AI Raises Homework Scores 18% — and Tanks Exam Results for Two Years
Agentic AIAI made students look better at homework while making them worse at applying what they studied — and the damage didn’t show up on high-stakes tests for two full years.
What the Study Found
Researchers analyzed 30 months of panel data covering more than 26,000 secondary school students (grades 7–12) in a county with one million residents in central China. Using a difference-in-differences design, the study tracked individual student performance before and after AI adoption, comparing monthly exams, homework scores, completion times, and high-stakes entrance exams.
The surface-level results looked positive. Within six months of first AI use, homework scores increased 18% and assignment completion time dropped from 64 to 45 minutes. Students were finishing faster and scoring higher on the work they submitted.
The closed-book exam data told the opposite story: monthly test scores fell 20%. High-stakes entrance exam scores declined 18–24%. The entrance exam effect took approximately two full years to fully manifest — which explains why it went undetected in earlier, shorter-window studies.
Why It Matters for Builders
The usage pattern that drove the damage was straightforward: about 81% of long-term AI users finished homework quickly with high grades but performed poorly on exams. The researchers concluded students were outsourcing the work rather than using AI to deepen understanding. Homework was no longer functioning as practice — it had become a deliverable the AI completed on the student’s behalf.
The subject-level breakdown is instructive: social sciences lost the most ground (27% decline), followed by STEM (22%), English (17%), and Chinese (9%). The subjects where AI-generated responses are hardest to distinguish from genuine understanding suffered the steepest declines.
Younger students lost more than older ones (24% vs. 17%), and boys showed larger negative effects than girls (21.6% vs. 18.4%). Neither finding is fully explained by differential AI access — the patterns held after controlling for usage rates.
The researchers’ conclusion: AI erodes the value of homework as a learning signal, and they recommend shifting emphasis toward in-class assessments that cannot be offloaded.
What This Means for Enterprise AI Adoption
The education context is a controlled version of a problem that appears anywhere AI is introduced to knowledge work: short-term output metrics improve while underlying capability development slows. If the work that was supposed to build skill is being completed by the tool instead, the skill doesn’t build.
For engineering teams deploying AI coding agents, this study suggests that productivity gains measured by tickets closed or code shipped may obscure a two-year lag in the engineering fundamentals that compound over careers. The 26,000-student study is the first large-panel evidence that the lag is real, measurable, and only visible at the two-year horizon.
The So What
AI doesn’t make students — or knowledge workers — worse immediately. It makes them look better in the short run while eroding the practice that builds durable capability. The study’s methodological contribution is the two-year time horizon: organizations measuring AI’s impact over quarters rather than years are likely missing the compounding cost. The anthropic-ai-skill-inequality-gap-2026 dynamic follows the same curve — AI amplifies existing skill, but only if the skill was built first.
Source: The Decoder
Content created with AI assistance and reviewed for accuracy.
Join the conversation
Stack Insiders is our free community for readers who want to go deeper — share resources, ask questions, and connect with others across every vertical we cover.
Join Stack Insiders →