AI Homework and Exam Scores: Homework Up 18%, Exams Down 20%

A large new study of AI homework and exam scores found the gap many teachers suspected: when students started using generative AI, their homework scores went up 18% and their homework time fell 30%, but their scores on closed-book monthly exams fell 20% within six months. On high-stakes entrance exams, scores fell 18% and 24%, and the full penalty took about two years to show. The study, a June 2026 discussion paper by David Strömberg, Victor Lei and Yanhui Wu, followed 26,811 students in grades 7 to 12 in China over 30 months. The most useful finding for parents is about who lost. The losses were concentrated among roughly 80% of AI users whose homework was both unusually fast and unusually good, a pattern consistent with outsourcing. Students who kept spending about as much time on homework as non-users lost little. Here’s the study read carefully, how it fits other research, and the one red flag it gives you.

Key Takeaways

• AI adoption raised homework scores 18% and cut completion time 30%, but lowered monthly closed-book exam scores 20% within six months.
• High-stakes entrance-exam scores fell 18% and 24%, “with the full penalty emerging only after about two years.”
• Losses were concentrated among about 80% of AI users whose behavior looked like homework outsourcing; users who kept normal homework time “experience small learning losses.”
• The paper is a CEPR discussion paper, not yet peer-reviewed, and comes from one country’s school system.
• A 2025 PNAS experiment in Turkey found the same shape: AI raised practice scores, then students did worse once AI was removed, unless the AI was designed to give hints instead of answers.

What the AI Homework and Exam Scores Study Found

“The Generative AI Learning Penalty: Evidence from Chinese Secondary Education” is CEPR Discussion Paper 21577, dated June 2026. A discussion paper is a working paper circulated for comment; it hasn’t passed journal peer review.

The data. Thirty months of records on 26,811 students in grades 7 to 12, combining monthly closed-book exams, high school and college entrance exams, and homework scores and completion times across nine subjects.

The method. Students adopted AI at different times. The authors used a difference-in-differences design, which compares how outcomes change for students after they start using AI with how outcomes change over the same period for students who haven’t started yet. That separates the effect of adoption from general trends, though it can’t rule out every difference between early and late adopters.

The results, from the abstract:

• Homework scores up 18%; completion time down 30%.
• Monthly exam scores down 20% within six months.
• Entrance-exam scores down 18% and 24%, with the full penalty emerging after about two years.
• The largest losses were in social science subjects, then STEM and languages, and were “especially large for junior students, high-achieving students, and boys.”

The finding that matters most: losses “are concentrated among roughly 80% of AI users whose behavior is consistent with homework outsourcing, as indicated by exceptionally short homework completion time coupled with high homework scores.” Meanwhile, “AI users who maintain similar homework completion time as non-AI users experience small learning losses.”

One caution on the numbers: the abstract reports percentage changes without saying exactly what they’re percentages of. The full paper defines them, and the precise magnitude may read differently in context.

Why Homework Scores Rose While Learning Fell

The mechanism isn’t mysterious. Homework exists for practice. Retrieving a formula, working a problem, drafting an argument: the effort is the learning. When an AI produces the answer, the homework score measures the AI, and the practice doesn’t happen.

That’s why the time signal works. A student who finishes math homework in a third of the usual time with a near-perfect score either got much better overnight or didn’t do the thinking. Across thousands of students, the study found the second explanation fits most cases.

The two-year delay on entrance exams is the uncomfortable part. Monthly tests caught the problem within six months, but the full cost on the exams that matter most took about two years to appear. A parent watching homework grades would have seen improvement the whole time.

How This Fits Other Research

The China study is observational. A randomized experiment points the same way.

In a 2025 PNAS study, Hamsa Bastani and colleagues gave nearly a thousand high school math students in Turkey access to one of two GPT-4 tutors: “GPT Base,” a standard ChatGPT-like interface, and “GPT Tutor,” prompted to give teacher-designed hints instead of answers. With access, grades on practice problems rose 48% for GPT Base and 127% for GPT Tutor. But when access was taken away, the GPT Base students “actually perform worse than those who never had access (17% reduction in grades).” Students had used it “as a ‘crutch.’” The safeguards in GPT Tutor “largely mitigated” the harm.

The OECD’s Digital Education Outlook 2026, as reported by Hürriyet Daily News, drew the same lesson: AI can raise short-term performance, but that “does not necessarily translate into lasting learning,” while tools designed for learning, with guided questions and hints, can support more durable gains.

And in a randomized study of 117 university students published in the British Journal of Educational Technology, Yizhou Fan and colleagues found the ChatGPT group improved essay scores the most, but “their knowledge gain and transfer were not significantly different” from other groups. They warned of “metacognitive laziness.”

There’s a reason students don’t notice. Physics students in a 2019 PNAS study by Louis Deslauriers and colleagues learned more from active, effortful classes but felt they learned less, taking effort “to signify poorer learning.” Easy feels like learning. It often isn’t.

