AI Use Raised Homework Scores 18% But Cut Exam Scores 20%

Generative AI use raised homework scores 18% and cut exam scores 20% over six months, per a CEPR working paper. The split matters for any school that grades homework as a proxy for learning. Closed-book exams, taken without the tool, are where the gains disappeared.

The paper is by David Strömberg of Stockholm University, with Victor Lei and Yanhui Wu of the University of Hong Kong. It tracked more than 26,000 Chinese secondary students.

Homework time for AI users fell from 64 minutes per assignment to 45, even as scores went up. Monthly closed-book exams told a different story, with scores down 20% relative to non-users.

Over the full two-year window, entrance exam results were down 18 to 24%. About 80% of AI users showed a pattern the researchers call homework outsourcing, finishing fast while still scoring high.

What the numbers don't settle

Subject-level detail is missing, including which classes took the biggest hit and whether stronger students used AI differently than weaker ones. It also doesn't test whether shifting homework grading toward closed-book checks would close the gap, or just move the outsourcing somewhere else.

The pattern held across the full two-year window, not a single semester's grades.