OECD AI Education Report Finds Students Who Use AI Usually Score Lower

The OECD AI education report has exposed a difficult conflict: most students using chatbots for schoolwork scored below students who avoided them. The pattern appeared across a 2025 assessment involving 760,000 15-year-olds in 91 countries and economies.

That headline does not mean every student should close ChatGPT or Gemini. Moderate users sometimes performed better than occasional or intensive users. Frequent users also gained a slight advantage when schools taught them to evaluate AI-generated information.

The real divide is therefore not AI users against non-users. It is cognitive outsourcing against guided practice. A chatbot can produce a polished answer without ensuring that its user understands the question, evidence, or reasoning behind it.

That distinction pressures schools, families, and AI developers to reconsider what educational adoption should measure. Faster homework and better-looking submissions do not automatically represent learning. Independent performance remains the harder and more important test.

What the OECD AI Education Report Actually Found

Students who avoided AI for schoolwork generally outperformed users, but the result describes an association rather than proven harm.

The OECD published its PISA 2025 results on September 8, 2026. PISA, the Programme for International Student Assessment, measures how 15-year-olds apply knowledge to unfamiliar problems.

The 2025 assessment was the program’s largest edition. It covered 760,000 students across 91 participating countries and economies. Science was the main subject, while reading and mathematics provided additional measures of foundational knowledge.

For the first time, PISA collected detailed information about how often students used AI chatbots for schoolwork. Its questions covered four common activities.

Students reported whether they used chatbots to help them learn, research a new topic, summarize assigned reading, or draft written assignments. The survey also distinguished occasional, moderate, weekly, and nearly daily use.

AI had already become normal for most students. Across OECD countries, only 14 percent reported never, or almost never, using chatbots for every schoolwork activity examined.

Yet non-users generally earned higher science scores than users. Students who did not use AI to draft assignments also outperformed students who reported using it for that purpose.

The pattern varied by activity and frequency. Students using AI to help them learn weekly achieved similar adjusted science scores to non-users. Those using it almost daily, or only occasionally, tended to score lower.

For summarization and preliminary research, moderate users performed better than both limited and intensive users. The OECD defined moderate use for these tasks as between once monthly and twice weekly.

Drafting showed a weaker version of that middle-use advantage. Students who avoided chatbot drafting still held the clearest general lead.

These comparisons adjusted predicted science scores for students’ socioeconomic status. That adjustment matters because AI adoption was not evenly distributed.

Frequent users tended to come from more advantaged backgrounds. Non-users were more likely to face socioeconomic disadvantages, including limited access to devices, reliable internet connections, or paid AI services.

Even after considering socioeconomic background, user behavior and learning dispositions remained difficult to separate. A struggling student might turn to AI because the work is already difficult. Heavy AI use might also reduce the effort needed to develop independent skills.

The report cannot determine which direction dominates. It offers cross-sectional evidence, meaning the survey captured usage and performance during the same assessment period.

The OECD therefore does not claim that opening a chatbot directly lowers a student’s score. Its stronger conclusion is more practical: access alone does not predict better learning.

That conclusion challenges a common assumption behind classroom AI adoption. Availability is easy to count, while understanding is much harder to measure.

Better Homework Can Hide Weaker Learning

Generative AI can raise the quality of completed work while weakening the practice that prepares students to work independently.

The OECD’s findings fit a wider body of experimental research. General-purpose chatbots often improve the immediate output produced during an assisted task. Those gains can disappear when assistance is removed.

The organization’s digital education outlook describes this problem as a gap between performance and learning. Performance measures whether a student completed today’s task successfully. Learning measures whether the student can apply that knowledge later.

A student can ask a chatbot to summarize a chapter and receive an accurate, readable response. That output says little about whether the student recognized the main argument, weighed the evidence, or remembered the material.

Drafting creates an even sharper conflict. A chatbot can turn a short prompt into a structured essay within seconds. The resulting submission might display vocabulary and organization that the student cannot reproduce independently.

The visible product improves, but the hidden learning process shrinks. The student spends less time retrieving information, resolving uncertainty, choosing evidence, and revising weak reasoning.

Those steps can feel inefficient. They are also where much of learning happens.

