The 'Moderate Use' Rule: Setting an AI Dose at Home

The most useful number to come out of PISA 2025 isn’t a ranking. It’s the shape of a curve. Students who used AI chatbots for schoolwork “moderately,” roughly once a month to once or twice a week, recorded higher average science scores than both students who barely used AI and students who used it almost every day. That single finding is what turns vague worry about AI use limits kids home into something you can actually set: a weekly dose, with named tasks. Here’s how to build one, what the OECD data supports, and what it doesn’t.

Key Takeaways

• PISA 2025 (released September 8, 2026) tested “over 760,000 students representing 33 million 15-year-olds across 91 countries and territories.”
• Moderate AI users, about once a month to once or twice a week, scored higher in science than both rare users and near-daily users. Almost-daily use was reported by about one in five OECD students.
• Students who never or almost never used AI to draft writing scored 509 on the science test; near-daily drafters scored 481. The gap was wider for AI summarizing.
• The OECD is explicit that this is correlation: these relationships “do not necessarily imply a negative impact of AI use on science performance.”
• Students who reported learning at school how to assess AI-generated information tended to score slightly higher, especially frequent users. Teaching evaluation appears to matter more than restricting access.

What PISA 2025 actually found about AI use

PISA is the OECD’s Programme for International Student Assessment, run every three years on 15-year-olds. The 2025 round, released September 8, 2026, covered over 760,000 students across 91 systems, and for the first time asked detailed questions about how students use AI chatbots for schoolwork.

The headline results were grim on their own terms. Education International’s summary of the report notes mathematics “fell by 22 points between 2015 and 2025 across OECD countries” and reading fell 28 points, with science “broadly stable.” Roughly one in five 15-year-olds is now a low performer across all three domains.

The AI findings are more interesting than the headlines suggested. Yes, the top-line association is negative: OECD reports that students who don’t use AI chatbots for schoolwork generally score higher than students who do. Fortune’s September 11, 2026 coverage notes that “86% of students across OECD countries had used AI chatbots for at least one of the schoolwork purposes examined,” and that “about one in five students reported using it almost every day.”

But the frequency curve isn’t a straight line down. Per analysis of the OECD data, “among students who did use AI for those tasks, moderate users — roughly once a month to twice a week — scored higher than students who used it only a few times a year or almost every day.” Weekly users who used AI for summarizing assigned reading or doing preliminary research on a new topic “tended to perform at about the same level as AI nonusers after accounting for socioeconomic status.”

Task-level numbers make it concrete. Students who said they never or almost never use AI to draft texts for writing assignments scored 509 in science; those using it for drafting every day or almost every day scored 481. For summarizing what they’d read, the gap ran close to 30 points.

And one finding gets almost no coverage but matters most for parents: students who reported learning at school how to assess AI-generated information tended to have slightly higher science performance, an effect that showed up especially among frequent users. Those opportunities were reported more often by socioeconomically advantaged students.

The correlation caveat, stated plainly

The OECD’s own framing: these relationships “do not necessarily imply a negative impact of AI use on science performance, but may reflect a complex mix of who adopts AI and how they use it.”

That sentence deserves respect rather than a wave-past. Students who reach for a chatbot every day for every assignment may already be struggling. Students who use it twice a week may be the organized ones. PISA cannot separate those stories, and the reverse-causation reading is at least as plausible as the forward one.

There’s a second caveat specific to the U.S. numbers. Chalkbeat reported on September 8, 2026 that over 40% of U.S. students didn’t complete the accompanying questionnaires, which limits what can be analyzed about student background. U.S. results carry a flag for incomplete school response rates.

What survives the caveats is modest but real: there is no evidence that maximizing AI use helps, and there is evidence that teaching evaluation skills correlates with better outcomes. That’s enough to build a household rule on.

The task table: what to allow, what to limit, what to hold back

This table maps the PISA task categories onto a home policy. The “PISA signal” column names what the data actually showed; the rule column is a judgment call built on it.

Task	PISA 2025 signal	Home rule	Why
Summarizing assigned reading	Near-30-point science gap for near-daily use; weekly users matched non-users	Once a week, after they’ve read it	The summary is the thinking. Use AI to check their own summary, not replace it
Drafting a writing assignment	509 (never/rarely) vs. 481 (near-daily)	Not for graded writing; allowed for outlines after their own brainstorm	Drafting showed the clearest negative gradient of any task
Preliminary research on a new topic	Weekly users performed about level with non-users	Allowed weekly, with one rule: every claim gets a second source	Lowest-risk category in the data
Explaining a concept they’re stuck on	Not separately reported; tutoring RCTs show gains with guardrails	Unlimited, if the AI asks questions instead of giving answers	Bastani et al. (2025) found guarded tutors avoided the harm
Checking their own answer after solving	Not separately reported	Encouraged, always after their attempt	Turns AI into a verifier, which builds the checking habit
Doing the problem set for them	Not measured; Common Sense found 63% of teen AI users get an answer	Never	This is the behavior the whole budget exists to prevent
Evaluating whether AI output is right	Associated with slightly higher science scores, especially for frequent users	Do this weekly, out loud, together	The only AI activity with a positive association in the data
Companion-style chatting	Outside PISA’s scope; regulated separately in several U.S. states	Separate conversation, separate rules	Different risk profile entirely

Building the weekly AI budget

Write the number down, then write the tasks down

“Moderate” in the PISA data means roughly one to two times a week. So start there: two AI sessions per school week for schoolwork, each with a named task from the table. Write it on paper, on the fridge. The reason to write it is not enforcement; it’s that a specified rule gets followed and a vague one gets negotiated nightly.

