OECD's 2026 Verdict: AI Lifts Task Scores, Then Learning Drops 17%
The OECD's 2026 Digital Education Outlook finds GenAI raises task completion by 48% but pupil performance falls 17% once the tool is removed, unless teaching is redesigned around it.
A performance illusion, quantified
The OECD's Digital Education Outlook 2026, published in January and drawing renewed attention as schools return this term, puts a hard number on something many teachers have suspected for a while: pupils who use generative AI look like they are learning more than they are.
Across the studies the OECD reviewed, students using GenAI were 48 percent more successful at completing set tasks than those working without it. But when the AI was taken away and the same students were tested again, their performance fell by 17 percent. The tool had inflated the appearance of competence without building the underlying skill.
The OECD's explanation is about metacognition, not motivation. When a chatbot supplies the reasoning steps, students spend less mental effort turning an answer into understanding. Task completion and genuine learning start to pull apart, and a mark scheme or a homework tracker will not catch the difference. Only the moment the scaffolding disappears does the gap show up.
"If designed or used without pedagogical guidance, outsourcing tasks to GenAI simply enhances performance with no real learning gains."
The adoption numbers behind the warning
The report lands alongside other 2026 data showing how far AI use has already spread. Across OECD countries, 37 percent of lower-secondary teachers reported using AI for their job in 2024, and 57 percent said it helped them write or improve lesson plans. Separately, Microsoft's 2026 education report, based on a survey of 3,345 respondents including UK educators, found 88 percent of teachers had already used AI for school-related purposes, with 76 percent reporting increased use over the past year.
That adoption is running well ahead of pedagogical design. The OECD notes that 72 percent of lower-secondary teachers believe AI can harm academic integrity by letting students pass off its work as their own, an integrity worry that sits uneasily alongside how routinely AI is now used in lesson prep and, increasingly, in pupils' own work.
What actually produces learning gains
The OECD is not arguing schools should retreat from GenAI. It reports that when AI is used with clear pedagogical intent, or when teaching is redesigned around its availability, studies show small-to-medium gains in subject learning and more substantial improvements in critical thinking and collaboration. The gains show up specifically when AI supports group work and discussion rather than replacing the back-and-forth between students.
The report frames three distinct classroom roles for GenAI, each carrying different risks. As a tutor, it can work well when built around questioning rather than answers, such as Socratic-style tools that probe reasoning instead of supplying it. As a partner, it can support collaborative tasks without displacing peer interaction. As an assistant, it saves teachers time on planning and admin, the use case where adoption is already highest and the evidence for benefit is least contested.
The distinction the OECD draws is between AI that does the cognitive work for a student and AI that is structured to make the student do the cognitive work themselves, with the tool checking, prompting or scaffolding rather than solving.
Why this matters more than another adoption survey
UK schools have spent the past year absorbing guidance on detection, disclosure and policy: JCQ rules on assessment malpractice, DfE product safety expectations, safeguarding duties under KCSIE. Almost none of that guidance addresses what the OECD's data speaks to directly, which is whether AI use in day-to-day classwork, entirely outside formal assessment, is actually building the skills it appears to build.
A department that lets pupils use a chatbot to check their work or get unstuck on homework, with no change to how that homework is set or reviewed, is precisely the scenario the OECD's 48/17 figures describe. The task gets done. The learning does not necessarily follow. And because the shortfall only appears once the AI is removed, a school could run an entire term believing its AI rollout is working, right up until an unaided assessment or exam disagrees.
What to do
Build in unaided checkpoints. If AI is available for homework or drafting, pair it with regular no-AI retrieval tasks so gaps in genuine understanding surface early, not at the exam.
Redesign tasks, don't just permit tools. The OECD's gains appear where teaching was adapted to AI's presence, not where AI was simply allowed into an unchanged task.
Favour Socratic and formative uses over answer-generation. Tools that question and prompt protect metacognitive effort in a way that answer-giving tools do not.
What to watch
Watch for the gap between coursework performance and unaided assessment results this year, particularly in departments with high AI uptake for homework and revision. A widening gap is the practical, classroom-level version of the OECD's laboratory finding, and it is the clearest early warning sign that use has outpaced design.
Sources: OECD Digital Education Outlook 2026, Generative AI in education: what the OECD 2026 data shows, Microsoft's New AI in Education Report