AI-assisted homework masks 20% drop in exam scores for Chinese students
A large-scale study reveals a 'learning penalty' where generative AI boosts short-term grades while eroding long-term knowledge retention.
Generative AI tools are creating a dangerous decoupling between student productivity and actual learning in Chinese secondary education. A massive study of nearly 27,000 students found that while AI significantly inflates homework performance, it correlates with a sharp decline in closed-book exam results.
Researchers David Strömberg of Stockholm University, along with Victor Lei and Yanhui Wu of the University of Hong Kong, tracked 26,811 students in grades 7 through 12. The data, collected over 30 months across nine different subjects, showed that AI adoption increased homework scores by 18%. Furthermore, the tools reduced the time required to complete assignments by approximately 30%. However, this efficiency came at a steep cost: students using AI scored 20% lower on monthly closed-book exams than their peers who did not use the technology.
The Productivity Paradox
The study utilized a difference-in-differences design to isolate the impact of AI on academic achievement. The results suggest that the immediate gains seen in homework grades are illusory, acting as a mask for a decline in cognitive retention. Because generative AI can provide immediate answers and structure, students are effectively outsourcing the mental struggle required to master complex material. This creates a productivity paradox where students appear to be excelling in their daily coursework while simultaneously losing the ability to perform independently.
Systemic Educational Risks
This 'learning penalty' extends beyond monthly assessments into high-stakes environments. The research found that scores on critical entrance exams fell by 18% and 24% among AI users. Notably, the full weight of this academic penalty did not emerge immediately, but rather materialized after approximately two years of usage. This suggests a cumulative erosion of foundational knowledge that may not be apparent in the first few months of adoption, potentially creating a systemic educational gap that is difficult to reverse once established.
The Path Forward
As AI adoption reaches high levels among students in various wealthy nations, the findings highlight an urgent need to redefine how educators measure progress. The reliance on homework as a proxy for understanding is increasingly unreliable in an era of generative tools. Future academic scrutiny will likely focus on whether specific pedagogical frameworks—such as AI-integrated classrooms that emphasize critical verification over answer-generation—can mitigate these losses. For now, the evidence suggests that without strict guardrails, the efficiency of AI may be trading long-term intellectual competence for short-term grade inflation.