Generative AI Learning Penalty

Generative AI may improve learning by acting as a personalised tutor. However, it may also undermine learning by allowing students to engage in homework outsourcing in which students use generative AI to complete assignments and bypass the cognitive effort that produces learning.

This large-scale study by Strömberg, Lei, and Wu (June 2026) examines how generative AI affects learning among secondary school students in China.

Using 30 months of panel data on 26,811 Chinese students in grades 7–12 (who are typically 12-17 years old), the study combines monthly closed-book exams, high-school and college entrance exams, and homework scores and completion time across nine subjects.

The central finding is stark and concerning: while AI dramatically improves homework performance on the surface, it causes substantial and lasting damage to genuine learning.

1. The Productivity–Learning Divergence

The most striking result is a sharp split between surface-level productivity and actual learning. After six months of use, homework scores rise by 18% of the baseline mean while completion time falls from 64 to 45 minutes. However, students’ scores in monthly school exams fall by 20% of the baseline mean, equivalent to 1.4 standard deviations.

In other words, students appear to be doing better — their homework looks polished and is completed faster — but they are actually learning significantly less.

2. Who Is Most Affected?

Lower secondary students are more negatively affected than upper secondary students, with the differences roughly 40% larger in both regular and entrance exams — a pattern consistent with teachers’ reports that high-school students faced more supervised coursework and more restrictions on AI use.

There is a clear dose-response relationship: students reporting 0–1 hours of weekly AI use experience an average 5% score decline, compared to 30% for those using AI five or more hours per week. Students who use AI more get lower grades.

On gender, boys experience a negative effect approximately 17% greater than girls, largely explained by boys’ higher reported intensity of AI use rather than fundamentally different usage patterns.

As the school terms progress, scores become lower.

3. Which Subject Is Most Affected?

The losses are largest in social science subjects, followed by STEM and languages.

4. GenAI Use Time & Scores

If learners use GenAI as personalised tutors – instead of outsourcing homework – their scores are only slightly lower than learners who do not use GenAI to complete homework.

Why Don’t Students, Teachers, or Parents React?

The study explores why such large learning losses go uncorrected. School administrators may react slowly because aggregate effects emerged gradually. Teachers typically observe student performance in only one subject, making a 20% drop seem less alarming than it truly is. Meanwhile, students who fall behind may not connect their declining exam scores to their AI use, as the deterioration unfolds gradually over six months.

Additionally, AI weakens the informational value of homework itself — among AI students with above-median homework scores, higher homework scores are actually associated with lower exam scores, creating a monitoring problem for teachers and parents where better homework performance may paradoxically reflect weaker learning.

Some Signs of Adaptation

There is a small silver lining. The estimated AI learning penalty fell from around 25% in early 2023 to around 16% by June 2025, a trend that holds even for a fixed sample of early adopters, suggesting that students or teachers are gradually adapting to generative AI — though substantial learning losses remain.

Policy Implications

The authors suggest several interventions: providing students with credible information about the long-run costs of homework outsourcing; increasing the weight placed on closed-book, in-person assessments to restore the link between effort and reward; and shifting teacher and parental monitoring from homework outputs (scores) to inputs (time and effort). The finding that AI students who spend the same amount of time on homework as non-AI students achieve similar exam scores does not necessarily imply that simply mandating longer homework time would restore learning for students engaged in outsourcing, as other factors such as intrinsic motivation and effective AI use strategies may also play important roles.

Broader Significance

This study stands out for its sheer scale, its 30-month longitudinal window, and its observation of both short- and long-run learning across nine subjects — advantages that allow it to reveal effects that shorter experimental studies consistently miss. The central message is clear: in realistic conditions where students freely choose general-purpose AI tools, the dominant behaviour is not tutoring but homework outsourcing, with serious and lasting consequences for learning.

This study reminds us to be more circumspect about using AI due to the substantial cognitive debt it incurs.