Updated September 9, 2026 3:54 pm
In short
OECD data still shows students who use AI generally score lower, but weekly use for general learning can slightly help—and the benefit is stronger when students are trained to evaluate AI output.
- OECD PISA data links student AI use with generally weaker science performance.
- Moderate, intentional AI use performs better than rare or daily dependence.
- Teaching students to judge AI output can flip the outcome from negative to mildly positive.
- AI use is more common among advantaged students, raising equity concerns.
Update — September 9, 2026 3:54 pm
The OECD report adds one important nuance: students who use AI weekly to broadly help them learn can edge ahead of nonusers, but only by a small margin.
That advantage becomes clearer when those students are also regularly taught to judge AI-generated information. In that setting, regular AI users for general learning purposes outperform students who do not use AI at all.
The report’s authors say the pattern points to moderate, intentional AI use as the most promising approach for schoolwork.
Students who use AI to help with schoolwork generally perform worse than peers who do not, according to a new OECD analysis of global test data released on September 9, 2026. But the report also finds a meaningful exception: when students are taught to evaluate AI-generated answers critically, regular use of AI can slightly improve results.
The findings come from the latest Programme for International Student Assessment, or PISA, which examined 15-year-olds in science, mathematics and reading across 91 countries and economies. Based on responses from more than 760,000 students collected in 2025, the report is the first major international snapshot of school-age AI use since generative tools became widely available.
The headline result is not a simple anti-AI verdict. Instead, the OECD paints a more complicated picture in which how students use AI, how often they use it, and whether schools teach them to judge its output all matter to academic performance.
What the OECD found about AI and grades
The OECD’s data suggests that students who never use AI tend to do better in science than students who do, even after adjusting for socioeconomic background. That gap is not uniform, and the report stresses that certain uses of AI are less harmful than others.
Students who turned to AI for narrow school tasks — such as summarizing reading material or drafting written assignments — showed weaker performance than non-users. By contrast, those who used AI more broadly for research or general learning support did not appear to fall as far behind.
The pattern is especially important because it challenges the popular assumption that any use of AI automatically gives students an edge. In the OECD’s view, the issue is not merely access to the technology, but whether the technology is used in a way that supports active learning rather than replacing it.
How often students use AI also matters
The report says the relationship between AI and achievement follows something closer to a U-shape than a straight line. Students who used AI only once or twice a year performed poorly, and so did those who used it every day. Better results were associated with monthly or weekly use.
That middle ground suggests the most beneficial AI use may be occasional and deliberate rather than constant. In practical terms, students who treat AI as a limited support tool may do better than those who lean on it too rarely or too heavily.
| AI use pattern | Typical performance trend | Interpretation from OECD data |
|---|---|---|
| Never uses AI | Generally stronger science results | Students preserve more direct engagement with learning tasks |
| Once or twice a year | Among the weakest results | Too little use may not build useful skills, and could reflect limited access |
| Monthly or weekly | Relatively better outcomes | Moderate use may support learning without replacing effort |
| Daily use | Among the weakest results | Heavy dependence may short-circuit productive learning |
Why critical thinking training changes the picture
Students do better with AI when schools teach them how to question what the tool produces. The OECD says that among students who regularly use AI for general learning, performance improves once they are also asked to assess the quality of AI-generated information.
That finding is the report’s clearest policy signal. It implies that the educational value of AI depends less on the software itself than on the skill set students bring to it.
When learners are trained to spot errors, identify weak reasoning, and compare AI output with other sources, the technology may reinforce learning instead of weakening it. Without those habits, students can become passive consumers of content rather than active participants in the thinking process.
OECD learning and skills director Andreas Schleicher argued that students do not improve simply by absorbing information, but by wrestling with material in a way that develops understanding. He said technology helps only when it supports that struggle, and harms learning when it bypasses it.
In his framing, AI is neither a guaranteed academic shortcut nor an automatic educational breakthrough. It is a tool whose impact depends on whether it strengthens the mental effort that real learning requires.
Who is using AI in school — and who is not?
The OECD also found that AI use is not evenly distributed. Students from more advantaged backgrounds are more likely to use AI, which likely reflects better internet access, stronger devices and greater access to paid tools.
That inequality matters because it means AI may widen existing gaps if schools and policymakers do not intervene. Students with more resources are often the first to encounter new digital tools, and they are also more likely to receive guidance on using them well.
Internationally, adoption rates varied sharply. In Vietnam, more than 95 percent of students reported using AI tools. In Japan, the share was closer to 60 percent.
Those differences highlight how uneven the educational AI landscape remains, even before questions of quality and effectiveness are considered. Access, school policy, local culture and national technology infrastructure all appear to shape whether students use AI at all.
Curiosity is rising — but not always grades
One of the report’s more encouraging findings is that curiosity appears to be linked to AI use, with the highest curiosity levels among daily users. That suggests many students are drawn to the technology because they want to explore, ask questions and learn faster.
But curiosity alone does not guarantee stronger academic outcomes. The OECD says that for now, the boost in interest is not consistently translating into better marks.
