The authors tracked 26,811 Chinese students in grades seven through twelve over 30 months, comparing students who adopted generative AI at different times to measure its effect on homework and learning across nine subjects, alongside monthly closed-book exams and high-stakes entrance exams.
Students using AI finished their homework 30 per cent faster and scored 18 per cent higher, but their learning declined. Within six months, monthly exam scores fell by 20 per cent. The penalty on high-stakes entrance exams reached 18 and 24 per cent, though this full drop took about two years to appear. The damage was worst in social sciences and hit junior students, high-achieving students and boys the hardest. These losses were concentrated among the roughly 80 per cent of users whose behaviour suggested they were outsourcing the work to AI, producing high grades in very little time. Students who used AI but spent the usual amount of time on their homework suffered only minor drops in their test scores.
When an automated tool can generate a perfect essay or problem set in seconds, the finished product no longer proves that a person did the underlying cognitive work. Any system that relies on visible output to measure human effort breaks down once generative AI enters the workflow. Institutions that grade the final product are simply measuring the software.
Today’s links: Assorted links for 28 September 2026.