OECD Digital Education Outlook 2026: Exploring Effective Uses of Generative AI in Education

Given the widespread use of GenAI tools by students, including outside school settings and without teacher guidance, and to a lesser extent by teachers themselves, important questions for the education sector include: how does the use of GenAI affect learning? And how can GenAI tools be used to enhance learning? The first part of the OECD Digital Education Outlook 2026 addresses these questions, which are further explored throughout the report.

Does the use of general-purpose tools automatically enhance students learning?

Gašević and Yan (2026) provide an overview of the emerging research on teaching and learning with GenAI and highlight that, depending on their use, GenAI tools can undermine or enhance learning. In some cases, the use of GenAI can be deceiving. For example, GenAI systems may enhance the apparent quality of student work (that is, their performance at educational tasks) without improving their actual learning (their knowledge and skill acquisition) – a paradox illustrated in several studies.

A study of 1 000 high school students in Türkiye in mathematics in grades 9, 10 and 11 examined three practice conditions across six 90-minute sessions: 1) with their course notes and textbook (as usual); 2) with a general-purpose LLM chatbot (“GPT base”); 3) with an educational LLM chatbot (“GPT tutor”). Educational LLMs are configured (or finetuned) to avoid the provision of direct answers and support learning (but there is no indication in this case that it was an adaptive learning tool). This randomised controlled trial analysed the results of the students during practice and noted that the percentage of correct answers to the exercises was much higher for the students using GenAI tools compared to those working by themselves, with a much higher performance for those using the educational chatbot. However, when their knowledge was assessed in a closed-book environment, the performance gains vanished: students who used the general-purpose GPT scored lower than those studying on their own (Bastani et al., 2024) (Figure 1.5). The students who used the educational chatbot performed about as well as their self-study peers. While they enhanced their GenAI skills, one would expect effective digital learning tools to enhance learning, not merely practice performance, which hints to the possible under-performance of self-declared “educational” GenAI tools.

Other studies show similar results in different circumstances, although with less statistical power. Their purpose lies more in explaining why enhanced performance when using general-purpose tools does not necessarily lead to learning. Two studies analysed the metacognitive processes of Chinese higher education students using general-purpose LLMs to revise an essay they had written in English without using a GenAI tool. In a first randomised control trial (Fan et al., 2025), students were assigned to revise their essay in four different ways: 1) alone; 2) with human expert advice; 3) using a check list (and digital tools); 4) with a general-purpose LLM. They found no statistically significant difference in motivation among the conditions, although the three groups with external support had slightly increased motivation. In terms of task performance, the group using the GenAI tool scored highest, but knowledge gains measured by a knowledge test did not improve. In terms of metacognitive processes or self-regulated learning, the group using GenAI performed less metacognitive tasks, especially evaluation and orientation. Another study compared differences between students who revised essays by asking for human advice or used a general-purpose LLM (Chen et al., 2025). They found that those who interacted with human experts sought help in a linear way, following the models of “help seeking” theory: diagnosing what they need help with, asking for help, evaluating the help received, iterating, and then implementing the final help. They showed that when interacting with a general-purpose chatbot, some students tended to ask directly for help and implement immediately the solution received. They would often skip the diagnosis, evaluation and iteration stages. The authors refer to this as “metacognitive laziness”, a way of describing “cognitive offloading”.

Another study is worth mentioning as it casts neuroscientific light on the above findings (Kosmyna et al., 2025). In the United States, students from five universities were asked to write a 20-minute essay under three working conditions: writing by themselves (“brain-only”), with a search engine, or with a general-purpose GenAI tool (ChatGPT). Afterwards, within one hour of the actual essay writing, only 12% of the LLM group could quote something from their essay (exact recall) as opposed to 89% in the two other groups. Even though the essays of the LLM group were well rated, the LLM group also had a lower ability to summarise their essay’s viewpoint, lower level of ownership, and more similar content across essays. Brain imagery suggested a shift in their executive tasks from generating content to supervising the AI-generative content, with lower neural connectivity and involvement. They also showed that writing alone first and then using the LLM preserved higher levels of activation and recall. In contrast, those who started with an LLM and then continued alone had low levels of activation and recall. These findings suggest that initial cognitive activation is crucial before using GenAI.

