Study of 27,000 students finds AI boosts homework but hurts exam scores

[SammyZayn]

A large study tracking about 27,000 Chinese students aged 12 to 18 over 30 months found that after adopting generative AI, homework scores rose by 18 percent while time spent on each assignment fell by roughly 30 percent, from 64 minutes down to 45 minutes on average. But when the same students sat monthly closed book exams without AI assistance, their scores fell by 20 percent within six months, and the effect compounded over time, with high stakes entrance exam scores dropping between 18 and 24 percent after two years

Researchers David Stromberg of Stockholm University, along with Victor Lei and Yanhui Wu of the University of Hong Kong, found the learning losses were concentrated among roughly 80 percent of AI users who specifically used the tools to speed through tasks rather than genuinely engage with the material, while students who kept a similar task completion time to non-users experienced only small learning losses. The researchers wrote that for students, completing tasks efficiently isn't actually the goal, learning from them is, and that the rapid spread of generative AI tools has created a substantial negative impact on genuine student learning

The pattern notably inverted a previously reliable predictor, where strong homework performance used to reliably signal strong exam performance too, but after AI adoption, students with the highest homework grades increasingly performed worse on exams, since AI had effectively started doing the actual thinking for them rather than merely supporting it. An OECD Digital Education Outlook report published in January reached a similar conclusion more broadly. Curious what people think schools should actually do given these findings, restrict AI use during homework entirely, or find ways to redesign assignments so genuine learning stays the actual goal

Faster homework, poor exam results: What AI is doing to students' learningwww.aljazeera.com

[Baz]

Curious how this specific pattern compares across different subjects, the study apparently found social sciences were hit hardest, worth understanding why certain subjects seem more vulnerable to this specific effect than others

[Arty Leah]

The 80 percent versus small minority split among AI users is the detail that actually matters most for any practical policy response, this isn't really about banning AI outright, it's about how specifically students are choosing to use it

[Coastal Otter]

Completing tasks efficiently isn't actually the goal, learning from them is feels like it should be printed on the wall of every classroom currently rushing to adopt AI tools without genuinely rethinking assignment design around them

[Foundry20]

China requiring dedicated hours of AI instruction while also producing this exact cautionary study is a genuinely interesting tension worth noting, the country leading adoption is also generating the clearest evidence of real risk

[Thomas]

Redesigning assignments around genuine understanding rather than simply banning AI outright seems like the more realistic long term path forward, the tools aren't disappearing regardless of how any individual school or teacher feels about them

[Gunther92]

27,000 students tracked over 30 months is a genuinely serious sample size and duration for this kind of study, this isn't a small speculative pilot, it's real longitudinal evidence worth taking seriously in policy conversations

[Luke_67]

That inverted relationship between homework and exam performance is genuinely the most striking finding here, a metric that used to reliably predict learning quality flipped into actively predicting the opposite once AI got involved