The Generative AI Learning Penalty — A 30-Month Study of 26,811 Chinese Students

Researchers from Stockholm University and the University of Hong Kong tracked 26,811 middle- and high-school students (grades 7-12) in a county in central China over 30 months, from September 2022 to June 2025, combining monthly closed-book exam results, high-stakes zhongkao and gaokao entrance-exam scores, and homework scores and completion times across nine subjects. About 80% of students used generative AI tools — including Doubao, DeepSeek, ChatGLM, Ernie Bot and Qwen — to complete homework.

AI use raised homework scores by 18% and cut average completion time from 64 to 45 minutes. But within six months, monthly exam scores among AI users had fallen 20%; by the two-year mark, zhongkao scores were down 24% and gaokao scores down 18%. The decline was largest in social-science subjects (-27%), followed by STEM (-22%) and English (-17%), and was most pronounced among junior students, higher-achieving students and boys. The authors estimate roughly 80% of the measured learning loss is concentrated among students whose usage pattern — very fast completion paired with high homework scores — indicates outsourcing rather than tutoring-style engagement; students who used AI at a pace similar to non-users showed no exam penalty.

This is a single-county, quasi-experimental panel study, not a randomised trial, and the exact county and province are not named in press coverage or the publicly available abstract. The finding is included here as a rigorous, real-world counterpoint to more promotional accounts of generative AI's classroom benefits — a reminder that unrestricted AI homework use, without deliberate limits on outsourcing, is associated with substantial measured harm to exam performance.