This mechanism helps explain the unusual frequency curve in the PISA data. Moderate use can complement an active learning process. Intensive use creates more opportunities to replace that process.

Occasional users can also score lower for several reasons. Some may open a chatbot only when they are confused or behind. Their usage frequency then reflects an existing difficulty rather than causing it.

The OECD Digital Education Outlook examined a controlled mathematics study involving high school students in Türkiye. Students received either a general GPT-4 interface or a specially designed tutoring version during practice.

The tutoring system improved practice performance by 127 percent relative to students without AI access. The general chatbot improved it by 48 percent.

However, students who practiced with the general chatbot scored 17 percent lower during a closed-book examination than students who never received AI access. The tutoring group avoided most of that penalty.

The underlying controlled math study offers an important comparison. Both systems used the same underlying model family, but their behavior encouraged different student actions.

The general interface supplied answers readily. The tutor included safeguards intended to guide students through problems without immediately completing the work.

That difference moves the debate beyond model accuracy. Even a correct answer can undermine learning when it arrives before a student forms and tests an independent approach.

Educational AI therefore needs another quality measure. Developers usually evaluate whether a model answers correctly, follows instructions, and satisfies users. A learning system must also preserve productive effort.

Productive effort does not require withholding every answer. It means prompting recall, requesting an explanation, identifying a misconception, and delaying assistance until the learner attempts the problem.

A conventional chatbot optimizes for immediate helpfulness. An educational tutor must sometimes resist that impulse.

This creates an uncomfortable product conflict. Students often prefer tools that reduce work, while schools need tools that improve independent capability.

Usage, completion speed, and satisfaction can reward the wrong design. A system that makes assignments effortless might look successful on a dashboard while weakening examination performance.

The PISA AI student scores make that measurement problem visible at global scale. Schools cannot judge educational technology solely through adoption or assignment quality.

They need evidence that skills transfer into situations where the chatbot is absent, restricted, or unable to supply a ready-made answer.

AI Literacy Changes the Result

Students benefited most when schools combined frequent AI use with explicit practice in questioning the information it produced.

Around six in ten students across OECD countries reported receiving classroom assignments that involved assessing AI-generated information. National shares ranged from 31 percent to more than 80 percent.

Among those students, 46 percent said they sometimes completed such activities. About 33 percent did so often, while 22 percent reported doing so very often.

Students received more practice evaluating ordinary online information. Roughly eight in ten said lessons had asked them to assess the quality of information found on the internet.

That distinction matters because AI output creates specific challenges. A chatbot produces fluent language even when its reasoning or evidence is incomplete.

Evaluating it requires ordinary reading skills and some understanding of how generative systems construct responses. Students must inspect claims, trace sources, recognize uncertainty, and compare answers with reliable material.

PISA found a modest but notable reversal among students receiving those opportunities. Frequent users who learned to assess AI information slightly outperformed non-users and students using AI twice monthly or less.

The result does not prove that evaluation lessons caused the improvement. Schools offering stronger AI instruction might also have better teachers, resources, or broader academic support.

However, it weakens the idea that AI use is uniformly harmful. The educational setting and the student’s behavior appear to shape the outcome.

The OECD’s AI literacy findings emphasize interpretation rather than passive consumption. That focus places reading ability at the center of effective chatbot use.

A useful classroom exercise might give students an AI-generated explanation containing accurate statements, unsupported assumptions, and one false citation. Students would then annotate and verify each claim.

Another assignment could require learners to write a first response without AI. They could ask a chatbot for criticism afterward, then explain which suggestions they accepted or rejected.

These activities retain student ownership. The model becomes an object of analysis or a source of feedback, rather than an automatic answer machine.

They also create evidence that teachers can inspect. A final essay alone may conceal how much reasoning came from the student. Drafts, prompts, source notes, and written judgments expose the process.

This is where personal knowledge practices can support responsible use. Students who preserve sources and their own notes can compare AI responses against an evidence trail.

A structured student workspace can help keep readings, notes, and questions connected. The learning value still depends on how the student uses that material.

Critical assessment cannot become a one-time AI safety lesson. It must appear inside ordinary science, history, writing, and research tasks.

