Generative AI may be helping Chinese students finish their homework faster and score higher on assignments, but the apparent gains come with a steep cost when students sit for exams without AI assistance, a large study has found.

Using 30 months of panel data on 26,811 students in grades 7–12, researchers found that generative AI increased homework scores by 18% and reduced homework completion time by 30%. Yet monthly closed-book exam scores fell by 20% within six months of AI adoption.

The study combined monthly exam results, high-school and college entrance exam scores, homework marks and completion times across nine subjects. Researchers used a staggered difference-in-differences design to estimate the effects of AI adoption.

The decline was even visible in high-stakes entrance examinations. Scores fell by 18% to 24%, with the full effect emerging only after around two years.

Homework gets easier, but learning suffers

The researchers found that the learning losses were largely concentrated among students whose behaviour suggested they were outsourcing their homework to AI.

About 81% of AI-using students completed homework in exceptionally short periods while receiving high scores, suggesting they were relying on AI to produce answers rather than using it as a learning aid.

By contrast, students who continued spending roughly the same amount of time on homework as non-AI users experienced only small learning losses. They also received higher homework scores.

The findings suggest that the problem may not simply be AI use itself, but how students use it.

Social sciences see the biggest losses

The negative effects appeared across all nine subjects studied, but their magnitude varied considerably.

Social sciences recorded the largest learning losses, with exam scores falling by about 27% from the baseline average. STEM subjects followed, with losses of around 22%.

Mathematics scores fell by 22%, equivalent to 0.84 standard deviations.

English and Chinese were comparatively less affected, with scores declining by 17% and 9%, respectively.

The researchers noted that the similarity of the effects across regular and entrance exams strengthens the case that the observed declines reflect a genuine learning effect rather than differences in individual examinations.

Younger and high-achieving students hit harder

Lower-secondary students experienced larger declines than their older counterparts.

Their regular exam scores fell by 24%, compared with 17% among upper-secondary students, a difference of roughly 40% in the size of the effect.

The study also found a clear relationship between AI use and learning losses. Students reporting five or more hours of AI use per week experienced an estimated 30% decline in exam performance, compared with 5% among those using AI for up to one hour a week.

High-achieving students were also more heavily affected.

Students in the highest pre-AI achievement group experienced a 24% decline in regular exam scores, compared with 16% among those in the lowest group. For entrance exams, the corresponding declines were 18% and 11%.

The researchers said this could compress the distribution of academic skills, but through a different mechanism from previous studies in which AI disproportionately helped lower-skilled workers or students.

Boys experience larger learning losses

The study also found a larger negative effect among boys.

Exam scores fell by 21.6% for boys, compared with 18.4% for girls, meaning the negative effect was about 17% greater for boys.

The researchers attributed most of this difference to boys reporting more intensive AI use. They found little evidence that boys were substantially more likely than girls to outsource their homework.

In a June 2025 survey, 82% of boys and 80% of girls reported having adopted generative AI, suggesting only a small difference in overall adoption.

The problem may be how students use AI

The researchers argue that the findings highlight a key distinction in the debate over AI in education.

While much research has focused on designing AI systems that can support learning without encouraging students to outsource their work, the study suggests that students often choose general-purpose AI tools that provide quick answers instead.

The researchers suggest that schools could respond by giving students clearer information about the long-term learning costs of outsourcing homework, placing greater weight on closed-book and in-person assessments, and encouraging parents and teachers to monitor study time and effort rather than simply looking at homework scores.

The study’s central finding is therefore less about whether students use AI and more about what happens when AI replaces the effort that homework is supposed to generate: homework becomes faster and looks better, while performance without AI can deteriorate.