OECD: Higher education already runs on GenAI, policy and governance lag behind
This month, the OECD published a new Education Spotlight, number 23 in the series, entitled "Policies supporting responsible and systematic GenAI adoption in higher education". It is a synthesis of the OECD Digital Education Outlook 2026, current surveys from dozens of countries, and an expert meeting held on 6 March 2026 in Bratislava. The conclusion is clear: in just three years, generative AI has gone from novelty to near-standard use in higher education, while policy and management at most institutions have yet to begin.
Figures that demonstrate the speed
In the United Kingdom, the proportion of undergraduate students using GenAI rose from 66% in 2024 to 92% in 2025 and 95% in 2026. In Germany, student usage increased from 63% in 2023 to over 90% in 2025. On average across the EU, 72% of students say they have used GenAI in the past three months, with 53% specifically for formal education. The differences between countries are significant: in Estonia and Slovenia, usage is around 87 to 89%, while in Romania and Türkiye it is around 52 to 53%.
Unevenly distributed and mostly superficial
Usage is not evenly distributed. Students in science, technology, and economics use GenAI more often than students in the arts and humanities—in Germany, 96% compared to 79%. There is also a gender and socioeconomic gap: men and students from higher social classes use AI more frequently. Most applications remain superficial, such as summarising, editing, and translating, rather than the kind of in-depth use that truly enhances learning. Worldwide, only 17% of lecturers describe themselves as advanced or expert users.
The recurring concerns
The report lists the familiar concerns. Data protection, as institutions often have no insight into what students and lecturers entrust to commercial tools. Academic integrity, since detection tools have proven unreliable. Equal opportunities, both between institutions with and without budgets and in the biases hidden within the models themselves. Hallucinations, where students and lecturers struggle to assess what is and is not accurate. Cognitive shift, where outsourcing routine work to AI can actually hinder the development of one’s own skills. And digital sovereignty, as a small number of large providers determine where the data of students and researchers ends up.
Three reasons why the current approach is unsustainable
The OECD cites three characteristics that make the current model unsustainable. First, students and lecturers mainly rely on free, general consumer tools, as institutions themselves offer very little. In the United Kingdom, in 2026 only 38% of institutions actively provided AI tools for students. Second, formal policy is lacking in many places; worldwide, only 19% of the institutions surveyed have an AI policy, and in a sample of 163 British institutions, fewer than two-thirds could produce a public policy document. Third, usage remains focused on completing tasks rather than learning, as the tools are designed to generate output, not to provide educational guidance.
What governments around the world are doing now
The OECD identifies five types of policy responses. Countries such as Australia, Ireland, and the European Commission are working on concrete guidelines for responsible use. Italy was one of the first countries to sign a national ChatGPT Edu agreement for all affiliated universities; France, through the ILaaS consortium of twenty institutions, is pooling shared computing and language model capacity; and in the Netherlands, SURF supports institutions in assessing AI suppliers for compliance and is working on a safe alternative to public chatbots. Additionally, countries are investing in AI literacy for lecturers, including in Germany, Switzerland, and South Korea, in evidence gathering through pilots in Australia and Canada, and in the development of education-specific AI tools such as Ethel from ETH Zürich, which bases its answers on its own course material rather than on the training data of a general model.
Why this is relevant for EduGPT and our other GPT products
What the OECD describes as a problem is exactly the gap that EduGPT aims to bridge, as we at CivAI recognise. Institutions that do not offer their own AI tools see students and lecturers turn to consumer versions without any underlying policy, thereby losing control over where data ends up and how AI is used in education. EduGPT runs on European hosting, gives institutions ownership of student and lecturer data, and has been designed from day one as an education-specific tool rather than a general language model with an educational layer on top. Precisely the combination of policy, management, and pedagogical design that the OECD itself identifies as the missing link.
Want to read more? The full report Policies supporting responsible and systematic GenAI adoption in higher education is available online for free, OECD, Education Spotlights number 23, July 2026, published under CC BY 4.0. You can find more about EduGPT at edugpt.nl.
Illustrative image, created with AI