No AI Homework Rules, Tested Against the Research

In February 2026, the Screenagers blog published three no AI homework rules that have since spread through school newsletters and parent group chats: no AI for writing, editing or annotating; no AI for problem solving; and no AI-made study materials. The post, by physician Delaney Ruston, borrows the rules from neuroscientist Jared Cooney Horvath’s book The Digital Delusion. It cites no specific studies. So I did what an engineer does with a spec that arrives without test data: I checked each rule against the published research. Short version? Two of the three hold up well. The third is half right, and the half that is wrong matters for how your kid studies for tests.

Key Takeaways

• Rule 1 (no AI for writing) is well supported. A 2020 meta-analysis of 56 experiments found that writing about school content improves learning in grades 1–12, with an effect size of 0.30.
• Rule 2 (no AI for problem solving) is supported for answer-giving chatbots. In a field experiment with nearly 1,000 high schoolers, students who practiced with plain GPT-4 later scored 17% worse than students who never had it. A hint-only version largely avoided the harm.
• Rule 3 (make your own study materials) is half right. Making your own flashcards helps. But summaries are a weak study method no matter who writes them, and the strongest method, self-quizzing, works even with questions someone else wrote.
• Kids will often say the AI way “feels like” learning more. Research on effort shows that feeling is an unreliable guide.
• A better household rule: the child must do the generating and the remembering. AI may give hints, never finished answers.

What the No AI Homework Rules From Screenagers Say

The no-AI homework rules are three limits on how children use generative AI for schoolwork, proposed by Jared Cooney Horvath and endorsed by Delaney Ruston, MD, on the Screenagers blog on February 3, 2026. In their wording:

1. “No AI for writing, editing or annotating. Essays should be planned and drafted offline.”
2. “No AI for problem solving: if your child gets stuck encourage them to ask you, a teacher, or a peer for help.”
3. No AI-generated study material: “Have your children make their own flashcards, study guides and summaries.”

Horvath’s reasoning is that “when novice children use these tools they short circuit effort and skip the thinking process entirely.” He goes further and suggests families “consider banning it from your home, at least when it comes to schoolwork.”

The post says “research suggests” AI can undermine learning but names no studies. That is not a criticism of the advice so much as a gap this article tries to fill.

Rule 1: No AI for Writing. Verdict: Strongly Supported

The research case for this rule rests on two well-replicated findings.

The first is writing-to-learn. Graham, Kiuhara and MacKay pooled 56 true and quasi-experiments with students in grades 1 to 12 in a 2020 meta-analysis in the Review of Educational Research. Students who wrote about content in science, social studies or math learned more than those who did not, with an average effect size of 0.30. The effect was similar across subjects and across elementary, middle and high school. In plain terms: the act of composing sentences about photosynthesis is part of how a kid learns photosynthesis.

The second is the generation effect. People remember information they produce better than information they read. A 2007 meta-analysis in Memory & Cognition by Bertsch and colleagues summarized 445 effect sizes from 86 studies and found an average benefit of 0.40 standard deviations for generating over reading.

Here is the mechanism, stripped of jargon. When a child drafts an essay, they must retrieve facts from memory, decide how ideas connect and notice what they don’t understand. Each of those steps strengthens the memory. When a chatbot drafts the essay, the child reads a finished product. Reading is the weaker condition in both bodies of research.

Where the rule goes beyond the evidence is editing. I found no strong research on whether, say, a grammar suggestion after a child finishes a draft harms learning. The concern is plausible (a kid who never fixes their own comma splices may not learn to), but it is a judgment call rather than a finding. For a deeper look at this, see our piece on the AI illusion of learning in homework.

Rule 2: No AI for Problem Solving. Verdict: Supported, With a Design Caveat

The best single study here is a field experiment by Bastani and colleagues, published in PNAS in 2025. Nearly a thousand high school math students in Turkey were given one of two AI tutors during practice sessions:

• GPT Base, which worked like standard ChatGPT.
• GPT Tutor, prompted to give teacher-designed hints instead of answers.