This evidence highlights that a proportion of students using general-purpose LLMs may take shortcuts, avoiding the productive struggle and cognitive effort needed for learning and durable knowledge and skill acquisition. This aligns with a “fast” rather than more productive “slow” and iterative use of GenAI, as argued by Beghetto (2026), and the notion of “lazy” use of GenAI (Vinchon et al., 2023).

Is it possible to enhance learning outcomes using pedagogical GenAI tools?

While evidence suggests GenAI tools sometimes enhance student performance at the cost of lasting skill and knowledge development, that does not mean positive outcomes are impossible. In fact, several chapters of this Outlook highlight some promising results too (Gašević and Yan, 2026; Li and Hu, 2026; Strauß and Rummel, 2026; Cukurova, 2026; Topali, Ortega-Arranz and Molenaar, 2026; Baker et al., 2026). Two types of uses of educational GenAI should be considered. Some LLM-based GenAI tools are repurposed for education. They have either been “finetuned”, that is to say, partly retrained based on education-relevant data, or “configured” through a series of instructions and prompts on how they should answer user requests.

For example, in the United States, a Stanford-developed GenAI tool called “Tutor CoPilot” finetuned GPT4 based on the ethnographic observations of good teachers providing feedback (Wang et al., 2024). The tool was integrated in an online tutoring platform to assist in real time 900 tutors (rather than teachers) working with 1 800 underserved pupils in the United States. The intervention raised student pass rates by 4% on average, with the largest gains among less experienced tutors (9%) and those previously rated as lower quality (7%). The use of the tool made less difference for the more effective and experienced tutors. Given the effectiveness of tutoring as a learning strategy, this robust study shows promise for GenAI tools to support tutors, and perhaps also less experienced teachers. Other promising uses are also being developed to make teaching more effective (Demszky, 2026).

Another randomised controlled trial in the United States compared learning in an undergraduate physics course at Harvard university delivered through in-person “active learning” versus an AI tutor based on GenAI that was configured to implement “active learning” pedagogical principles online. The idea was to have the same pedagogical practices in both conditions, only the delivery differed. The study found that students learnt significantly more in less time when using the AI tutor, and also felt more engaged and motivated (Kestin et al., 2025) (Figure 1.6). In China, a GenAI tool configured to implement problem-based learning in reading led to enhanced reading performance and motivation compared to the conventional, less personalised approach (Huang et al., 2025).

Strauß and Rummel (2026) reviewed different studies that use LLMs to support collaborative learning. They noted that most studies do not use a general-purpose LLM but have instead usually configured the model to control the behaviour of the LLM and assign it some specific role and behaviour, following previously identified good practices in research on computer-assisted collaborative learning.

While evidence is limited and often lacks statistical power, these studies show that GenAI can enhance learning when tools have been configured to be educational and grounded in expert teaching practice and learning science.

Is it possible to enhance learning outcomes using general-purpose GenAI tools with pedagogical purpose?

A key question is whether we should move exclusively to educational GenAI tools that have been designed from the outset to support learning. While these tools are in principle more likely to support learning than general-purpose models, a pedagogical use of general-purpose GenAI tools can also develop students’ knowledge and skills. In addition, the use of general-purpose GenAI tools will foster students’ AI literacy in a broader way, allowing them to engage with some of the tools they will likely have to use in the labour market.

For example, in the United Kingdom, a study on creativity provides some lessons that could be applied to education (Doshi and Hauser, 2024). The study measured how the use of a general-purpose LLM can enhance people’s creative output as well as the quality of their writing (communication). Participants were asked to write an original eight-sentence creative story and assigned to three groups: 1) they had to write independently, with no GenAI support; 2) they could ask for one GenAI idea (the GenAI was configured to provide a 3-sentence idea); 3) they could ask for 5 GenAI ideas (same principle repeated). The groups that could brainstorm with GenAI outperformed those working alone both in terms of creativity and in terms of quality of the writing, with those receiving more ideas performing best (Figure 1.7). Importantly, compared to the studies presented before, in this case they all wrote the story without using GenAI so that the benefits did not stem from cognitive (and production) offloading. However, as in the “brain imagery” study mentioned earlier, the stories written from GenAI-assisted groups were more similar, pointing to a possible negative side effect and a drop in collective creativity.