That disconnect helps explain why educators are wary. AI may make learning feel easier or more engaging, but if it replaces independent thinking, the long-term academic cost could outweigh the short-term convenience.
How the PISA study was conducted
The OECD’s PISA program is one of the world’s most influential international education assessments. It surveys and tests 15-year-olds every few years, allowing researchers to compare academic outcomes and related behaviors across countries.
This latest cycle is notable because it captures the first wave of student AI use after generative tools became mainstream. The 2025 data includes responses from more than 760,000 students in 91 countries and economies, making it one of the largest international datasets yet on AI in education.
Unlike a classroom experiment, the PISA findings do not prove that AI directly causes lower scores. The OECD is careful to frame the results as correlations rather than simple cause-and-effect. Students who use AI may differ from non-users in ways that are harder to measure fully, even after statistical adjustments.
Still, the dataset is large enough to identify clear trends and to suggest how schools might respond as AI becomes more embedded in everyday study habits.
What the data can and cannot prove
The report shows patterns, not final judgments. It indicates that AI use is associated with weaker science performance in many cases, but it cannot isolate every background factor, school policy or home environment that may influence results.
That limitation is important for parents, teachers and policymakers. The data should not be read as proof that AI harms every student, nor as evidence that it helps no one. Rather, it shows that outcomes vary depending on use case, frequency and training.
In other words, the report measures a complicated education ecosystem, not a laboratory test.
Why the OECD says learning still requires effort
The strongest theme running through the report is that education works best when students actively process material instead of passively receiving it. AI can either support that effort or replace it.
When students use the technology to brainstorm, check ideas or test their understanding, it may reinforce the mental work involved in learning. When they use it to summarize a chapter, write an essay draft or answer a question for them, it may reduce the struggle that actually builds knowledge.
That distinction helps explain why moderate users often do better than heavy or minimal users. A little AI, used for clearly defined tasks, may help students stay engaged. Too much AI may encourage dependency. Too little may leave learners without useful support or digital literacy.
The OECD’s central message is that education systems should teach students not just how to use AI, but how to use it wisely.
What this means for schools and families
The report arrives as schools around the world are still trying to define their rules for generative AI. Some districts have moved quickly to ban or restrict use, while others are experimenting with classroom integration and teacher-led guardrails.
For families, the findings suggest a practical takeaway: AI should not be treated as a shortcut for homework, especially when the goal is long-term learning. Used carelessly, it may weaken the development of reading, writing and problem-solving skills. Used with guidance, it may support them.
For schools, the policy implications are clear:
- Teach students how to check AI output against reliable sources.
- Use AI for limited, clearly defined learning tasks rather than full substitutions.
- Make sure access to AI tools does not depend entirely on family income or device quality.
- Train teachers to identify both the strengths and the weaknesses of AI-assisted work.
The findings also suggest that blanket bans may miss the point. If AI is already widely available outside school, then the more durable solution may be to teach judgment, not just impose restrictions.
How should policymakers read the findings?
Policymakers should read the OECD report as a warning against both hype and panic. The study does not show that AI automatically improves education, and it does not show that students should avoid the technology entirely.
Instead, it points to a middle path: AI can be educationally useful when students know how to interrogate it and use it sparingly. That means education policy may need to move beyond access alone and focus more on digital literacy, source evaluation and critical thinking.
The distribution data also raises a fairness issue. If better-off students are the first to learn how to use AI effectively, then schools risk turning a new technology into an amplifier of existing inequality. That is especially true in systems where devices, broadband and premium AI access are unevenly distributed.
In that sense, the report is about more than test scores. It is about who gets to benefit from the newest educational technology — and under what conditions.
The bottom line
OECD data does not suggest that AI is useless in school, but it does suggest that more AI is not automatically better. Students who use it heavily or casually for shortcut-style tasks tend to do worse, while those who use it moderately and are trained to question its answers can see small gains.
For now, the report’s message is straightforward: AI may help education, but only if schools preserve the hard thinking that learning depends on.
| Key finding | What the OECD observed | Why it matters |
|---|---|---|
| General AI use | Often linked to lower science scores | AI does not automatically improve academic performance |
| Moderate use | Monthly and weekly users often did better than extreme users | Intentional use may be more effective than constant dependence |
| Critical assessment training | Improved outcomes for regular users | Digital literacy can turn AI into a learning aid |
| Access gap | Use was more common among advantaged students | AI could widen inequality without school intervention |
Frequently asked questions
Does AI use in school lower student grades?
Yes, generally it does in the OECD data. Students who use AI for school tasks tend to score worse than non-users after background factors are adjusted, although the effect varies by how and how often the tool is used.
Can AI ever help students learn better?
Yes, it can. The OECD found that students who used AI moderately and were trained to assess its quality could outperform non-users, suggesting that thoughtful use can support learning rather than replace it.
Which students use AI the most?
Students from more advantaged backgrounds are more likely to use AI, according to the OECD report. That likely reflects better access to devices, internet service and paid AI tools, which can widen existing inequality.
What kind of AI use is most risky for learning?
AI used as a shortcut for summarizing readings or drafting assignments appears most risky. The OECD says these uses are associated with weaker performance than using AI for research, practice or general help learning.
