Major Study Reveals 20% Exam Score Drop Among Students Who Relied on AI Tools

A six-month study of more than 26,000 secondary school students in China has found that those who used generative AI tools for homework saw their exam scores fall by 20 per cent compared with classmates who did not use the technology. The drop occurred even as the same students' homework scores rose by 18 per cent over the same period.

The findings come from a working paper by David Strömberg of Stockholm University and Victor Lei and Yanhui Wu of the University of Hong Kong, published through the Centre for Economic Policy Research. Researchers tracked 26,811 pupils aged 12 to 18 across a county in central China over 30 months, comparing AI users against a control group who did not use the tools.

A Productivity Boost With a Hidden Cost

Homework completion times fell sharply among AI users, dropping from an average of 64 minutes to 45 minutes per assignment. Around 80 per cent of the students reported using generative AI platforms such as Doubao and DeepSeek.

Within six months, however, monthly closed-book exam scores among AI users had dropped by 20 per cent relative to non-users. The effect deepened over time, with entrance exam results falling by 18 to 24 per cent once researchers measured the full two-year period.

Homework Outsourcing Behind the Losses

The study attributes most of the damage to a specific pattern of use. Roughly 80 per cent of AI users showed signs of what researchers describe as homework outsourcing, finishing assignments unusually quickly while still scoring highly.

Losses were largest in social science subjects, followed by STEM and languages, and were most pronounced among younger students, high achievers and boys. Students who used AI but took as long on their homework as non-users experienced comparatively small learning losses.

In China specifically, the effect extended to national examinations. The zhongkao, the high school entrance exam, recorded a 24 per cent decline among affected AI users, while the gaokao, the university entrance exam, fell by 18 per cent.

AI as Tutor or Substitute

Not every AI user suffered the same losses. Pupils who treated the tools as a study aid rather than a shortcut retained most of their learning gains.

The Chinese findings echo a separate study by Zara Contractor and Germán Reyes of Middlebury College, in which undergraduates using a chatbot to learn an unfamiliar topic performed better on tests, with the advantage persisting a week later. The distinction points to a widening gap between AI as a tutor and AI as a substitute for effort, one that researchers say schools have yet to fully address.

The findings land as AI adoption in classrooms accelerates worldwide, with recent polling putting usage among undergraduates at 94 per cent in Britain and 93 per cent in Germany. For schools and policymakers, the data suggests the question may not be whether to allow AI in classrooms, but how to teach pupils to use it without letting the tools do the thinking for them.

The 20 per cent exam score drop identified in this study is not a one-off result either. It builds on a growing body of research on AI and learning outcomes, and it arrives at a moment when schools across the UK and elsewhere are still drafting policy on classroom AI use, making the timing of these findings especially relevant to parents and educators weighing up the trade-offs.