AI lifts Chinese homework by 18%, then cuts exams 20% within six months

If you needed proof that “more practice” can still mean “less learning,” this study brings the receipts. Among 26,811 Chinese students in grades seven through 12, using generative AI for homework increased homework scores by 18% and reduced completion time by 30%. Then reality showed up fast: within six months, monthly exam scores decreased by 20%.

The same pattern kept worsening after that. College entrance exam practice results fell by 18% to 24%, and the scores hit their worst after two years, according to the research reported by Fortune. The study is published by the Centre for Economic Policy Research and tied the long and short-term drop to a specific student behavior: outsourcing homework by getting AI to produce accurate answers quickly, not necessarily learning from the work.

That distinction matters because it cuts against the most intuitive argument for AI in school. Homework, in theory, is supposed to reinforce learning. AI, in practice, can function like a turbocharger for output. The research suggests it can raise grades and speed up task completion while undermining the later cognitive step that exams measure. Researchers from Stockholm University and the University of Hong Kong point to about 80% of the students in this “outsourcing” profile as driving the poor test outcomes.

In other words, the study draws a line between performance now and capability later, and it implies that the incentive structure in the homework ecosystem is pulling students toward the easier win. The researchers explicitly frame the problem as an outcome mismatch: for students, finishing tasks efficiently is not the goal; learning from them is. Their conclusion, as summarized in Fortune, warns that generative AI has a substantial negative impact on student learning, even if it improves homework productivity.

This is where the debate stops being abstract and becomes operational for decision-makers. Education leaders, district administrators, and AI vendors are not just choosing “tools” anymore. They are choosing incentive systems. If homework completion becomes faster and superficially more accurate, but exam performance deteriorates, then policies that only track assignment completion or short-term accuracy will miss the metric that matters: transfer, retention, and the ability to apply knowledge outside the AI “container.”

The article connects this to a much older scientific problem known in educational research circles as the transfer problem. Neuroscientist Jared Cooney Horvath, who wrote in testimony to the U.S. Senate Committee on Commerce, Science, and Transportation, argues EdTech has more than 100 years of evidence behind it showing automation can hinder learning. He cites the 1924 “teaching machine” invention by Ohio State University psychology professor Sidney Pressey, where students got better at using the device but struggled to generalize knowledge without it. Three decades later, behaviorist B.F. Skinner built his own version, and the pattern reportedly persisted, with both projects abandoned before school implementation.

Horvath frames the modern AI risk as déjà vu. He argues AI can individualize learning by generating answers to specific queries, but it may not create the friction and critical processing needed for learning subject matter. The mechanism is not just cheating in a moral sense, but dependency in a learning sense. As Horvath puts it in Fortune’s reporting, using tools that help experts can teach novices dependency rather than expertise. That is a board-level issue for any organization supporting AI adoption in educational settings because “learning” is harder to measure than “completion,” and it is easier to optimize the wrong objective.

The policy and social context around this is already tense. The story notes that AI use in schools has proliferated globally. It cites a CollegeBoard survey of more than 1,000 U.S. high schoolers reporting 84% use AI for homework. It also describes documented misuse patterns: Jacob Shelley, an associate professor of health law at Western University, told Fortune in May he was convinced his students cheated on a final exam that included using AI, with 8% getting a perfect score on multiple choice but struggling on the essay portion, including submitting answers with content not in the curriculum. The article also says economic data has yet to show an impact from AI on labor markets or productivity, but that pressure is still shaping student behavior.

Why does that matter for executives beyond classrooms? Because education outcomes feed the talent pipeline, and the pipeline is a competitive advantage. The Fortune report ties student anxiety about AI replacement to labor-market concerns: it says almost 90% of graduates from the class of 2026 are worried AI or automation could replace entry-level jobs, according to Monster. That atmosphere, combined with the incentives around grades and speed, can make outsourcing feel rational even when it harms long-term performance.

For founders and operators building education tech, and for investors and boards underwriting AI deployment strategies, the strategic stakes are clear. If AI boosts homework scores by 18% but cuts exam performance by 20% within six months, then “accuracy” and “productivity” are not the whole story. The study points to a subgroup-driven dynamic, where rapid AI-produced completion does the damage. The question executives need to answer is not whether AI can generate correct answers. It is whether the learning system can prevent offloading, measure transfer, and align incentives so students practice the skill exams actually test.