Students also need permission to challenge a fluent answer. Chatbots communicate with confidence, and younger users can mistake polished language for authority.

Teachers can counter that effect by rewarding verification and correction. A student who identifies a model’s unsupported claim should receive credit for reasoning, not just for reaching the expected answer.

The goal is not distrust for its own sake. It is calibrated trust based on evidence, task type, and uncertainty.

That approach better matches the real AI impact on learning. The same model can either shorten thinking or provide feedback that extends it.

The Score Gap Does Not Prove AI Caused the Decline

The OECD data identify a warning signal, but they cannot isolate chatbots from the many forces affecting student performance.

Headlines about AI users scoring worse can easily outrun the evidence. PISA did not randomly assign hundreds of thousands of students to use or avoid chatbots.

Students described their own behavior. Self-reported frequency can be imprecise, and respondents may interpret “help me learn” differently.

One student might use AI to generate completed answers. Another might request practice questions, feedback, or an alternative explanation. Both can enter the same broad usage category.

Science scores also represent accumulated knowledge, not only activities completed during 2025. Students arrived at the assessment with different schools, teachers, family support, motivation, and prior achievement.

Reverse causality remains plausible. Lower-performing students might rely on chatbots more often because they need additional help. That pattern would produce lower scores among users without proving that AI created the gap.

The OECD openly acknowledges this limitation. It says the relationships do not necessarily indicate that AI negatively affects science performance.

Broader achievement trends also began before widespread generative AI adoption. Across OECD countries, reading scores fell 28 points between 2015 and 2025. Mathematics scores fell 22 points.

The OECD equates those changes with approximately one and a half years of learning in reading and slightly more than one year in mathematics.

ChatGPT became publicly available only in late 2022. It cannot explain a decade-long decline by itself.

Pandemic disruption, attendance, family engagement, digital distraction, curriculum changes, and learning conditions all belong in the analysis. The PISA results do not assign a precise share of the decline to each factor.

Science performance remained stable between 2022 and 2025 across the OECD average. That stability further complicates claims that chatbot adoption produced an immediate overall collapse.

At the same time, the absence of causal proof should not erase the warning. The observed pattern repeatedly favors non-use, moderate use, or guided use over intensive, unguided assistance.

Controlled studies also provide mechanisms that make a learning penalty credible. When students receive answers before attempting retrieval and reasoning, they practice fewer of the skills later measured without assistance.

The strongest position therefore sits between panic and complacency. PISA does not prove that AI makes students less intelligent. It shows that ordinary adoption has not delivered a general academic advantage.

That finding should change the burden of proof. Schools should not assume a chatbot is educational merely because students use it for schoolwork.

AI developers also need evidence beyond engagement. If a company markets a learning product, it should demonstrate durable gains through delayed and independent assessment.

Such testing should compare general chatbots, constrained tutoring tools, teacher-guided use, and no-AI practice. It should also track whether benefits persist across subjects and student groups.

Equity adds another uncertainty. Disadvantaged students were less likely to receive classroom tasks involving assessment of AI-generated information.

They were also more likely to be non-users, potentially because of limited access. Expanding access without expanding guidance might replace one divide with another.

The central risk is not simply that some students have better models. It is that some students learn how to interrogate models while others receive only instant answers.

Schools Must Redesign Tasks, Not Just Police Chatbots

The findings shift pressure from blanket access decisions toward the design of assignments, assessments, and AI products.

Schools initially treated generative AI mainly as an academic integrity problem. That concern remains substantial.

In a separate OECD teacher survey, 72 percent of lower secondary teachers said AI can harm integrity by allowing students to present generated work as their own.

Detection alone offers a weak response. AI detectors estimate whether text resembles machine-assisted writing, but they do not reliably reconstruct how a student produced an assignment.

A student might use AI for brainstorming, grammar feedback, unsupported drafting, or full substitution. The final text rarely reveals those differences cleanly.

Assessment design can capture more useful evidence. Teachers can request outlines, source annotations, interim drafts, oral explanations, or short supervised follow-ups.

These methods make reasoning visible. They also remain useful when a school permits some forms of AI assistance.

