OECD warns AI may raise grades while weakening real learning

(Photo by Mira Kireeva)

Generative AI can lift student performance fast, but those gains can disappear once the tool is removed. That is basically the central warning in the OECD’s “Digital Education Outlook 2026: Exploring Effective Uses of Generative AI in Education” a report which I discussed already. In this article I want to go a bit deeper into this aspect. The report argues that schools and universities should stop confusing stronger task output with stronger learning, especially as chatbots become routine in classrooms and homework.

The OECD anchors that warning in a field experiment in Türkiye. In that study, access to GPT-4 improved short-term performance by 48% with a standard interface and by 127% with a tutoring version. But once access was removed, students performed 17% worse. The report presents that result as evidence that AI can boost apparent success while undermining the knowledge and skills that deep learning depends on.

That conclusion lands at a moment when generative AI is already embedded in daily education work. The OECD says 37% of teachers in its 2024 Teaching and Learning International Survey already use generative AI for work-related tasks such as summarising topics and supporting lesson planning. The same report also cites evidence from England showing a 31% reduction in time spent on lesson and resource planning by secondary science teachers.

What the OECD report found about AI and student learning

The OECD report does not argue that generative AI has no educational value. It however argues that the effect depends on how the tool is built and how students use it. In the Türkiye mathematics trial, around 1,000 high school students in grades 9 to 11 worked across six 90-minute sessions under three conditions: standard study materials, a general-purpose chatbot, or an educational chatbot configured to support learning rather than deliver direct answers. Students using AI performed much better during practice, especially with the tutoring version. But when they later sat a closed-book assessment, the gains largely vanished, and the students using the general-purpose chatbot did worse than peers who had studied on their own.

The OECD report separates “performance at educational tasks” from actual learning, meaning the acquisition of knowledge and skills. A polished answer or faster homework submission may look like progress, but that does not prove a student has understood the material well enough to retain it or apply it later.

Can AI improve grades but hurt learning?

The OECD’s answer is yes. Its report says overreliance on tools that provide direct answers can reduce active engagement and displace the cognitive effort needed for durable learning. In plain terms, students can complete more tasks correctly while doing less of the mental work that helps ideas stick.

That is why the report keeps returning to the distinction between AI as a shortcut and AI as a teaching tool. Used as a shortcut, generative AI can flatten the struggle that often produces understanding. Used more carefully, especially in tutoring systems designed to prompt, question and guide, it can support learning (rather than replace it). The OECD does not frame the technology as inherently harmful, however it does say that misuse and poor design are the real risk here.

Where teachers are already using generative AI

The report also makes clear that AI adoption is there alright. Teachers are already using these systems, mainly for preparation and productivity. Among teachers who use AI, 68% report using it to learn about or summarize topics they teach, while 64% use it to generate lesson plans. A smaller share use it to review student participation or performance data, or to assess and grade student work.

That use is not free of tension. Around 40% of teachers, on average, say AI helps them support students individually, and around half say it helps create or improve lesson plans. At the same time, seven in ten teachers believe AI could help students misrepresent others’ work as their own. Around four in ten say AI may amplify bias, reinforce misconceptions, or compromise privacy and security. The same report says three in four teachers report lacking the knowledge or skills to teach using AI.

Taken together, the OECD’s position is certainly not anti-AI. Instead it offers a labor and pedagogy warning: AI may save time, but it can also erode teacher autonomy and professional judgement when schools start treating automation as a substitute for expertise.

Why the OECD wants education-specific AI systems

One of the clearest policy conclusions in the report is that education should not rely too heavily on generic chatbots. The OECD calls for a shift toward educational generative AI systems designed with teachers, so educators can monitor student interactions and shape how AI is used in learning. It also says human judgement, feedback and oversight should stay at the centre of AI use in education.

That is a direct response to the pattern the report documents. General-purpose tools can make students faster and more fluent in the moment. But the OECD argues that systems built for education should be structured to support skill acquisition, not just answer generation. The report points to intelligent tutoring systems and guided educational chatbots as more promising directions than open-ended answer engines.

What schools and universities may need to change in 2026

The report points toward three immediate changes. First, schools may need to rethink assessment. If AI can raise the quality of visible output without improving understanding, then educators will need more closed-book tasks, oral assessments, staged drafting, and other formats that test comprehension rather than surface polish.

Second, teacher training is necessary, they need to learn the best practices when using AI in the classroom and we need to address faculty training gaps at universities. Not surprisingly the OECD says many teachers still feel unprepared to use AI well, even as adoption rises. That gap matters because the report’s central finding is about use rather than about access. The same tool can either support learning or weaken it, depending on how it is introduced, framed and monitored.

Third, policymakers will need governance that goes beyond enthusiasm. The OECD calls for practical guidance, stronger teacher capacity, and policy frameworks that address access, privacy, ethics and bias. In that sense, the report is less a celebration of classroom AI than a demand for restraint and design discipline.

---

---