PISA 2025 AI Use Students Data: Moderate Beat Heavy and Zero

The PISA 2025 AI use students data is the most interesting thing in the entire report, and almost nobody has explained the shape of it. Here is the finding: 46% of OECD 15-year-olds use AI chatbots weekly or more to help them learn. Students who never use AI for a specific schoolwork task generally score higher in science than students who do. But among the students who do use it, the curve is not a straight line down. For summarizing reading and doing preliminary research, moderate users, defined as once a month to twice a week, outscored both the rare users and the near-daily users.

That U shape matters more than the headline “AI users score lower,” because most kids are already using it. The question is not whether. It is how often, for what, and with what instruction.

Key Takeaways

• 46% of OECD students use AI chatbots weekly or more “to help me learn.” Only 14% never use AI for any schoolwork purpose.
• Weekly users of AI “to help me learn” score about the same as non-users once socio-economic status is accounted for. Daily users and occasional users score lower.
• For summarizing and preliminary research, moderate users (monthly to twice weekly) outscore both limited and frequent users. For drafting text, the moderate-use advantage mostly disappears.
• Students who get taught to evaluate AI output in class and use AI weekly slightly outperform their peers. About six in ten OECD students report ever being asked to assess AI-generated information.
• Non-users skew socio-economically disadvantaged; frequent users skew advantaged. The OECD warns of “a growing AI divide.”

What the PISA 2025 AI use students questions actually asked

PISA 2025 asked 15-year-olds how often they use AI chatbots such as ChatGPT for four purposes: to summarize a text they had to read, to conduct preliminary research on a new topic, to draft text for writing assignments, and a general “to help me learn.” Response options ran from “never or almost never” to “every day or almost every day.” The results are in Chapter 4 of PISA 2025 Results (Volume I), released September 8, 2026.

Business Standard’s breakdown of the same tables makes the point plainly: usage is already the norm. On average across OECD countries, only 14% of students reported never (or almost never) using AI for any of the four purposes. “To help me learn” is the most common use, followed by preliminary research, then drafting, then summarizing.

The OECD’s summary of the performance relationship is worth quoting in full because it is more careful than most coverage: “Students who do not use AI for specific schoolwork tasks outperform those who do, while weekly users and non-users of AI for general purposes perform similarly. Students tend to perform slightly better when frequent AI use is paired with opportunities to develop AI literacy skills.”

The U-shaped curve, task by task

Here is the pattern the OECD reports for science performance, adjusted for students’ socio-economic status. “Moderate” means between once a month and twice a week. “Frequent” means almost every day or more. “Limited” means twice a year or less.

AI use for…	Non-users vs. users	Where moderate users land	Strength of the U shape
Summarizing assigned reading	Non-users score higher	Above both limited and frequent users	Clear
Preliminary research on a new topic	Non-users score higher	Above both limited and frequent users	Clear
Drafting text for writing assignments	Non-users score higher	Advantage “less pronounced”	Weak
General “to help me learn”	Weekly users ≈ non-users	Weekly users highest among users	Moderate

Two observations that the headlines skip.

First, the U shape is strongest for the tasks where AI is doing something adjacent to thinking, and weakest for drafting, where AI is doing the thinking. Summarizing a text you have already read, or orienting yourself in a new topic, leaves the hard part to the student. Having a model write your paragraph does not.

Second, “to help me learn” behaves differently from the specific tasks. Weekly users of that general category perform about the same as non-users. That is the closest thing in the data to a neutral verdict on AI as a study aid.

Why this is association, not causation, and why that matters here

I want to be direct about this because it is where most coverage goes wrong. PISA is a cross-sectional survey administered on one day. It cannot tell you whether AI use lowered a score or whether lower-performing students reach for AI more. The OECD says so explicitly: “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.”

The report then hands you three reasons the arrow might point backwards:

Access. Students who do not use AI “tend to come from more socio-economically disadvantaged backgrounds,” partly because of device and connectivity limits. Frequent users skew advantaged. Socio-economic status was controlled for in the score comparisons, but imperfectly measured confounders remain.

Help-seeking. Among students using AI “to help me learn,” frequent users report seeking help more than both non-users and light users. That is a disposition, not an effect of the tool.

Motivation. Intrinsic motivation to learn is generally highest among weekly users, regardless of purpose. Kids who are already motivated may be the ones using a chatbot once or twice a week rather than every day.

A parent should read the U-shape as a description of which students are where, not as a dosage instruction. There is no study establishing that moving your child from daily to weekly AI use would raise a science score. What the data do support is the narrower claim that heavy, unguided use is not associated with better outcomes anywhere in the report.

The one finding that looks closest to actionable: teach evaluation

Box I.4.2 of Volume I contains the most useful number in the AI section. About six in ten OECD students report having been asked, in class, to assess the quality of AI-generated information; the range across systems is 31% to over 80%. Among those students, 46% did so “sometimes,” 33% “often,” and 22% “very often.”

And here is the pattern: “When students have learning opportunities related to AI in school, those who reported using AI to help them learn frequently (once a week or more) tend to slightly outperform their peers, both AI non-users and AI users about twice a month or less.”

In other words, frequent use plus instruction on evaluating output looks better than frequent use alone. The OECD adds the caveat that this “cannot determine causality” and the warning that disadvantaged students are less likely to get that instruction, which “risks a growing AI divide.”

This lines up with what US teen survey data show. Common Sense Media’s “Teens in the AI Era” (August 18, 2026, n=1,017 teens aged 13–17) found 70% use AI for schoolwork but only 27% say a teacher ever discussed safe AI use with them. The demand exists; the instruction largely does not.

