← Back to feed

A study covering 27,000 middle and high school students in China documents a consistent pattern: AI-assisted learning tools drove significant short-term homework score improvements, followed by serious declines in proctored exam performance. The study was led by David Strömberg of Stockholm University and published in The Economist.

Students aged 12 to 18 were tracked. Around 80% reported using Chinese AI tools — a market where AI adoption in education has accelerated faster than in most Western countries. The scale makes this one of the largest empirical records of AI’s effect on secondary school performance.

The Split

Homework performance improved markedly with AI access. Exam performance — proctored, without AI — declined significantly for the same students over time.

High achievers were disproportionately affected. Students who performed well without AI assistance saw the steepest exam score drops after AI adoption. The mechanism is straightforward: AI tools substitute for the cognitive effort that builds durable knowledge. Students who had the most to lose from outsourcing that effort lost the most.

How This Study Differs From Earlier Work

Prior AI education research has found mixed results. A June 2026 ALEKS study of 3.2 million math records found AI lifted completion speed 23% while dropping proctored retention 25% — a smaller and narrower finding. A July 2026 study tracked accuracy alongside confidence and found AI access collapsed the “I don’t know” rate from 44% to 3% while dropping accuracy by two-thirds.

The China study is notable for scale (27,000 is large for this domain), breadth (all subjects, not just math), and for explicitly comparing the same students’ homework versus exam performance across time. It captures the full cycle: performance with AI support, then performance without it.

The Mechanism

The pattern matches what cognitive scientists call retrieval practice effects. Retrieval practice — pulling knowledge from memory under pressure — builds durable, accessible knowledge. Homework completed with AI bypasses retrieval. The student gets the right answer without the encoding process that makes it retrievable later.

AI tools, in this framing, are not failing at education. They are succeeding at homework completion while hollowing out the underlying knowledge-building process.

Policy Pressure

This data will surface in regulatory discussions. Several countries are debating AI access restrictions in educational settings. The scale and clarity of the China study gives policymakers the clearest large-scale empirical case to date that unstructured AI access in secondary education can improve measurable short-term outputs while degrading underlying competency.

The harder problem has no clean answer: homework and exams were designed assuming students had to do the work themselves. The intervention that would actually close the gap — redesigning assessment to integrate AI, or restricting access entirely — is not straightforward in either direction. What the study confirms is that the current middle ground, where AI is accessible but exams are not, produces the worst of both outcomes.