The OECD Digital Education Outlook 2026 sums up the case for and against generative AI in schools in one distinction: a tool can raise how well a student performs a task without raising what the student learns. Its executive summary says generative AI can scale personalized support, improve feedback, and save teachers time, but that heavy reliance on tools that hand over answers lowers the mental effort that turns answers into understanding. Evidence is strongest for teacher time savings and weakest for long-term student learning.
Key takeaways
- The OECD report, published in 2026, finds that general-purpose chatbots can improve task output without matching learning gains, and that tools built with explicit teaching methods look more promising.
- A randomized trial of 259 teachers in 68 English secondary schools found a 31 percent cut in planning time, from 81.5 to 56.2 minutes a week, with no loss in lesson quality.
- Student use has grown quickly: 55 percent of French higher education students used generative AI in 2023 and 82 percent in 2025, in the studies the OECD cites.
- On average, seven in ten teachers in the OECD teacher survey believe AI could let students pass off others' work as their own.
What the evidence supports on the benefit side
The OECD report separates benefits by who gains. For teachers, the best-documented result is time. The randomized trial it cites, led by Roy and colleagues in 2024, gave practical guidance to 259 teachers and found planning time fell by an average of 31 percent without lowering the quality of lesson plans and resources. For students, the report describes tutoring, collaborative learning, and creativity support as areas of promise, and says the evidence is still emerging. It also describes a large experiment in rural Brazil in which AI gave feedback and guidance despite intermittent connectivity, a sign that some uses do not depend on fast internet. For school systems, it points to faster classification of courses, admissions support, and exam item generation, with human oversight still required.
What the evidence says on the cost side
The core risk is metacognitive: the planning, checking, and revising that turn an answer into knowledge. The OECD summarizes a randomized trial of Chinese university students revising an English essay in four ways: alone, with human expert advice, with a checklist, or with a general-purpose chatbot. Task performance was highest in the chatbot group, yet knowledge test results did not improve, and that group carried out fewer evaluation and orientation steps. A second study found that students working with a chatbot tended to request help and apply it at once, skipping diagnosis and checking. The authors of that work call this metacognitive laziness. The report treats such findings as a reason for caution, not as proof that every use harms learning, and notes that some studies show gains in both outputs and learning.
How widely students and teachers already use it
Because use is already common, the pros and cons are not hypothetical. The OECD cites a German survey of more than 23,000 higher education students in which 94 percent used AI in 2025, including 65 percent daily or weekly, and a French study in which use among higher education students rose from 55 percent in 2023 to 82 percent in 2025. In the United States, it reports that about 68 percent of teenagers aged 15 to 17 used AI chatbots in 2025, citing Pew Research Center. On the teacher side, the 2024 OECD Teaching and Learning International Survey found that on average 36 percent of lower secondary teachers had used AI in the previous 12 months, from around 75 percent in Singapore and the United Arab Emirates to under 20 percent in France and Japan. Among teacher users, 68 percent used it to learn about and summarize topics, and 64 percent to generate lesson plans.
Pros and cons as the report weighs them
| Claim | What the OECD report cites | Strength of evidence |
|---|---|---|
| Saves teachers time | Randomized trial, 259 teachers, 68 schools, 31 percent less planning time | Moderate: one trial, quality held |
| Personalized tutoring | Prototypes and early trials | Emerging, results mixed |
| Improves student output | Several studies of general-purpose tools | Consistent for task performance |
| Improves student learning | Gains seen with tools built on teaching methods; not with answer-giving tools | Depends on design |
| Weakens effort and skills | Metacognitive findings; teacher concerns about deskilling | Plausible and observed in studies, long-term effects unknown |
| Enables misrepresented work | Seven in ten teachers surveyed worry about it | Perception data, not measured rates |
Teacher concerns that sit beside the time savings
In the same teacher survey, about four in ten teachers agreed that AI may amplify biases, reinforce student misconceptions, or compromise data privacy and security. On average three in four teachers who had not used AI said they lacked the knowledge or skills to teach with it, and one in ten reported that their school prohibits its use in teaching. The OECD also warns that offloading too much marking, feedback, or planning to a tool could erode teachers' own skills and their relationship with students, and notes that students sometimes prefer feedback from a human teacher even when they rate the AI feedback as better.
Questions for a school weighing a tool
- Does the tool give students worked answers, or does it ask them questions and hold back solutions?
- Is there evidence on learning measured without the tool, not only on task scores while using it?
- Can teachers see and adjust how students interact with it?
- What student data does it collect, and where is that data stored?
- Who decides which tasks stay AI-free, and how are students told?
FAQ
Is AI in education mostly good or mostly bad?
The OECD report does not give a single verdict. It finds promise when tools are designed around teaching methods and risk when they act as a shortcut, which is why it frames the policy goal as a learning partner, not a learning shortcut.
Which benefit has the best evidence?
Time saved on lesson and resource planning, based on a randomized trial of 259 teachers in England. Benefits for student learning have less settled evidence.
Do these findings apply to Canadian and American schools?
The report draws on research from many countries, including the United States, and lists Canada (through CMEC) among the delegations that commented on drafts. Provincial and state rules on student data and tool approval still apply to any classroom decision.
Does the report recommend banning chatbots?
No. It recommends moving from generic chatbots toward purpose-built educational tools, co-designed with teachers, while building students' AI literacy.
Related reading
- Challenges and limitations of AI in education
- Impact and potential of AI in education
- AI in the learning experience
- AI and academic integrity
More in AI in education.
Last verified: October 2026
Sources
- OECD Digital Education Outlook 2026: Exploring Effective Uses of Generative AI in Education, OECD Publishing, Paris, 2026
- Results from TALIS 2024: The State of Teaching, OECD, 2025 (as summarized in the Outlook)