A study tracking 26,811 Chinese secondary school students over 30 months has produced a set of numbers that pull in opposite directions: homework written with generative AI scored 18% higher, and the time spent on a single assignment dropped from 64 minutes to 45. But within six months, the same students' monthly exam scores fell 20%.
The paper, titled "The Generative AI Learning Penalty: Evidence from Chinese Secondary Education," is co-authored by David Strömberg of Stockholm University and Victor Lei and Yanhui Wu of the University of Hong Kong. It was released in June as CEPR Discussion Paper DP21577, and The Economist's chart-of-the-day column revisited it on August 18. The sample covers grade 7 to grade 12 students in central China, with data collected from September 2022 through June 2025; roughly 80% of students in the sample reported using generative AI at some point.
The bill comes due two years later
The short-term numbers all look good. Homework scores went up, time was saved, and from where parents and students sit, that reads as a clear efficiency win.
Stretch the window to two years, though, and another set of numbers surfaces: scores on the high school entrance exam (zhongkao) fell 24%, and scores on the college entrance exam (gaokao) fell 18%. Breaking the losses down by subject, social sciences dropped the most, followed by STEM and language subjects; breaking them down by group, middle schoolers, previously high-achieving students, and boys were hit hardest.
The fact that previously high-achieving students lost more is more striking than the headline number. The usual intuition is that stronger students know how to use tools better; the result runs the other way — they were already converting high-intensity independent practice into scores, so once that practice got replaced, the absolute loss was largest for them too.
Where did the saved time go
Do the rough math: 19 minutes saved per assignment, once a day, over roughly 200 school days a year, adds up to about 63 hours a year — more than 100 hours over two years. The study didn't track where that time went, but it does establish something else: the private-tutoring effect of AI and the time-saving effect coexist, and even added together they didn't hold up exam scores.
In other words, the problem isn't that kids are slacking off. Homework itself, as a form of practice, got skipped — the working-through, the trial and error, the getting stuck and figuring a way out. Those are the moves that have to be reproduced in a closed-book exam room, and they didn't happen in an AI-assisted homework workflow. Homework scores measure output; exam scores measure ability. The gap between the two indicators has never been this wide.
The researchers also cite a counterexample: college students who used AI for active learning scored higher on tests, and retained more a week later. The difference is in how it's used, not in the tool itself.
Free subscriptions are pushing the other way
The timing is oddly pointed. The same month this paper resurfaced, Google rolled out a free year of Gemini for eligible students, OpenAI launched a version of ChatGPT aimed at teenagers, and several domestic Chinese vendors pushed their education products hard for back-to-school season. The tools named in the study are the Chinese counterparts of these products — as of June 2025, among students who used AI, Doubao accounted for 47%, DeepSeek 36%, ChatGLM 14%, Ernie Bot 9%, and Qwen 8%.
Some regulators have already made the opposite call. Norway banned AI use in schools for students aged 6 to 13 starting in August, citing the same underlying concern this study points to.
The paper is unlikely to change vendors' marketing pace — the acquisition cost and retention value of the student market speak for themselves. But it does push the conversation past "does AI help learning" and into "which uses help, and which uses hurt." Telling the two apart requires teachers to be able to see students' homework process, not just the result — which happens to be exactly what today's homework systems are worst at doing.
Sources: CEPR Discussion Paper DP21577, CocoLoop, The Economist chart column, South China Morning Post; the percentage changes in homework and exam scores, sample size, and tool market shares have been checked against the paper's original figures.