A randomised controlled trial in nine secondary schools in Nigeria provided access to Microsoft Copilot based on GPT-4 in an after-school programme while the control group did not have access to it (De Simone et al., 2025). The students in the treatment group worked in pairs and received teacher instructions on how to use Copilot, including prompts. While this was a general-purpose LLM, the students were pedagogically guided on how to use it effectively and put intentionally in a particular pedagogical scenario (peer learning). The study found a positive, medium-sized impact on learning.

In Indonesia, Darmawansah et al. (2025) compared the effects of using ChatGPT rather than search engines as a support tool to prepare and implement argumentative speeches in English (as a foreign language). Far from being left alone, the ChatGPT group were provided with some initial training on prompting, and looked for initial information with ChatGPT. In a second phase, both groups were provided with a collaboration script guiding their collaborative learning. In a third phase, groups produced an argument mapping, with the help of ChatGPT in that group, before the performance of their argumentative speeches. The GenAI group outperformed the “search engine” group in learning gains on argumentation, and in self-reported levels of “critical thinking awareness” and “collaborative tendency”. The GenAI group provided more “backing” (factual information) to its arguments and benefited from the “rebuttals” provided by the GenAI, with results dependent on the quality of student prompting and of their English proficiency. A study in Vietnam making students use ChatGPT and CharacterAI in a collaborative learning setting in English as a foreign language exhibited similar findings (Bui, Ngo and Mai, 2025).

Other studies show that the use of general-purpose LLMs in a pedagogical manner can improve students’ learning outcomes. For example, studies indicate better knowledge acquisition when using an LLM as a teachable agent in computer science (as opposed to studying alone) (Chen et al., 2024).

Can GenAI improve learning outcomes by providing feedback to students?

GenAI holds strong promise for enabling the rapid and scalable generation of feedback to students. Good formative assessment relies on frequent, timely, targeted, and individualised feedback on student work. Given classroom sizes, teachers cannot always provide detailed, personalised feedback to all students, which makes AI-generated feedback a plausible driver of learning outcomes. Gašević and Yan (2026) review the research literature on GenAI-generated feedback and argue that GenAI can support teachers to give better feedback, although it cannot replace human feedback.

The research literature comparing feedback generated by LLMs, usually configured based on marking rubrics and examples of good answers, finds that AI-generated feedback matches human feedback in quality, while acknowledging the shortcomings of human feedback. Heinrich et al. (2025) don’t find much difference on the grading of short answers in political science. Chevalier, Orzech and Stankov (2024) found that students who received feedback based on GPT-4 had similar learning gains as those who always received feedback from human instructors, as is also shown by a meta-analysis of AI feedback covering all forms of AI (Kaliisa et al., 2025). Dai et al. (2024) compared the quality of human and GenAI-generated feedback on task (correctness), process (learning strategies), self-regulation (monitoring learning), and self (personal traits and motivation). On average, GenAI produces more readable and stylistically polished feedback than human educators on written essays. Human feedback tends to be more succinct. GenAI also slightly outperformed humans in the frequency of provision of feedback on process and self-regulation, which support deeper learning and learner autonomy. Despite similar performance on formative feedback, humans and GenAI can have low levels of agreement on the strength and weaknesses of student work (Dai et al., 2024), and thus on the marking of student work (Chevalier, Orzech and Stankov, 2024).

Does this equivalent performance imply that formative feedback on students’ tasks should be delegated to GenAI? Not necessarily. Feedback is only effective if taken seriously and acted upon, which partly depends on its quality but also on its credibility and “motivational” dimension. As noted by Gašević and Yan (2026), comparable performance does not imply pedagogical interchangeability. The growing research literature on the perception of AI feedback by students finds that students perceive human feedback as more credible and meaningful, which makes it more likely to influence their motivation, evaluative judgement, and trust – relational dimensions that remain difficult for GenAI systems to reproduce.