Also in the data: distraction is the clearer signal

If you want a finding with a stronger evidence base than the AI curve, it is device distraction, highlighted in the OECD’s September 8 release. On average, 28% of OECD students report that classmates are distracted by digital devices in most or every science lesson. In B-S-J-Z (China), Japan, and Korea, that figure is under 10%, and a comparison of the two top systems puts B-S-J-Z at just 5%. And the OECD reports that “in schools with clear policies on mobile phone use, whether through bans or well-defined guidelines, fewer students are distracted by digital devices.”

The broader digital finding is consistent: “limited or moderate use of digital devices for learning at school is often associated with better outcomes,” while leisure use during school “reverses sharply” the relationship. Same U shape, wider evidence base.

What to actually do at home

Separate the four uses, then set different rules

Do not make one AI rule. Make four. Summarizing something your child has already read, and orienting in a new topic, are the uses where moderate frequency looked best in the data. Drafting is the use where the moderate advantage nearly vanished. Treat drafting as the one that needs the most supervision.

Add the evaluation step, because school probably will not

The single association that pointed in a good direction was frequent use combined with instruction in assessing AI output. You can supply that in five minutes: after your child uses a chatbot, ask them to find one thing in the output that is wrong, unsupported, or too vague. If they cannot find anything, they did not read it carefully.

Make the chatbot go second, not first

The strongest reading in the data is that AI helps least when it replaces the effortful part. A workable house rule: your child writes or attempts first, then asks the model. The OECD’s own framing is that AI “should add to students’ effortful learning rather than replace it, since real understanding comes from doing the thinking, not from having a tool do it instead.”

Fix distraction before you fix AI

28% of students see classmates distracted by devices in most science lessons, and schools with clear phone rules report less of it. A phone in another room during homework has better evidence behind it than any AI policy you can write this week.

What not to do

Do not ban AI outright on the strength of this report. Non-users scored higher on average, but non-users also skew disadvantaged and differ in help-seeking and motivation, and the report explicitly declines to claim a causal effect. A ban is a defensible parenting choice; it is not a finding.

For the wider context on how different school systems are handling this, see our look at why the top PISA performers use very different amounts of AI, and our parent’s guide to the full PISA 2025 results.

What to Watch For Over the Next 3 Months

• Week 4: Ask your child which of the four uses they actually do and how often. Most parents guess wrong, usually assuming less use than is happening.
• Month 2 red flags: Homework that reads better than your child talks, or a kid who cannot explain a paragraph they submitted. That is drafting-mode use without the evaluation step.
• Month 3 self-check: Your child can name one error they caught in AI output this month. If they cannot, the evaluation habit has not taken.

Frequently Asked Questions

Did PISA 2025 find that AI makes kids worse at school?

No. It found that students who use AI for specific schoolwork tasks score lower in science, on average, than students who do not, and that weekly “help me learn” users score about the same as non-users. The OECD explicitly says these associations may reflect who adopts AI and how, not a causal effect.

What counts as “moderate” AI use in the PISA data?

Between about once a month and twice a week for a given task. “Frequent” is almost every day or more; “limited” is twice a year or less. Moderate users outscored both groups for summarizing and preliminary research.

How many students use AI for schoolwork?

Across OECD countries, only 14% reported never or almost never using AI for any of the four schoolwork purposes PISA asked about, and 46% reported using it weekly or more to help them learn. Use is already the default, not the exception.

Does teaching kids to evaluate AI output actually help?

The association is positive. Among students who get classroom practice assessing AI-generated information, frequent AI users slightly outperform both non-users and light users. The OECD says the finding cannot establish causality, but it is the one pattern in the section that points clearly upward.

Should I let my 12-year-old use ChatGPT for homework?

The PISA data cannot answer that for an individual child. What it supports: keep frequency moderate rather than daily, keep AI out of the drafting step, and add an explicit evaluation habit. If your child cannot explain the work, the tool did the learning.

---

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. (2026). “PISA 2025 Results (Volume I): Future-Ready Students,” Chapter 4 and Box I.4.2. OECD Publishing. https://www.oecd.org/en/publications/pisa-2025-results-volume-i_73451bc5-en.html
2. OECD. (2026). “PISA 2025: Students’ reading and mathematics performance declined sharply across the OECD.” OECD press release, September 8, 2026. https://www.oecd.org/en/about/news/press-releases/2026/09/pisa-2025-students-reading-and-mathematics-performance-declined-sharply-across-the-oecd.html
3. Common Sense Media. (2026, August 18). “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
4. Hong, J. (2026, September 11). “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.” Fortune. https://fortune.com/2026/09/11/students-ai-school-test-scores-moderate-use-vs-daily-users/
5. Business Standard. (2026, September 14). “PISA 2025: How 15-year-olds use AI chatbots for schoolwork.” https://www.business-standard.com/industry/news/ai-helps-students-with-work-but-more-use-doesn-t-mean-better-scores-126091400473_1.html
6. National Center for Education Statistics. (2026). “Program for International Student Assessment (PISA).” NCES. https://nces.ed.gov/surveys/pisa/
7. Iftikhar, F. (2026, September 9). “China’s PISA toppers use less AI. Singapore uses more. Both score high.” ThePrint. https://theprint.in/india/education/chinas-pisa-toppers-use-less-ai-singapore-uses-more-both-score-high-whats-the-lesson/3038191/