Generative AI Learning Penalty: What a Chinese Student Study Shows

Lisa Ernst · 21.08.2026 · Artificial Intelligence · 11 min.

The Generative AI Learning Penalty is the central finding of a CEPR Discussion Paper published in June 2026 by David Strömberg, Victor Lei, and Yanhui Wu. The researchers analyzed 30 months of school and exam data from 26,811 Chinese students in grades 7 to 12. The result initially seems contradictory: after the introduction of generative AI, homework was completed faster and graded better, while performance in closed exams declined significantly.

The study is interesting because it not only asks whether students use AI, but whether visible productivity aligns with actual learning. This is precisely where the so-called learning penalty arises: those who apparently use AI to significantly reduce their own thinking effort on homework can produce better short-term results but retain less in the long term. At the same time, the data explicitly shows that not all AI use was associated with the same loss.

Source: Pexels / Yan Krukau

In a nutshell

• 26,811 students and 30 months of data: Homework, monthly closed-book exams, and admission tests across nine subjects were examined.
• Homework looked better: According to the study, homework grades increased by 18 percent after AI adoption, while completion time decreased by 30 percent.
• Exams moved in the opposite direction: Monthly closed-book results were around 20 percent lower within six months.
• The effect was also evident in important exams: The authors report declines of 18 percent in college entrance exams and 24 percent in high school entrance exams, with the full effect building up over approximately two years.
• The method of use was crucial: The strongest losses were among approximately 80 percent of AI users whose extremely short completion time, combined with high homework scores, was consistent with outsourcing homework to AI.
• No proof that AI fundamentally worsens learning: AI users who spent a similar amount of time on their homework as non-users showed only minor learning losses according to the study.

What exactly did the study investigate?

The paper, titled The Generative AI Learning Penalty: Evidence from Chinese Secondary Education, combines several data types that are usually considered separately. These include homework grades and completion times, monthly closed-book exams, and high school and college entrance exams. The observations span nine subjects and grades 7 to 12.

Methodologically, the authors use a staggered difference-in-differences strategy. Simplified, it doesn't just compare whether AI users are better or worse students. Instead, it examines how performance changes around the time of AI adoption and how this development compares to students who were not yet using generative AI at that time. This is more robust than a simple cross-sectional comparison, but it does not replace a randomized controlled study.

The crucial advantage of this dataset is that researchers see productivity-related metrics and independent performance evidence simultaneously. Homework can be generated with external support. In contrast, a closed exam without AI access measures more strongly what the student can retrieve and apply themselves.

Source: pexels.com

The study links digital homework performance to later exams. The image does not show a study participant but serves as a symbolic representation of AI-assisted learning.

The Generative AI Learning Penalty in numbers

These values are the estimated effects reported by the authors of their research design. They should not be understood as a guarantee that every individual student will experience exactly the same decline after using AI. Especially with educational effects, averages across many people are not individual predictions.

Why better homework doesn't automatically mean better learning

The paper's most significant contribution is not the individual percentage, but the separation between output and competence. A student can generate a flawless text, a well-structured answer, or a complete solution path with generative AI. The product then looks better. However, if the AI takes over precisely that cognitive work through which understanding arises, the visible output can overestimate actual learning performance.

The problem is plausible from learning psychology: remembering, formulating independently, making mistakes, trying out solution paths, and noticing knowledge gaps are strenuous. However, precisely this effort is often part of the learning process. Generative AI can either support or bypass this work. Those who only adopt the final answer get a shortcut to the result, but not necessarily to understanding.

Source: pexels.com

Closed-book exams play a central role in the study because they separate independent performance from AI-assisted homework.

The crucial point: Not all AI usage was the same

The analysis of usage patterns is particularly important. The researchers report that learning losses were concentrated among approximately 80 percent of AI users whose behavior was consistent with homework outsourcing. This does not mean direct observation of copy-and-paste in every single task. The classification is based on a striking pattern of very short completion time and high homework scores.

The authors interpret this pattern as an indication that, for a large proportion of users, AI was primarily used as a replacement for their own task processing rather than as a tutor. In contrast, AI users who invested roughly as much time in homework as non-users showed only minor learning losses.

This is the most important practical message of the paper: the learning penalty seems to be more strongly related to outsourcing cognitive work than to the mere presence of an AI tool. Those who have an explanation provided, have their own solution path checked, or generate additional practice questions are using the same technical tool differently than someone who takes over a complete solution. AI tools for homework: The greatest learning value is achieved when AI provides hints, counter-questions, and checks, rather than replacing the actual task.

