The Sweet Spot for Kids Using AI

When the OECD published its AI-and-schoolwork numbers from the September 2026 PISA study, headlines ran with the flat version: students who use AI score lower in science. That headline is true for one specific group. It’s also missing the part of the finding that matters most if you have a kid doing homework tonight.

What did the new PISA study find?

PISA (the Programme for International Student Assessment) tested roughly 760,000 15-year-olds across 91 countries and economies in its 2025 round, the largest PISA edition yet, according to the OECD’s own published results. Students who reported using AI to draft written assignments every day scored an average of 481 in science. Students who said they rarely or never used AI that way scored 509, a 28-point gap.

The real shape of the AI-use curve

Here’s the part most coverage skipped: that 28-point gap is between the two extremes, not a straight downward line. According to the OECD’s own explanation of the findings, students who used AI moderately, roughly once or twice a week, outperformed both the daily users and the near-never users. The relationship is a curve, not a slope: too little AI use and too much AI use both correlate with lower scores, and the middle is where students did best.

AI use frequency	Average science score
Rarely or never	509
Moderate (about 1-2 times a week)	Highest of the three groups
Daily, for drafting assignments	481

The OECD’s blog didn’t publish an exact score for the moderate-use group, only that it topped the curve.

Why heavy AI use hurts science scores

PISA also tested critical-thinking-linked reading skills specifically: evaluating evidence, telling fact from opinion, and handling ambiguous information. Those are exactly the skills that decline most among the heaviest AI users. A student who lets AI produce a finished answer never practices judging whether that answer is any good, and that muscle atrophies like any other one that goes unused.

Does this mean AI is bad for learning?

No, and the OECD’s own director for education, Andreas Schleicher, has been explicit about this on the organization’s blog: the outcome depends heavily on whether a student is taught to question and check AI output, or simply taught to accept it. 86% of PISA students reported using an AI chatbot for at least one schoolwork purpose, so “should kids use AI for homework” is no longer a live question. “How are they being taught to use it” is the one that matters.

What makes the difference between AI help and AI harm

Classrooms that explicitly taught students to interpret and check AI-generated answers, rather than just consume them, saw better outcomes among their AI-using students than classrooms that didn’t, per the OECD’s framing of the results. Good teachers have always asked the same three questions about any source: does this make sense, what’s it missing, how would you check it. Pointing those questions at an AI’s answer instead of a textbook’s is the entire skill gap this data is describing.

How many students are already using AI for schoolwork?

Almost all of them, in some form. The OECD figure, 86% of students having used an AI chatbot for at least one schoolwork purpose (help learning a concept, research, summarizing, or drafting), makes clear that the debate over whether to allow AI in schoolwork is largely already settled by students themselves. The open question is how well each classroom is teaching the checking-and-questioning skill that separates the moderate-use group’s results from the daily-use group’s.

What should parents do about this?

Two practical takeaways come out of this data. First, daily, unsupervised AI use for every assignment is the pattern most associated with the biggest score gap, so it’s worth noticing if that’s becoming a habit rather than a tool for specific tasks. Second, banning AI outright isn’t what the data supports. Asking the kid to explain, out loud, whether they think the AI’s answer is right and how they’d check it does more: that’s the habit behind the gap between the moderate-use and daily-use groups.

Three questions worth asking a kid before an AI-assisted assignment goes in the folder:

• Can you explain this answer in your own words, without looking at the screen?
• What part of this did the AI just guess at, and how would you check it?
• Would you have written this differently, and why or why not?

Where AI helps students learn

We’ve written before about the specific reading-skills concern this study echoes in The Reading Crisis Before AI, and about age-appropriate AI use in AI and Your Kid’s Reading and ChatGPT for Teens: Parent’s Guide. If you’re choosing a tool rather than deciding on the ground rules, Best AI for Students in 2026 covers what’s built specifically for schoolwork.

Frequently asked questions

Does using AI for homework lower test scores?
It depends on how much and how. Daily, unsupervised use correlates with lower science scores in the PISA data. Moderate use, about once or twice a week, correlated with the highest scores of any group, including students who never use AI at all.

Is the OECD saying schools should ban AI?
No. The OECD’s own explanation of the findings frames the issue as how AI is taught and used, not whether it’s used. 86% of students already use AI chatbots for some schoolwork purpose.

What specifically got worse for heavy AI users?
Reading skills tied to critical thinking: evaluating evidence, telling fact from opinion, and navigating ambiguous information. Those are the skills a student skips practicing when AI hands them a finished answer.

How many students were in this study?
Roughly 760,000 15-year-olds across 91 countries and economies, making PISA 2025 the largest edition of the assessment to date.

What can a parent do with this information?
Ask your kid to explain and check the AI’s answer rather than just turning it in, and use the three questions above as a starting point.

Sources

OECD: PISA 2025 Results (Volume I) (September 2026)

OECD: PISA findings on artificial intelligence use, reading skills and learning (September 2026)

Fortune: Students, AI, and school test scores (moderate use vs. daily users)

MLQ: OECD’s PISA data link student AI use to lower science scores, with important exceptions

You might also like

The Reading Crisis Before AI
AI and Your Kid’s Reading
ChatGPT for Teens: Parent’s Guide
Best AI for Students in 2026
AI for High School Students: Study Smarter, Not Harder
AI for Math Students
Do AI Detectors Work? What to Know