During practice, both helped a lot: grades improved 48% with GPT Base and 127% with GPT Tutor. Then the AI was taken away. On the unaided exam, students who had used GPT Base scored 17% lower than students who never had AI at all. The authors describe students using GPT-4 as a “crutch.” The hint-based tutor largely avoided that drop.

That result supports Horvath’s rule for the AI most kids actually have, a general-purpose chatbot that will hand over a full worked solution if asked. It also shows the rule is stricter than the evidence requires. The harm came from answer-giving, not from AI as such.

A 2025 randomized trial by Kestin and colleagues in Scientific Reports points the same way. In a college physics course, students using a custom AI tutor built on teaching best practices learned more, in less time, than students in an active-learning class. Note the limits: these were college students, the tutor was carefully designed by instructors, and one course is one course. It does not show that a 10-year-old with an open chatbot will learn more. It does show that tool design is the variable that matters.

So “ask a person” remains excellent advice. People tend to give hints. Chatbots, by default, give answers.

Rule 3: Make Your Own Study Materials. Verdict: Half Right

This is the rule where the details change what your kid should do on Thursday night.

The most cited review of study techniques is a 2013 monograph by Dunlosky and colleagues in Psychological Science in the Public Interest. They rated ten techniques. Two got “high utility”: practice testing (quizzing yourself) and distributed practice (spacing study over days). Five got “low utility,” including summarization, highlighting and rereading.

That creates two problems for rule 3 as written:

• Summaries are weak whether a human or a chatbot writes them. Telling a child to write their own summary instead of asking AI for one swaps a weak method for a slightly better weak method. Self-quizzing would beat both.
• Retrieval practice works even when someone else writes the questions. In Roediger and Karpicke’s 2006 experiments in Psychological Science, students who took recall tests remembered much more a week later than students who restudied the same material, even though restudying made them feel more confident. The benefit comes from pulling the answer out of memory, not from authoring the question.

Making your own flashcards still has value. Writing a good question requires understanding the material, and that is generation again. But a kid who spends 45 minutes making beautiful cards and five minutes quizzing has the ratio backward. If an AI-made practice quiz gets a child to spend the full hour retrieving, that may be the better study session. The non-negotiable part is that the child answers from memory before looking.

The Rules Side by Side

Rule	What the research says	Verdict	Sharper version for home
No AI for writing, editing or annotating	Writing-to-learn (ES 0.30, 56 experiments) and the generation effect (0.40, 86 studies) both favor kids producing their own text	Strongly supported for drafting; editing is a judgment call	Child plans and drafts alone. Any feedback comes after a complete draft, as questions, not rewrites
No AI for problem solving	Answer-giving GPT-4 boosted practice but cut later exam scores 17%; hint-only tutor avoided most of the harm	Supported for general chatbots; stricter than needed for hint-only tools	No full solutions, ever. A hint after a real attempt is acceptable if the tool or teacher allows it
Make your own flashcards, guides, summaries	Practice testing and spacing are high utility; summarization is low utility; testing works with others’ questions	Half right	Child must answer from memory before checking. Prioritize self-quizzing over summaries, regardless of who wrote the questions

How to Put These Rules to Work at Home

Write the “who generated it” test on the fridge

Every homework session can be checked with one question: who did the generating and who did the remembering? If the answer is “the AI,” the session probably did less than it looked like. This is simpler to enforce than a list of banned apps, and it works across tools.

Expect pushback that sounds like evidence

Your child may say the AI way helps them understand better. Take that seriously, and then know this: in a 2019 PNAS study by Deslauriers and colleagues, physics students randomly assigned to active learning learned more but felt they had learned less than peers in polished lectures. The authors concluded that students read the effort of active learning as a sign of poorer learning. Effortful study feels worse while it is working.

