The OECD’s Verdict on AI in Schools: It Works—But Only When Teachers Lead It

The most authoritative review of generative AI in education published to date has a clear, evidence-based conclusion: AI can transform learning. Without pedagogical guidance, however, it produces a mirage: better outputs, no deeper understanding. Here is what the research actually says, and what schools and governments should do about it.

For years, every educational technology wave followed the same script: launch with extraordinary promise, deploy with insufficient preparation, disappoint at scale, and leave teachers to pick up the pieces. Interactive whiteboards. MOOCs. Tablet programmes. The pattern was so consistent that the education research community developed a shorthand for it — the “EdTech hype cycle” — and most experienced educators learned to wait it out. Generative AI, the OECD says in its most comprehensive assessment of the field to date, is different. Not because it automatically works. But it can work under specific, identifiable conditions that governments and schools can actually create.

The OECD Digital Education Outlook 2026, published in January and already the most cited education policy report of the year, is the result of a systematic synthesis of emerging international research, policy analysis, and design experiments across dozens of countries. Its core message is neither the uncritical enthusiasm of the EdTech marketing sector nor the reflexive scepticism of commentators who distrust anything algorithmic in a classroom. It is more useful than either: a careful, evidence-based verdict that GenAI has genuine educational potential, that this potential is being realised in specific contexts, and that how AI is designed and used matters more than whether it is used at all.

The most striking and consequential finding in the Outlook is also the simplest to state: completing a task with GenAI does not mean a student has learned anything. The OECD synthesizes multiple studies showing that students who use general-purpose GenAI tools — standard chatbots, not purpose-built educational systems — consistently produce higher-quality outputs than those working without AI. Essays are more polished. Code runs. Problems are “solved.” But when the AI is removed — during an examination, a practical exercise, an assessment conducted without technological support — the advantages disappear. In some cases, they reverse.

The research reviewed by the OECD is clear about why this happens. Cognitive load theory and decades of learning science establish that genuine understanding requires students to engage in the effortful mental process of connecting new information to existing knowledge — retrieving, elaborating, and applying it. When AI performs those steps on behalf of a student, the neural encoding that constitutes learning either weakens or does not occur. The output is impressive. The learning is absent. This distinction is invisible in a standard assessment of final products, which is one reason it has been missed in many early AI-in-education evaluations.

“When designed or used without pedagogical support, outsourcing tasks to GenAI will only enhance student performance without leading to real learning gains.”

OECD Digital Education Outlook 2026 — core finding, page summary

The Outlook is not a critique of AI in education. It is a guide to what forms of AI use produce genuine learning gains — and the evidence base here is genuinely encouraging. The OECD identifies three productive roles for GenAI in educational settings, each grounded in different evidence streams and each requiring different levels of teacher design and oversight.

Across all three roles, the OECD’s synthesis identifies a common design principle: hybrid systems that combine GenAI with explicit pedagogical models show more promise than general-purpose chatbots. An AI tutor that applies structured tutoring strategies and evidence-centred assessment design outperforms a standard chatbot given the same task. A teacher using AI to design formative feedback based on a clear understanding of student misconceptions outperforms a teacher deploying AI without that diagnostic framing. The human expertise is not replaced — it is amplified.

A consistent theme throughout the Outlook — and one that the OECD has built a dedicated conceptual framework around — is the centrality of teacher agency. This is not a diplomatic gesture toward a profession that might feel threatened by AI. It is an evidence-based finding about what determines whether AI produces learning or produces the mirage of learning.

The report’s framework for “teacher-AI teaming” distinguishes between two types of AI deployment in education. In the first, AI is handed to students or classrooms as a general-purpose tool with minimal pedagogical framing — the approach that produces the crutch effect. In the second, a teacher designs the context in which AI is used: specifying what role the AI plays, what cognitive work the student must still do independently, how AI outputs will be interrogated rather than accepted, and how assessment will be structured to value process over product.

The difference is not in the technology. The same AI model can produce the crutch effect in one classroom and genuine learning gains in another, depending entirely on how the teacher has structured the encounter. This means that teacher professional development — specifically, development in AI pedagogy, not just AI tool familiarity — is the highest-leverage investment any school system can make right now.

One of the most practically significant implications of the OECD’s findings concerns assessment. The traditional model — grade the final essay, the completed problem set, the finished project — is designed to evaluate products. When AI can generate high-quality versions of those products in seconds, grading them no longer reliably measures what a student knows or can do. The assessment is still measuring something. It is no longer measuring learning.

The Outlook advocates for a shift toward what it calls process-oriented assessment: evaluation that focuses not on what a student produced but on how they produced it. How did the students interact with AI? Did they accept its output or interrogate it? How did they revise AI-generated drafts? How do they explain their reasoning when questioned? What decisions did they make and why? This shift is not simply a response to AI — it reflects a learning-science principle that metacognitive engagement (thinking about your own thinking) is one of the strongest predictors of genuine understanding. AI has made ignoring that principle more expensive.

Implementing process-oriented assessment at scale is technically and logistically challenging. It requires different tools, different teacher training, and, in many systems, different examination frameworks that may require policy changes at the national level. The OECD acknowledges that thisthat this is not a short-term transition. But it is also not optional: as general-purpose AI becomes more capable, assessment systems that do not adapt will increasingly measure AI performance rather than student learning.

The Outlook concludes with specific policy recommendations, structured as four action areas for governments and education systems. These are not aspirational statements. They are derived from the evidence of what conditions produce genuine educational benefit from AI, and translated into the kinds of decisions that ministries, school authorities, and institutions can make.

The OECD’s findings land at a moment when the pace of AI capability development is outrunning the pace of educational adaptation. ChatGPT users doubled between 2024 and 2025. General-purpose AI tools are already more capable than the purpose-built educational systems the Outlook recommends. Students are using them anyway, largely outside institutional control, because they are free, intuitive, and enormously effective at producing the outputs that traditional assessment rewards. The question of whether those outputs reflect genuine learning is one that most assessment systems are not yet equipped to answer.

The OECD’s verdict is not pessimistic. The evidence base for educational AI genuinely working in well-designed intelligent tutoring systems, in collaborative learning contexts, and in teacher productivity is real and growing. A 9-percentage-point improvement in student pass rates when novice tutors received AI pedagogical support is not a trivial finding. A 31% reduction in lesson planning time, if it translates into better lesson quality rather than just faster production, represents a genuine shift in what teaching can look like. The potential is there. The conditions to realise it require deliberate investment.

The harder truth, implicit throughout the Outlook, is that education systems have spent decades struggling to implement even straightforward evidence-based practices — formative feedback, retrieval practice, spaced repetition — at scale. GenAI does not make that implementation challenge simpler. It raises the stakes. Systems that invest in teacher AI pedagogy, process-oriented assessment, and purpose-built educational tools will see genuine learning gains. Those who deploy general-purpose AI without a pedagogical structure will produce a generation of students who are better at generating content and no better at understanding it. The OECD has given governments the evidence they need to choose. The choice now is theirs.