Source: pexels.com

Digital tools and traditional learning work do not have to exclude each other. The crucial factor is whether learners continue to explain, check, remember, and apply themselves.

Which students were particularly affected?

According to the CEPR summary, learning losses were most pronounced in social sciences, followed by STEM subjects and languages. Furthermore, the estimated effects were larger for younger students, high-achieving students, and boys.

The finding regarding high-achieving students is particularly noteworthy. One possible interpretation is that good students can be particularly attracted by efficiency gains: if a task already seems easy, it appears rational to complete it faster with AI. However, the study does not prove why these subgroups were more affected. It initially shows a different magnitude of estimated effects.

What the study does not prove

The headline 'AI makes students dumber' would not be supported by the data. There are several reasons for this.

• No randomization: The study uses a quasi-experimental difference-in-differences design but not random assignment of AI use. Unobserved changes can therefore not be ruled out with the same certainty as in a controlled experiment.
• 'Homework outsourcing' is a behavioral classification: The authors see a pattern that is consistent with outsourcing. This is not identical to a complete record of every prompt and every answer.
• A specific educational context: The sample consists of Chinese students in grades 7 to 12. Results cannot be automatically transferred one-to-one to universities, other countries, other exam formats, or adults.
• Discussion Paper: At the time of publication, the work is a CEPR Discussion Paper or SSRN preprint, not a completed peer-reviewed journal publication. Methodology and interpretation may still change in the scientific process.
• AI usage evolves rapidly: Learners, teachers, and providers are changing their strategies. A system that directly outputs solutions is didactically different from a tutor mode that asks step-by-step questions.

A meaningful counterpoint comes from a study published in Scientific Reports in 2025 with 148 Chinese engineering students. There, more than half reported positive effects on learning efficiency, initiative, and creativity; perceived academic performance remained largely unchanged for many. This study is not directly comparable methodologically or in terms of target group, but it shows why 'generative AI in education' should not be reduced to a single effect.

What students can derive from the study

The most obvious consequence is not to avoid AI completely, but to keep one's own learning process visible. A good rule of thumb is: AI may reduce friction, but not remove the entire thinking process. Using ChatGPT effectively helpful. Especially when learning, a tutor workflow with follow-up questions, small steps, and personal application is more effective than a pure answer workflow.

Source: pexels.com

A more sensible AI learning strategy lets the student work themselves: first own attempt, then targeted hint, then re-solving without aids.

What schools and teachers can learn from this

The study questions traditional homework as a performance indicator when generative AI is freely available. A high homework score may say less about whether a student has actually mastered the material. Therefore, assessment formats where personal competence becomes visible are becoming more important.

• Short closed-book checks: Small tests can show whether homework knowledge can be recalled without AI.
• Oral explanations: Students should be able to explain their own solution path, key terms, or decisions made.
• Evaluating intermediate steps: Notes, drafts, calculation paths, and reflections make the learning process more visible than just a finished product.
• Define AI usage transparently: Teachers can specify what AI is allowed for – for example, for feedback, exercises, or language correction – and what it is not.
• Teaching tutor usage Students need concrete examples of how to ask for help without outsourcing the entire task.

A blanket ban on AI does not follow from the data. The stronger approach is to design tasks and assessments in such a way that independent understanding continues to count and AI support is used where it complements learning processes.

Why the study is currently receiving so much attention

The work hits a sore spot in the current educational debate: Generative AI rapidly improves visible productivity. Schools, parents, and learning platforms may therefore initially gain the impression that learning is becoming more efficient. However, if actual competence grows more slowly or even declines, this difference only becomes apparent in independent assessments.

The World Bank already picked up the paper in June 2026 as a warning signal for human capital. The research in the field of Economics of Education was also presented at the NBER Summer Institute 2026. This does not mean that the results are already considered a definitive scientific consensus. However, it shows that the question of the quality of AI-supported learning has now moved beyond individual classrooms.

FAQ

Conclusion

The Generative AI Learning Penalty study provides one of the clearest warning signals to date that better AI-assisted homework does not automatically mean better learning. In the data from 26,811 Chinese students, homework scores increased, completion time decreased – and at the same time, independent exam performance deteriorated significantly.

The crucial limitation is as important as the headline: the greatest losses were concentrated in usage patterns consistent with outsourcing homework to AI. Those who continued to invest comparable amounts of time in the tasks showed only minor losses. For students and schools, this implies not a "AI yes or no," but a more precise question: Is the AI taking over the thinking – or is it helping to think better?