PISA 2025 Singapore China AI Use: Two Opposite Playbooks

If you saw a headline saying the top PISA 2025 performers succeed because they use less AI, the PISA 2025 Singapore China AI use numbers break that story cleanly. B-S-J-Z (China), the four Chinese jurisdictions that topped math and science in the OECD’s September 8 release, reports 41% of students using AI chatbots weekly to help them learn. Singapore, which topped reading and came second in math and science, reports 66%. The OECD average is 46%.

Two systems at the top. One below average on AI use, one well above. Whatever is producing those scores, it is not the chatbot dial.

Key Takeaways

• B-S-J-Z (China): 597 science, 612 math, 527 reading, with 41% weekly AI use “to help me learn.” Singapore: 560 science, 563 math, 535 reading, with 66%.
• On specific tasks the gap is wider: 17% of B-S-J-Z students use AI for preliminary research versus 46% in Singapore (OECD 31%).
• Singapore teaches evaluation: 81% of its students report being asked in class to assess AI-generated information, against an OECD average of 63%.
• B-S-J-Z reports only 5% of students seeing classmates distracted by devices in most science lessons, against 28% across the OECD, and 83% attend schools restricting phones (OECD 50%).
• Neither approach is proven to cause the scores. Both systems were already at the top before chatbots existed.

PISA 2025 Singapore China AI use, side by side

PISA 2025, released September 8, 2026, asked students how often they use AI chatbots for four purposes: summarizing assigned reading, preliminary research on a new topic, drafting text, and a general “to help me learn.” Here is how the two leading systems compare, with figures compiled by ThePrint from the OECD tables.

Measure	B-S-J-Z (China)	Singapore	OECD average
Science score	597	560	482
Mathematics score	612	563	463
Reading score	527	535	461
Computational problem solving	560	563	500
Weekly AI use “to help me learn”	41%	66%	46%
AI for preliminary research	17%	46%	31%
AI for summarizing texts	19%	41%	30%
AI for drafting writing	21%	45%	29%
Taught to assess AI-generated info	81%	81%	63%
Classmates distracted by devices most lessons	5%	n/a	28%
In schools restricting phones	83%	n/a	50%
Daily device use for learning	1.2 hrs	n/a	1.7 hrs
Top performers in math (Level 5–6)	55.5% (any subject)	37% in math	8% in math

Read the AI rows and the score rows together and the causal story collapses. If low AI use produced high scores, Singapore should not be at the top. If high AI use produced high scores, B-S-J-Z should not be.

What they do share

Two things show up in both systems, and neither is about chatbots.

Evaluation instruction. In both systems, 81% of students report being asked in class to assess the quality of AI-generated information, against an OECD average of 63%. Singapore’s OECD country note records that alongside 37% of its students reaching Level 5 or 6 in math. This is the one AI-related variable in the whole PISA report that points consistently upward: the OECD found that students who use AI frequently and get classroom practice evaluating its output slightly outperform both non-users and light users.

Attention protection. B-S-J-Z reports 5% of students seeing classmates distracted by devices in most or every science lesson, versus 28% across the OECD. 83% attend schools that restrict phone use, against an OECD average of 50%. Daily device use for learning is 1.2 hours versus 1.7. Japan and Korea, both in the top 10 across all three domains, also report under 10% device distraction.

So the pattern that survives across both playbooks is not “how much AI.” It is “how much focused attention, and whether anyone teaches you to check the machine.” That is an actionable pair, and it is what our piece on the U-shaped PISA 2025 AI-use curve works through in detail.

Why “less AI equals better scores” is a bad inference

Let me be direct about the analytical error, because it will show up in a lot of coverage this year.

These systems were already on top. B-S-J-Z led PISA math in 2018 and 2022, before ChatGPT existed. Singapore has led or nearly led since it began participating. A variable that only came into existence in late 2022 cannot explain a ranking that predates it.

Country-level correlations are the weakest form of evidence. With roughly 80 systems, comparing an average AI-use rate to an average score confounds everything: curriculum, teacher training, tutoring culture, household income, testing conditions, and how students interpret the survey question. Vietnam reports 69% weekly AI use and Singapore 66%, but their science scores differ by more than 100 points (457 versus 560).

As Business Standard noted in its summary of the same tables, more use does not mean better scores in either direction.

Self-reported use is a soft measure. “Weekly” means different things to a 15-year-old in Shanghai and one in Singapore, and cultural norms about admitting tool use in schoolwork vary.

Access is confounded with advantage. The OECD warns that frequent AI users skew socio-economically advantaged and non-users skew disadvantaged, which shows up in both directions across countries at different income levels.

What can be said is narrower and still useful: at the system level, there is no evidence that restricting AI is necessary for top performance, and no evidence that adopting AI produces it.

What the top systems’ math results actually reflect

Since the AI story does not explain the scores, what does? PISA cannot answer causally, but it can rule some things out and point at others.