Closed-book exams provide one measurement of independent capability. They should not become the only answer, because learning also includes research, collaboration, revision, and tool use.

A balanced system can assess both modes. Students should demonstrate foundational knowledge without AI and responsible application with AI.

The OECD’s global release argues for purposeful, moderate technology use. It also warns against replacing attention, effort, and understanding.

That principle gives schools a workable standard. An assignment should specify which cognitive work belongs to the student and where AI can contribute.

For example, a teacher might prohibit AI during initial problem solving but allow it during revision. Students could then compare their approach with the model’s response.

A history assignment could permit topic exploration while requiring every factual claim to connect to an approved source. Students would submit a short explanation of any rejected AI suggestions.

A writing course could use chatbots as critics rather than authors. The model might flag unclear passages, but the student would decide how to revise them.

These designs make AI use slower than copying an answer. That friction is valuable when it forces learners to inspect their own thinking.

Developers face a parallel design challenge. Educational modes need more than friendly tone, simplified explanations, or age-specific interface controls.

A credible tutor should diagnose misconceptions, ask students to commit to an answer, and adjust hints gradually. It should avoid completing the central intellectual task too early.

Teachers also need visibility into the interaction. A useful classroom system could summarize the concepts attempted, hints requested, and errors corrected without exposing unnecessary personal data.

Privacy and safety remain important, especially for minors. Schools need clear rules covering data retention, model training, age appropriateness, and access to student conversations.

Teacher preparation is another pressure point. The OECD reports that 37 percent of lower secondary teachers used AI for their work in 2024.

Teachers cannot guide student behavior through policy documents alone. They need subject-specific examples showing when AI supports learning and when it removes essential practice.

The right approach will vary by discipline. Calculator-like assistance might be appropriate after a mathematics concept is mastered but harmful during its introduction.

Language learners can benefit from conversational practice. They still need opportunities to retrieve vocabulary and construct sentences without automatic completion.

The question is therefore not whether every classroom should adopt or ban AI. It is whether each use preserves the mental action the lesson intends to develop.

Three Signals Will Show Whether Educational AI Is Improving

The next phase should be judged through independent learning, guided-use access, and product behavior rather than raw chatbot adoption.

The first signal is performance without assistance. Schools and researchers need delayed assessments that test whether students retain and transfer skills after using AI.

Immediate assignment quality is insufficient. A useful system should improve results when the tool is unavailable and when the question differs from the practice example.

If future randomized studies show sustained gains across subjects, the OECD warning will weaken. If polished homework continues to precede weaker examinations, it will become stronger.

The second signal is access to explicit AI evaluation instruction. PISA found that about six in ten students received some classroom practice assessing AI-generated information.

That share should rise, but participation alone is not enough. Researchers should measure the quality, frequency, and subject integration of those lessons.

They should also watch the socioeconomic gap. Disadvantaged students currently report fewer opportunities to assess AI output during school lessons.

If that gap closes while guided users improve, AI literacy will look like a meaningful intervention. If access expands without independent gains, schools will need deeper instructional changes.

The third signal is whether education products resist answer substitution. Developers should explain how tutoring modes preserve student reasoning and how they measure learning after assistance ends.

Useful evidence would include independent evaluations, transparent study designs, and comparisons against ordinary chatbots. Engagement figures cannot answer the educational question.

Product behavior also matters during everyday use. Does the tutor ask for an attempt before supplying a solution? Does it request evidence and reveal uncertainty?

Can teachers configure the type of help available for each task? Can students review how their reasoning changed instead of receiving only a finished response?

Those choices will determine whether AI becomes a shortcut, a tutor, or an uneven mixture of both.

The OECD AI education report does not justify pretending chatbots will disappear from school. Most surveyed students had already used them, and enforcement cannot reverse that reality.

It does justify a higher standard. Students should leave an AI-assisted task knowing more than they knew before it, not merely holding better text.

Parents and learners can apply the same test immediately. After using a chatbot, close it and explain the idea, solve a related problem, or defend the answer with sources.

If the understanding disappears with the chat window, the tool completed the work but did not support learning. The next question is whether schools and developers will design for that distinction.