AI Homework Help Raises Grades But Wrecks Exam Scores
The Story
Artificial intelligence helps students finish homework faster and score higher on assignments, but new research shows it can quietly damage exam performance.
Economists David Stromberg of Stockholm University with Victor Lei and Wu Yanhui of the University of Hong Kong tracked nearly 27,000 Chinese pupils aged 12 to 18, roughly 80 percent of whom used chatbots such as Doubao and DeepSeek.
After six months the AI users raised their average homework scores by 18 percent and cut assignment time from 64 minutes to 45.
Those same students then scored 20 percent below classmates who never used AI when they sat their exams.
The penalty fell almost entirely on pupils who rushed their homework, while students who kept spending the same amount of time on assignments paid little price.
A separate study by Zara Contractor and German Reyes of Middlebury College found undergraduates who learned an unfamiliar topic with a chatbot tested better, and held that advantage a week later.
The evidence points to a single lesson: AI can raise learning productivity, but only when students think a problem through before reaching for the answer.
Economists David Stromberg of Stockholm University with Victor Lei and Wu Yanhui of the University of Hong Kong tracked nearly 27,000 Chinese pupils aged 12 to 18, roughly 80 percent of whom used chatbots such as Doubao and DeepSeek.
After six months the AI users raised their average homework scores by 18 percent and cut assignment time from 64 minutes to 45.
Those same students then scored 20 percent below classmates who never used AI when they sat their exams.
The penalty fell almost entirely on pupils who rushed their homework, while students who kept spending the same amount of time on assignments paid little price.
A separate study by Zara Contractor and German Reyes of Middlebury College found undergraduates who learned an unfamiliar topic with a chatbot tested better, and held that advantage a week later.
The evidence points to a single lesson: AI can raise learning productivity, but only when students think a problem through before reaching for the answer.
Why It Matters
The damage surfaces where it hurts most. Researchers at Stockholm University and the University of Hong Kong tracked 26,811 students across one central China county from September 2022 to June 2025, recording entrance exam falls of 24 percent on the zhongkao and 18 on the gaokao, with top performers down 24 percent against 16 for weaker classmates, The Star reports. The slide hid in plain sight: average scores moved just 3.4 percent over two years while adoption surged to 80 percent. The pattern is a century old, per Fortune: in 1924 Sidney Pressey abandoned his teaching machine because pupils mastered the device, not the material.
Go Deeper
Read the original reporting at The Economist.
Read Full Story at The Economist →