AI boosts homework speed but harms exam performance, new study shows
A large-scale study of Chinese students found that generative AI raised homework scores while exam results fell sharply.
A joint investigation by David Stromberg, Victor Lei and Yanhui Wu monitored about 27,000 Chinese students for two and a half years, comparing homework outcomes, monthly exams and entrance test scores across nine subjects. The adoption of generative AI lifted homework marks by 18 % while cutting average task time from 64 to 45 minutes. In contrast, students’ scores on AI-free monthly exams dropped 20 % after half a year, and performance on high-stakes entrance exams deteriorated over a longer horizon.
The OECD’s Digital Education Outlook 2026 echoed these findings, warning that unstructured AI use can inflate grades without genuine learning, a phenomenon dubbed “metacognitive laziness.” European surveys by Ipsos for the Vodafone Foundation and the HEPI/Kortext Student Generative AI Survey 2026 revealed that most pupils use AI for quick information or explanations, with fewer employing it for personalized learning or adaptive content, and mixed effects on feelings of loneliness.
Why it matters
Understanding AI's impact on learning helps educators shape policies that protect genuine knowledge acquisition.
In this story