Spending is not it. The OECD notes that “national wealth accounts for just half of the variation in countries’ performance” and offers Luxembourg, the world’s highest spender at USD 288,521 per student, scoring below Ireland, France, and Italy in science.

Equity plus performance is achievable together. The report names B-S-J-Z (China), Canada, Estonia, Ireland, Japan, Korea, and Macao (China) as systems demonstrating “both socio-economic fairness and strong performance.”

Depth of curriculum gets a specific endorsement. The OECD’s preface argues for “a smaller set of core topics well enough for genuine mastery” rather than rushing a long list.

And the top-performer share is stunning where it exists. 55.5% of B-S-J-Z students are top performers in at least one subject, against an OECD average of 11.9%. In math alone, 37% of Singapore’s students reach Level 5 or 6, versus an OECD average of 8%. For how that compares with US patterns, see our look at Singapore versus US math and what the actual difference is.

What to actually do at home

Copy the evaluation habit, not the usage rate

The variable both top systems share is teaching kids to check AI output. 81% of their students get that practice in class versus 63% across the OECD. You can supply it in two minutes after any chatbot use: “Find one thing in there that is wrong or unsupported.”

Copy the attention protection

B-S-J-Z’s 5% device-distraction figure is the most striking number in this comparison. You cannot change your child’s classroom, but a phone-free homework hour is the household version of the same policy.

Stop trying to pick the “right” AI dose from country data

There is no country-level dose to copy. Singapore at 66% and B-S-J-Z at 41% both work. What differs between them is everything else about their systems, which you cannot import.

Ask about math depth, not math speed

Both top systems score far above average on top-performer share. The OECD’s own read is that depth beats coverage. When you look at your child’s math work, ask whether they have mastered fewer things well or seen many things once.

What not to do

Do not use “the top countries use less AI” as a reason to ban tools at home, and do not use “Singapore uses more AI” as a reason to encourage them. Both are the same mistake in opposite directions: treating a country-level correlation as a parenting instruction.

What to Watch For Over the Next 3 Months

• Week 4: Whether anyone at your child’s school has taught them to evaluate AI output. Ask directly; 63% of OECD students report it, and the US questionnaire data are thin because over 40% of US respondents did not receive the student questionnaire, per NCES and OECD documentation.
• Month 2 red flags: Coverage or school communications that treat AI-use rates as a causal explanation for scores. That inference does not survive the Singapore-China comparison.
• Month 3 self-check: Does your child’s homework hour look more like B-S-J-Z’s 5% distraction figure or the OECD’s 28%?

Frequently Asked Questions

Do the top PISA 2025 countries use less AI?

One does and one does not. B-S-J-Z (China) reports 41% weekly AI use “to help me learn,” below the OECD average of 46%. Singapore reports 66%, well above it. Both finished in the top group across all three core domains, so AI-use rate cannot be what separates them from other systems.

Which system actually scored highest in PISA 2025?

B-S-J-Z (China) had the highest mean scores in science (597) and mathematics (612). Singapore was highest in reading (535). Macao (China) led computational problem solving at 572, with Singapore at 563 and B-S-J-Z at 560.

What do both top systems do that the OECD average does not?

Two things: 81% of their students report classroom practice assessing AI-generated information, against an OECD average of 63%; and B-S-J-Z reports very low device distraction (5% versus 28% OECD) with 83% of students in schools that restrict phones.

Does banning phones raise test scores?

PISA cannot establish that. It reports that schools with clear phone policies, whether bans or well-defined guidelines, have fewer distracted students, and that distraction is associated with lower performance. That is a correlational chain, not a causal estimate.

Should I copy Singapore’s approach or China’s?

Neither wholesale. What is copyable at home is the overlap: teach your child to evaluate AI output, and protect focused attention during study time. The usage rate itself is not the lever, and neither system’s broader context transfers to a single household.

---

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,” Table I.1 and Chapter 4. OECD Publishing. https://www.oecd.org/en/publications/pisa-2025-results-volume-i_73451bc5-en.html
2. Iftikhar, F. (2026, September 9). “China’s PISA toppers use less AI. Singapore uses more. Both score high. What’s the lesson?” ThePrint. https://theprint.in/india/education/chinas-pisa-toppers-use-less-ai-singapore-uses-more-both-score-high-whats-the-lesson/3038191/
3. OECD. (2026). “PISA 2025 Results (Volume I): Country Note, Singapore.” OECD Publishing. https://www.oecd.org/en/publications/pisa-2025-results-volume-i-country-notes_2d4ff9ea-en/singapore_80497e2f-en.html
4. 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
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. Hong, J. (2026, September 11). “Students are using AI in school and test scores are down.” Fortune. https://fortune.com/2026/09/11/students-ai-school-test-scores-moderate-use-vs-daily-users/