Artificial intelligence (AI) boosts homework scores but cuts exam results by 20%, and this “brain drain” effect takes two years to fully emerge, a new study has found.
Generative AI is rapidly transforming classrooms as much as workplaces, with students increasingly turning to chatbots to draft essays and solve problem sets. But is the technology a personalised tutor or a slow-acting cognitive poison?
Research by scholars from Stockholm University and the University of Hong Kong, who tracked the academic performance of more than 26,000 middle school and high school students across a county in central China over 30 months, offers a quantified answer.
From September 2022 to June 2025, the team tracked homework grades, completion times, monthly test scores and entrance-exam results. By comparing AI users with non-users, they found a clear gap: AI boosted short-term efficiency but harmed long-term learning.
Around 80% of the students reported using generative AI, with Doubao, DeepSeek, ChatGLM, Ernie Bot and Qwen emerging as the most popular tools.
Qwen is developed by the cloud computing arm of Alibaba, which also owns the South China Morning Post.
AI use boosted homework scores by 18% and cut homework time from 64 to 45 minutes – a clear productivity leap.
But within six months, the same students’ monthly exam scores had dropped by 20%, and after two years their high-stakes entrance exam results fell sharply, with a 24% drop in the zhongkao high school entrance exam and an 18% drop in the gaokao, or National Higher Education Entrance Examination.
The findings are detailed in the discussion paper The Generative AI Learning Penalty: Evidence from Chinese Secondary Education, published on the Social Science Research Network repository last month.
The researchers warned that previous short-term studies missed the real damage: because only the final exams at the end of school tested comprehensive knowledge, the negative effect built slowly and only became fully visible after two years.
The conclusion may not come as a surprise, but the underlying mechanism is telling.
Behavioural analysis reveals that 81% of AI users outsourced homework, generating direct answers to finish faster and score higher. The remaining 19%, however, spent similar time on homework as non-users, treating AI as a personalised tutor to aid thinking, and their exam results stayed stable.
The researchers suggested AI was not inherently harmful, but instead urged a focus on how it affected the cognitive effort needed for real learning. For students, finishing tasks quickly is not the objective – learning from them is.
The decline also varied across subjects and student groups: social sciences (politics and geography) fell most, by 27%, followed by STEM at 22% and languages (English at 17% and nine per cent for Chinese); a pattern likely driven by the ease with which AI can generate entire humanities essays, according to the paper.
Junior secondary students saw sharper declines than their senior-secondary counterparts owing to looser oversight, while boys underperformed girls by 17%, reflecting more intensive AI use.
And most tellingly, top-performing students dropped 24%, outstripping the 16% decline among lower achievers – possibly because stronger students leaned more heavily on AI to chase the “optimal” answer.
The researchers asked why, given the magnitude of the documented learning penalty, there had been no stronger pushback against such learning losses, with the answer possibly in the delayed effect: over the first two years, the average score fell only 3.4%. By June 2025, however, the total induced decline had grown to nearly 10% only after AI adoption surged to 80%.
Teachers might not react promptly because they typically observe performance within a single subject, where normal fluctuations mask the underlying trend, the paper said. The researchers said students themselves, meanwhile, might experience a deceptive sense of fluency – AI-made tasks feel effortlessly accomplished, reinforcing the subjective impression of heightened productivity.
In June, a Beijing science forum discussed how children should learn in the AI era. Researcher Gao Wenbin with the Institute of Psychology of the Chinese Academy of Sciences argued that strengthening the brain’s core processing power was a priority.
“Children need four basic abilities: gathering information, switching tasks, sustaining focus and multitasking. These underpin higher mental functions – cognition, emotion, willpower and behaviour – all of which directly affect learning outcomes.”
The real challenge, he said, was safeguarding the mental effort that truly enabled learning. – South China Morning Post