Three Studies, One Pattern

Study	Who	AI condition	While using AI	When tested without AI
Strömberg, Lei & Wu, 2026 (CEPR discussion paper)	26,811 students, grades 7–12, China	Real-world adoption, self-chosen	Homework scores +18%, time −30%	Monthly exams −20%; entrance exams −18% and −24%
Bastani et al., 2025 (PNAS)	~1,000 high school math students, Turkey	GPT Base vs. hint-only GPT Tutor	Practice grades +48% (Base), +127% (Tutor)	Base: −17% vs. never-users; Tutor: harm largely mitigated
Fan et al., 2024 (BJET)	117 university students	ChatGPT vs. human expert vs. tools vs. none	Largest essay-score gains	No difference in knowledge gain or transfer

Across very different settings, the shape repeats: performance with AI goes up, learning measured without AI stays flat or falls, and design (hints vs. answers) or behavior (time spent) separates the students who are harmed from those who aren’t.

What Parents Can Do

Watch homework time, not just homework grades

This is the study’s gift to parents. If your child’s homework suddenly takes much less time and the scores go up at the same time, ask how. It isn’t proof of outsourcing, but it’s the exact pattern the study flagged.

Ask for the thinking, not the answer

“Show me how you got that” works for any subject. A child who did the work can walk you through it. One who copied an AI’s answer usually can’t. For practical scripts, see our guide on what to do when kids use ChatGPT for homework.

Steer toward hint-mode, not answer-mode

The PNAS experiment showed the design of the AI matters. Many chatbots now offer study or tutor modes that ask questions instead of giving answers. Set those up, or teach your child to prompt, “Don’t give me the answer, give me a hint.”

Expect the effort to feel worse

Tell your kid what Deslauriers found: hard practice feels like you’re learning less, even when you’re learning more. That framing helps a frustrated student keep going.

What not to do

Don’t rely on homework grades as proof that things are fine. In the China study, homework grades rose while exam scores fell for months. And don’t ban AI so completely that your child never learns to use it in hint mode; the difference between harm and help in these studies was how the tool was used. Our article on the AI illusion of learning goes further into that gap.

What to Watch For Over the Next 3 Months

• Week 4: Has homework time dropped sharply since your child started using AI? Can they explain one recent assignment step by step?
• Month 2 red flags: Higher homework grades alongside lower quiz or test grades; homework done in minutes; your child unable to start an assignment without opening a chatbot.
• Month 3 self-check: Compare the most recent closed-book test with the homework grades in the same subject. If the gap is widening, sit down with your child and the teacher to talk about how AI is being used.

Frequently Asked Questions

Does using AI for homework lower test scores?

In a study of 26,811 Chinese students in grades 7 to 12, adopting AI raised homework scores 18% but lowered closed-book monthly exam scores 20% within six months. The losses were concentrated among students whose homework was unusually fast and high-scoring, a pattern consistent with outsourcing. Students who kept normal homework time lost little.

Is this study peer-reviewed?

Not yet. It’s a CEPR discussion paper from June 2026, which is a working paper circulated for comment. Its findings line up with a peer-reviewed 2025 PNAS field experiment in Turkey, which found AI raised practice scores but lowered performance once AI access was removed.

How can I tell if my child is outsourcing homework to AI?

The study’s signal is homework that suddenly takes far less time while scores rise. Asking your child to explain their work step by step is the simplest check. Falling test scores in the same subject are another warning sign.

Is there a safe way to use AI for homework?

The research suggests yes. In the PNAS study, a tutor designed to give hints rather than answers largely avoided the learning loss. In the China study, students who kept spending normal time on homework had small losses. Use AI to explain, quiz and give hints, not to produce answers.

Why did it take two years for the full effect to show?

The paper doesn’t fully explain it. A likely reason is that entrance exams draw on knowledge built over years, so skipped practice accumulates. Monthly exams caught a drop within six months, which is why closed-book tests are a better early warning than homework grades.

---

About the author

Ricky Flores is the founder of HiWave Makers and an electrical engineer with 15+ years of experience building consumer technology at Apple, Samsung, and Texas Instruments. He writes about how kids learn to build, think, and create in a tech-saturated world. Read more at hiwavemakers.com.

---

Sources

1. Strömberg, D., Lei, V., & Wu, Y. (2026). “The Generative AI Learning Penalty: Evidence from Chinese Secondary Education.” CEPR Discussion Paper 21577 (not peer-reviewed). https://econpapers.repec.org/RePEc:cpr:ceprdp:21577
2. Bastani, H., Bastani, O., Sungu, A., Ge, H., Kabakcı, Ö., & Mariman, R. (2025). “Generative AI without guardrails can harm learning: Evidence from high school mathematics.” Proceedings of the National Academy of Sciences. https://doi.org/10.1073/pnas.2422633122
3. Hürriyet Daily News. (2026, February). “OECD warns AI boosts grades but not lasting learning.” https://www.hurriyetdailynews.com/amp/oecd-warns-ai-boosts-grades-but-not-lasting-learning-218598
4. Fan, Y., Tang, L., Le, H., et al. (2024). “Beware of metacognitive laziness: Effects of generative artificial intelligence on learning motivation, processes, and performance.” British Journal of Educational Technology. https://doi.org/10.1111/bjet.13544
5. Deslauriers, L., McCarty, L. S., Miller, K., Callaghan, K., & Kestin, G. (2019). “Measuring actual learning versus feeling of learning in response to being actively engaged in the classroom.” Proceedings of the National Academy of Sciences. https://doi.org/10.1073/pnas.1821936116
6. Common Sense Media. (2026). “Teens in the AI Era: Schoolwork and Skills That Matter.” https://www.commonsensemedia.org/research/teens-in-the-ai-era-schoolwork-and-skills-that-matter