Hint ladders, not answer keys

When your child is stuck on a math problem, try a three-step ladder before anyone (human or machine) shows a solution: first ask what the problem is asking; then point to the relevant example in their notes; then give one next step only. This copies what made the PNAS hint-only tutor safer. More on the trade-offs is in our guide to when AI homework tools help and when they hurt.

Turn study time into quiz time

Ask your child to close the book and tell you, or a blank page, everything they remember about the topic. Then check. Repeat two days later. This is practice testing plus spacing, the two high-utility techniques, and it costs nothing.

What not to do

Don’t treat a ban as the finish line. A rule that kids cannot use AI at home, enforced by checking browser history, misses the more useful skill: knowing which kind of help builds your brain and which replaces it. Kids will meet AI tools at school and at work. A teen who understands why answer-giving hurts learning will make better choices when you are not watching.

What to Watch For Over the Next 3 Months

• Week 4: Look at a draft your child wrote alone. Is it rougher than their AI-assisted work? That is expected and fine. Rough and theirs beats polished and borrowed.
• Month 2 red flags: Homework grades stay high but quiz and test scores drop, or your child can’t explain an answer they turned in. That gap is the pattern the Bastani study found.
• Month 3 self-check: Ask your child to explain one concept from each subject without notes. If they can, your rules are doing their job, whatever they are.

Frequently Asked Questions

Are no AI homework rules backed by research?

Mostly. Writing about content and generating answers yourself are well-established learning boosters, and a 2025 PNAS experiment found answer-giving chatbots hurt later exam performance. The weakest rule is the one about making your own summaries, because summarizing is a low-utility study method regardless of who does it.

Should I ban AI from homework entirely?

It is a reasonable default for younger children with general chatbots, since those give full answers by default. For older kids, the research supports a narrower rule: no finished answers or drafts, ever, but structured hints after a real attempt may be fine, particularly in tools designed by teachers.

Is it okay to use AI to make practice quizzes?

Probably, if the child answers from memory before checking. The benefit of practice testing comes from retrieval, not from writing the questions. Check that the quiz is accurate, because AI-generated questions can contain errors, and keep some self-made cards in the mix.

My kid says AI explains things better than the teacher. Is that bad?

Not necessarily. The problem is not explanation, it is replacement. If your child reads an AI explanation and then solves new problems unaided, that can help. If the AI solves the assigned problems, it hurts. Research also shows students often mistake easy-feeling study for effective study.

What age should these rules start?

As soon as a child has access to a chatbot for schoolwork. Younger children are the “novices” Horvath worries about most, since they have the least background knowledge to judge or correct an AI’s output.

---

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. Ruston, D. (2026). “Should Kids Use AI for Homework? Here Are the Rules I Recommend.” Screenagers blog, February 3, 2026. https://www.screenagersmovie.com/blog/should-kids-use-ai-for-homework
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. Kestin, G., Miller, K., Klales, A., Milbourne, T., & Ponti, G. (2025). “AI tutoring outperforms in-class active learning: an RCT introducing a novel research-based design in an authentic educational setting.” Scientific Reports. https://doi.org/10.1038/s41598-025-97652-6
4. Graham, S., Kiuhara, S. A., & MacKay, M. (2020). “The Effects of Writing on Learning in Science, Social Studies, and Mathematics: A Meta-Analysis.” Review of Educational Research. https://doi.org/10.3102/0034654320914744
5. Bertsch, S., Pesta, B. J., Wiscott, R., & McDaniel, M. A. (2007). “The generation effect: a meta-analytic review.” Memory & Cognition. https://doi.org/10.3758/BF03193441
6. Dunlosky, J., Rawson, K. A., Marsh, E. J., Nathan, M. J., & Willingham, D. T. (2013). “Improving Students’ Learning With Effective Learning Techniques.” Psychological Science in the Public Interest. https://doi.org/10.1177/1529100612453266
7. Roediger, H. L., & Karpicke, J. D. (2006). “Test-enhanced learning: taking memory tests improves long-term retention.” Psychological Science. https://doi.org/10.1111/j.1467-9280.2006.01693.x
8. 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