As a result, the consensus among experts points towards a hybrid approach to feedback. The differences in performance between GenAI and human instructors create new opportunities to complement and enhance the effectiveness of teachers on feedback provision. For example, GenAI could generate initial feedback on student work, which teachers can use as suggestions to enhance their own feedback drafts. GenAI could also be used to assess the quality of the feedback provided based on research on high quality feedback. To the extent that their work is done on digital platforms, GenAI could also provide feedback on the process of learning, an important form of feedback that is usually out of reach for teachers. Ultimately, getting this feedback vetted by humans is essential for it to be trusted and effective for students.

What are the initial take-aways on the effects of GenAI on teaching and learning?

In conclusion, using general purpose GenAI tools carries risks, particularly if students use them to avoid the cognitive effort that educational tasks are designed to elicit. However, learning science also recognises that learners may need “scaffolding”, support that is gradually removed as students become more proficient. When used pedagogically, GenAI tools can serve as scaffolding (Strauß and Rummel, 2026). This highlights the importance of teachers to develop pedagogical competences that include GenAI in their teaching and assignments.

Studies that found a positive effect in the use of general-purpose GenAI with pedagogical purpose included some development of students’ GenAI literacy. Countries are starting to develop strategies to this effect (Box 1.1). The OECD and the European Commission have developed an AI literacy framework to support these efforts (European Commission and OECD, 2025).

Even when teachers do not integrate GenAI as part of their teaching, they may still have to adapt their teaching practices as their students can use these tools independently. How should teachers adapt their teaching and assignments so that they continue to yield positive learning results, even for students who use GenAI as a productivity, rather than a learning, tool? While many of these attempts at pedagogical change are undocumented, a few studies offer some insights. For example, Kosar et al. (2024) document how they redesigned their computer science course by: 1) changing their home assignments so they could not be directly completed by a GenAI chatbot; 2) using lab time for oral defences of the produced code, asking for “understanding” questions; 3) turning their mid-term exams into a paper and pencil exam focused on conceptual understanding. In these conditions, the groups with and without access to GenAI had similar learning outcomes. Documenting these efforts of pedagogical redesign internationally would support a more rapid sharing of knowledge around the use of GenAI.

Most countries address students’ GenAI literacy as part of AI literacy, which is often included in digital competence or transversal skills frameworks. Rather than introducing GenAI as a standalone subject, education systems integrate it across subjects and educational levels or, in some cases within a specific subject. Across countries, AI literacy increasingly encompasses understanding the limitations and risks of generative models, not only their use or the development of prompting.

Belgium (Flemish Community) integrates AI-related learning objectives into attainment targets, including prompt formulation and critical evaluation of digital sources. Czechia embeds generative AI within mandatory digital competence across subjects, supported by extensive methodological materials. In France, AI literacy related to generative tools is supported through the Pix platform, which includes modules on prompting, system functioning and data issues. In the United States, Arizona developed a roadmap aligning generative AI literacy goals with state academic standards, explicitly integrating generative AI into subject-level curricula rather than treating it as an add-on.

Some countries have also included GenAI literacy in the “computer science” part of their curriculum. Lithuania includes AI, including generative applications, within informatics for upper secondary education. In Ireland, GenAI-related topics will appear in the Junior Cycle Digital Media and Literacy Short Course (from 2026) and in optional senior-cycle computer science curricula.

In the 2025 European Schoolnet’s survey, 19 out of 20 European countries cite teacher’s GenAI literacy as one of their priorities (European Schoolnet, 2025). GenAI literacy should be understood as the ability to effectively use GenAI to prepare and deliver their teaching as well as having a high-level understanding of GenAI technology. Nearly all systems addressing GenAI provide professional development through online courses, workshops, webinars and communities of practice, although TALIS 2024 results show big variations in the uptake among countries.

Finally, while GenAI holds promise for formative assessment and the quality of the feedback given to students, it may mainly be used to support human instructors rather than to fully automate feedback. The human relationship remains a core element of teaching and learning, and AI-generated feedback does not come with the same credibility and motivational drive as human feedback, even when its quality is equivalent.

While most research has so far focused on general-purpose GenAI tools, future studies should examine the effects of GenAI tools designed for education. These tools seem more promising. Nevertheless, their efficacy in terms of improving learning outcomes or pedagogical competence should be evaluated as a minimal requirement for their adoption. As shown in the reviewed evidence, educational GenAI tools can be both student-facing, teacher-facing, or both. In practice, many tools combine these roles.