What one recent study of AI and homework does and doesn’t tell teachers and homeschool parents, and one idea of mine for what to try.
The scary part isn’t that AI can do the homework. It’s that the homework score went up while the exam score went down.
What the study found
A working paper from economists David Strömberg, Victor Lei and Yanhui Wu (CEPR Discussion Paper 21577, published June 2, 2026) followed 26,811 students in grades 7 to 12 in one county in central China for 30 months. The researchers had weekly homework records and monthly closed-book exams, so they could compare what happened after students started using AI with what happened to students who hadn’t started yet.
Six months after students began using AI, homework scores had risen 18% and the time per assignment had dropped from about 64 minutes to about 45. Monthly closed-book exam scores had fallen 20%. Both figures are relative to the average score before students adopted AI, not percentage points. The authors also estimate that the effect on high-stakes entrance exams grows over about two years, to 18% of the pre-AI average for the college entrance exam and 24% for the high-school entrance exam. They note this evidence is less rigorous than the monthly-exam results, and the entrance-exam gaps they measured directly so far are smaller (roughly 5-7%).
Among students who used AI, the authors found that higher homework scores went with worse exam scores. In their words, “a rapid increase in homework scores sends a warning signal of learning decline.”
How much to trust it
Hold this one loosely.
- It’s a working paper, not peer-reviewed, and it dates from June 2026. It isn’t new this week.
- It’s a before-and-after comparison using school records (a difference-in-differences design), not a randomized trial.
- Students’ AI start dates were self-reported.
- It’s one county in central China, grades 7 to 12, using tools like Doubao and DeepSeek. It doesn’t tell us what would happen in a US classroom or a homeschool. The authors themselves write that “we cannot directly extrapolate effect sizes to other settings, but they should serve as a warning.”
The part worth taking home: outsourcing, not cheating
The paper doesn’t say AI ruins learning across the board. It says the losses are concentrated among roughly 80% of AI users, after several months of use, whose behavior is consistent with outsourcing the homework: exceptionally short homework time and high scores. The authors infer this from the pattern. They didn’t see what any student actually did.
The rest of the AI users kept spending about as much time on homework as students who didn’t use AI, and they had small losses. The authors caution that this doesn’t mean making students spend more time would fix things, because students who choose different ways of using AI may also differ in motivation, information and skill.
So the useful question isn’t “is my student cheating?” It’s “what is the finished page evidence of?”
What the authors suggest
For teachers and parents, the authors suggest watching inputs, such as homework time and study effort, instead of outputs such as homework scores. They also point to frequent short in-class quizzes on the homework material, and to giving more weight to closed-book, in-person assessment. They add that supervised homework time combined with guidance on productive AI use, and information about the learning losses from outsourcing, may be needed.
A separate randomized field experiment points the same way. Bastani and colleagues (PNAS, 2025) studied about 1,000 high-school math students in Turkey in one math course. Students who practiced with an answer-giving AI scored about 17% lower on a later closed-book exam once the AI was taken away, compared with students who never had it. That drop was largely mitigated with a version built with safeguards.
My own interpretation: homework the AI can’t hand over
This part is mine. No study I’ve cited tested it.
If a finished page is no longer good evidence of learning, one response is to design work where the finished page isn’t the only evidence. A few things I’d try:
- Ask the student to explain the work out loud, in their own words, before it counts as done.
- Follow it with one closed-book question the next day, with no tool in reach.
- Ask for the messy middle: what they tried first, what they got wrong, what changed their mind.
In the Auxesis lens, the question is whether AI is augmenting the student’s thinking or replacing it, and whether the work deepens understanding or shortcuts it. The study doesn’t use that frame. I find it a useful way to read its numbers.
Try this
When a finished page comes in, what’s your first check: how long it took, or one closed-book question? Tell me which, and what you saw. (Please don’t share names or identifying details about your kids.)
Continue this in the community: https://auxesislab.com/community/
Sources
- Strömberg, Lei and Wu, “The Generative AI Learning Penalty: Evidence from Chinese Secondary Education,” CEPR Discussion Paper 21577 (June 2, 2026): https://cepr.org/publications/dp21577
- The authors’ VoxEU column: https://cepr.org/voxeu/columns/generative-ai-learning-penalty-secondary-school
- Bastani et al., “Generative AI without guardrails can harm learning: Evidence from high school mathematics,” PNAS (2025): https://www.pnas.org/doi/10.1073/pnas.2422633122
Leave a Reply