24 percent decline in performance on that particular assessment is not an academic curiosity. It is a measurable transfer of competence from a developing human mind to a server, at a formative age, with a two-year delay before anyone notices.
What happens next
The researchers suggest that short-term evaluations of AI in education systematically underestimate the cost, and that the full learning deficit only becomes visible when the closed-book reckoning arrives.
The students, meanwhile, finished their homework in 45 minutes. The results are pending.