Economist Reports Study of 27,000 Chinese Students: AI Use Raises Homework Scores 18% but Cuts Closed-Book Exam Scores 20%

The Economist reported on August 18 on a study covering 26,811 middle and high school students in a county in central China: after using generative AI, homework scores rose 18% on average and homework time fell from 64 minutes to 45 minutes, but closed-book monthly exam scores were 20% lower than peers who did not use AI. The study was released in June by three economists from Stockholm University and the University of Hong Kong, drawing on five middle schools and four high schools tracked from September 2022 to June 2025, a total of 30 months. It also found that after five months of adoption, the share of students outsourcing homework rose from 58% to 81%, and the penalty took about two years to fully materialize.

The study provides verifiable quantitative evidence for AI education product design, school homework evaluation systems, and parental judgment: homework scores may lose their ability to indicate real competence under heavy AI use, while closed-book exams still expose the gap. For AI education product developers, products that only optimize homework completion speed and scores may worsen skill hollowing; for schools, submission records and time data can serve as clues to identify excessive outsourcing. The study also shows the penalty narrows over time but has not reached zero, suggesting adaptation on the tool and teaching sides may partially mitigate the problem. Schools and teachers need to reassess evaluation methods centered on homework scores and submission neatness, as the divergence between closed-book exam and homework performance may widen. AI education products that market speed and score gains may face parental and school doubts about long-term skill degradation, requiring features such as process display or limits on direct answer generation.