Pair every session with a verification step

The single positive association in the PISA data was learning to assess AI output. So make it structural: no AI session ends without one check. “Which part of this do you know is right? How?” If they can’t answer, the session doesn’t count as done. This takes two minutes and it’s the highest-value two minutes in the whole routine.

Put drafting behind their own first draft

The 509-versus-481 drafting gap is the sharpest task-level signal PISA reported. The workable compromise isn’t a ban, it’s an order of operations: their outline first, on paper, then AI for feedback on that outline. This preserves the generative step, which is where the learning sits, and still gets them the tool’s benefit.

Track it for four weeks, then adjust once

Use a tally on the fridge or a note on their phone. Count sessions, not minutes. After four weeks, look at the pattern with them: which sessions felt useful, which felt like avoidance? The 2010 habit-formation study by Lally, van Jaarsveld, Potts, and Wardle in the European Journal of Social Psychology found a median of 66 days to reach maximum automaticity, with a range from 18 to 254 days. Four weeks is enough to see a pattern, not enough to have a habit. Expect to keep at it.

What not to do

Don’t set the budget as a screen-time number. The U.S. Department of Education’s August 20, 2026 Dear Colleague letter asked districts to judge technology by “purposeful, instructionally sound” use “contributing to measurable student outcomes” rather than minutes. That logic applies at home: two hours of AI-assisted concept practice and two hours of AI-written essays are not the same thing, and a timer can’t tell them apart.

What to Watch For Over the Next 3 Months

• Week 4: Are sessions being used on the named tasks, or drifting toward “just check my whole assignment”? Drift is the thing to catch early.
• Month 2 red flags: Your kid can’t explain the work they turned in, or the verification step has quietly disappeared. Both mean the budget is a formality.
• Month 3 self-check: Ask them to summarize a chapter without AI, then with it, and compare. If their own version is getting worse, cut the summarizing allowance.

Frequently Asked Questions

What exactly did PISA 2025 say about AI and test scores?

Students who don’t use AI chatbots for schoolwork generally scored higher than those who do, but the relationship isn’t linear: moderate users (about once a month to twice a week) scored higher than both rare users and near-daily users. The OECD stresses this is association, not proof of causation.

How much AI use is “moderate”?

In the PISA data, moderate meant roughly once a month to once or twice a week for a given schoolwork task. About one in five OECD students reported using AI almost every day, which is the group that scored lowest in science.

Should I just ban AI for homework?

The data doesn’t support a ban as the optimal choice. Non-users didn’t score highest; moderate users did. And the one factor positively associated with science performance was learning to assess AI-generated information, which requires some supervised use.

Which AI task is most harmful?

Drafting showed the clearest negative gradient: 509 in science for students who never or rarely used AI to draft writing assignments, versus 481 for near-daily use. Summarizing assigned reading showed a gap approaching 30 points.

Does this apply to younger kids too?

PISA tests 15-year-olds, so strictly the findings apply to that age group. For younger children, several districts have gone further than a budget. New York City adopted a one-year moratorium on student-facing generative AI for grades 2-K through 8 on September 2, 2026, affecting nearly 600,000 students.

What if my kid’s school hasn’t taught any of this?

That’s common. Common Sense Media’s August 2026 survey found “only 27% of teens say a teacher has ever talked with them about how to use AI safely; 73% say they have not.” The weekly verification step in this article is the home version of that missing lesson.

---

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. OECD. (Sept 8, 2026). “PISA 2025 Results (Volume I).” https://www.oecd.org/en/publications/pisa-2025-results-volume-i_73451bc5-en.html
2. Education International. (Sept 2026). “New PISA report reveals decline in student performance.” https://www.ei-ie.org/en/item/32952:new-pisa-report-reveals-decline-in-student-performance-amid-chronic-teacher-shortage-and-shrinking-attention-spans
3. Fortune. (Sept 11, 2026). “Students are using AI in school and test scores are down, but the OECD finds that those who use it once or twice a week perform similarly to nonusers.” https://fortune.com/2026/09/11/students-ai-school-test-scores-moderate-use-vs-daily-users/
4. Chalkbeat. (Sept 8, 2026). “PISA results 2025 show US gaining as other countries decline.” https://www.chalkbeat.org/2026/09/08/pisa-results-2025-show-us-gaining-as-other-countries-decline/
5. Common Sense Media. (Aug 18, 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
6. 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.” PNAS, 122(26). https://doi.org/10.1073/pnas.2422633122
7. Lally, P., van Jaarsveld, C. H. M., Potts, H. W. W., & Wardle, J. (2010). “How are habits formed: Modelling habit formation in the real world.” European Journal of Social Psychology, 40(6). https://onlinelibrary.wiley.com/doi/10.1002/ejsp.674
8. NYC Mayor’s Office. (Sept 2, 2026). “Mayor Mamdani and Chancellor Samuels announce generative AI moratorium in schools.” https://www.nyc.gov/mayors-office/news/2026/09/mayor-mamdani-and-chancellor-samuels-put-students-first-with-nat
9. NCES. (2026). “Program for International Student Assessment (PISA).” https://nces.ed.gov/surveys/pisa/

Related reading: PISA 2025 results explained for parents, the moderate-AI-use finding in detail, and when AI homework tools help versus hurt.