AI homework tools slash exam scores 20%, study of 26,000 students finds
Synopsis
A 30-month study of 26,000 Chinese students found AI homework tools boost short-term scores by 18% but slash exam results by up to 24% over two years — a 'brain drain' effect that only becomes visible long after students are hooked on the technology.
Key Takeaways
26,000+ students across a county in central China were tracked from September 2022 to June 2025 by researchers from Stockholm University and the University of Hong Kong .
AI use raised homework scores by 18 per cent and cut completion time from 64 to 45 minutes , creating a misleading short-term productivity signal.
Monthly exam scores fell by 20 per cent within six months of AI adoption, with the full 'brain drain' effect taking two years to emerge.
High-stakes exam results dropped 24 per cent on the zhongkao and 18 per cent on the gaokao among AI users.
Around 80 per cent of students used generative AI tools, with Doubao , DeepSeek , ChatGLM , Ernie Bot , and Qwen the most popular.
The research has been circulated on the Social Science Research Network and involved the Institute of Psychology of the Chinese Academy of Sciences .
A large-scale study tracking more than 26,000 middle and high school students across a county in central China has found that generative AI tools boost homework scores by 18 per cent but cause exam results to fall by 20 per cent within six months — a 'brain drain' effect that takes two years to fully materialise. The research, conducted by scholars from Stockholm University and the University of Hong Kong, offers one of the most quantified assessments yet of AI's double-edged impact on student learning.
What the study measured
Researchers tracked academic performance from September 2022 to June 2025 — a period of 30 months — monitoring homework grades, completion times, monthly test scores, and high-stakes entrance exam results. By comparing AI users with non-users within the same cohort, the team isolated a consistent pattern: short-term efficiency gains traded against measurable long-term cognitive costs.
Around 80 per cent of the students reported using generative AI for schoolwork. The most popular tools were Doubao, DeepSeek, ChatGLM, Ernie Bot, and Qwen — all products of major Chinese technology firms. Qwen is developed by the cloud computing arm of Alibaba.
The productivity trap
On the surface, AI use delivered clear efficiency gains: homework scores rose by 18 per cent and average homework completion time fell from 64 minutes to 45 minutes. For students and parents, these metrics likely read as unambiguous improvements. But the study's longer timeline tells a starkly different story.
Within six months, monthly exam scores among AI users had dropped by 20 per cent. Over two years, the damage compounded: scores on the zhongkao — China's high school entrance examination — fell by 24 per cent, while results on the gaokao, or National Higher Education Entrance Examination, declined by 18 per cent. Both exams are pivotal determinants of students' academic futures in China.
Why it matters
The findings arrive as generative AI tools spread rapidly through classrooms worldwide, with students increasingly using chatbots to draft essays and solve problem sets. This study is notable for its scale, its longitudinal depth, and its focus on high-stakes outcomes rather than self-reported learning quality. The researchers describe the phenomenon as a 'brain drain' — a gradual erosion of independent problem-solving capacity that only becomes visible over time.
The implications extend beyond China. As AI homework assistants proliferate globally, educators and policymakers face a fundamental question: whether the productivity metrics that make these tools appealing in the short term mask a slower-acting harm to deep learning.
The competitive backdrop
The five AI tools identified in the study — Doubao, DeepSeek, ChatGLM, Ernie Bot, and Qwen — represent the front line of China's consumer AI market, each backed by a major technology conglomerate. Their rapid adoption among students under the age of 18 raises regulatory questions that authorities in Beijing and elsewhere have yet to fully address.
What's next
The study, which has been circulated on the Social Science Research Network and involved researchers affiliated with the Institute of Psychology of the Chinese Academy of Sciences, is expected to intensify debate among education ministries and AI developers over guardrails for student-facing AI products. Whether platform operators will voluntarily introduce friction — such as hints-only modes or usage limits — or whether regulators will mandate them, remains the key question to watch.
Point of View
While the cognitive costs accrue invisibly over semesters. The 'brain drain' framing is significant — it suggests AI homework tools may be functioning less like tutors and more like cognitive offloading devices that atrophy the skills they appear to support. What mainstream coverage tends to miss is that the five tools flagged — Doubao, DeepSeek, ChatGLM, Ernie Bot, Qwen — are all products of companies under pressure to grow their consumer AI user bases, creating a commercial incentive that runs directly counter to pedagogical best practice. Regulators in China, which already tightly controls the gaokao ecosystem, may find this study difficult to ignore as they weigh AI governance frameworks for minors.
NationPress
14 Jul 2026
Frequently Asked Questions
What did the study find about AI and student exam performance?
The study found that AI homework tools boosted short-term homework scores by 18 per cent but caused monthly exam scores to fall by 20 per cent within six months, with high-stakes exam results dropping up to 24 per cent over two years. Researchers from Stockholm University and the University of Hong Kong described this as a 'brain drain' effect.
Which AI tools were most used by the students in the study?
Doubao , DeepSeek , ChatGLM , Ernie Bot , and Qwen were identified as the most popular generative AI tools among the students. Around 80 per cent of the 26,000+ students tracked reported using at least one of these tools for schoolwork.
How long did the AI brain drain effect take to appear?
According to the study, the full negative impact of AI tool use on academic performance took two years to fully emerge. The first signs appeared within six months , when monthly test scores began to decline despite improved homework grades.
What are the gaokao and zhongkao, and how were they affected?
The gaokao is China's National Higher Education Entrance Examination and the zhongkao is the high school entrance exam — both are high-stakes assessments that heavily influence students' academic trajectories. The study found AI users scored 24 per cent lower on the zhongkao and 18 per cent lower on the gaokao compared to non-users.
Does this study apply to students outside China?
While the study focused on students in central China , the tools involved — including ChatGLM and globally available models — and the underlying cognitive mechanisms are not unique to any country. Education researchers and policymakers worldwide are likely to treat the findings as a cautionary signal for AI adoption in classrooms globally.
