Homework scores up, exam scores down: the 30-month school records study
In the records of 26,811 secondary school students in one county in China, taking up AI went with homework scores 18% higher and monthly exam scores 20% lower. Students who kept spending a normal amount of time on homework had only small losses. This is a 2026 working paper based on school records. It is not an experiment and it is not peer reviewed, so it shows a pattern, not a proven cause.
The study in one table
| Title | The Generative AI Learning Penalty: Evidence from Chinese Secondary Education |
|---|---|
| Who | Stromberg, Lei and Wu |
| Status | Working paper. CEPR Discussion Paper 21577, June 2026. Not peer reviewed |
| People | 26,811 students in grades 7 to 12, in one county in central China |
| Design | School records followed for 30 months. Students were not assigned to groups |
| Source | ideas.repec.org/p/cpr/ceprdp/21577.html |
What the records show
| Measure | After students took up AI |
|---|---|
| Homework scores | 18% higher |
| Time to finish homework | 30% shorter |
| Monthly exam scores | 20% lower after six months |
| Entrance exam scores | 18% to 24% lower, with the full effect after about two years |
A note on these figures. We read these figures in the paper's abstract and in the authors' own summary column. We have not read the full paper. Treat them as reported, and check the source before you quote them.
The detail that matters most
About four in five students using AI finished their homework in under 50 minutes. Students who used AI but kept spending a normal amount of time on homework had similar exam scores to non-users. The paper's abstract calls their losses small.
That suggests one reading. The tool was not the problem by its mere presence. The loss went with the time and effort that students stopped putting in.
Why "not an experiment" matters
In an experiment, researchers decide at random who gets the AI. That makes the groups alike in everything else, so a difference in results can be put down to the AI.
Here nobody was assigned. The researchers studied what had already happened. So the students who used it may have differed from the others in ways the records do not capture.
- Who chose AI. A student who is already losing interest might reach for a shortcut and also do worse in exams. The shortcut would then be a sign, not the cause.
- What else changed. Over 30 months, teachers, timetables and exams change too.
- How use was measured. Teachers handed out surveys, and students looked up the dates they registered for AI tools. That relies on students reporting honestly.
Careful statistics can reduce these doubts in records data. They cannot remove them the way random assignment does.
What it did not show
- Not proof that AI caused the lower scores. It is a pattern in records.
- Not a result for every school. One county, in one country, with its own exam system.
- Not a finished paper. Working papers are often revised before a journal publishes them, and numbers can change.
- Not that all AI use costs marks. Students who still put in the time had only small losses.
Why it still deserves attention
It is large, and it is long. Most experiments on AI and learning last a few sessions. This is the only source we know of that covers more than two years. And it agrees with an experiment: in a randomised field trial with about 1,000 students, practice scores rose with a plain chatbot and exam scores were 17% lower. When records and an experiment point the same way, each makes the other more believable.
A fair one-line summary. "A large working paper on Chinese school records linked AI homework shortcuts to lower exam scores. Students who kept spending time on homework had only small losses. It is not an experiment."
What to do with this
Watch the clock in the other direction. If homework that used to take an hour now takes fifteen minutes, ask who did the thinking. Attempt each question first, then use the AI to check and explain. Parents can agree this as a house rule.
Questions
Does using AI for homework lower exam scores?
In these school records, AI use went with monthly exam scores 20% lower after six months. The study cannot prove the AI was the cause, because students were not assigned at random.
Did every student who used AI do worse?
No. Students who kept spending a normal amount of time on homework had only small losses.
Is this study peer reviewed?
Not as of 13 November 2026. It is a discussion paper from June 2026.
Why did homework scores go up?
The likely reason is that the AI was doing part of the homework. A homework score then shows what the student and the AI can do together, not what the student can do alone in an exam.