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

While general-purpose GenAI tools can support learning when used with clear educational purpose, current evidence suggests that the development of education-specific GenAI tools may hold even greater promise for improving teacher practice and student learning outcomes. This raises an important question: what does an educational GenAI tool look like?

At a minimum, any educational GenAI tool should generate safe and age-appropriate content, respect users’ privacy and data protection, be explainable and transparent, and mitigates algorithmic bias to the extent possible (OECD, 2023; OECD, 2023). Beyond these safeguards, it should also be “educational” and help teachers to teach more effectively and enable students to learn more or to catch up with peers. This report provides several examples of prototypes or early implementation of such educational GenAI tools.

How can GenAI tutors support personalised learning?

Adaptive learning systems are one of the most used digital learning tools in education systems. They illustrate the “personalisation” agenda associated with AI by providing learners with the possibility to practice and expand their knowledge. These systems usually assess students’ initial knowledge and skills as well as their misconceptions, diagnose types of problems that students should be performing, and adjust the difficulty of problems depending on how students perform (OECD, 2021). Impact evaluations show that these systems are overall effective for learning. Intelligent tutoring systems usually provide feedback and support students to learn, rather than simply telling them whether they were right or wrong. These rule-based AI tutors struggled with unanticipated student inputs or questions though, which limited the scalability and the richness of tutoring interactions. GenAI models, particularly LLMs, overcome these constraints, enabling more engaging and versatile tutoring experiences.

Recent research that compares legacy intelligent tutors to next-generation LLM-driven systems (e.g. Hu, Xu and Graesser, 2025) help articulate a vision for GenAI pedagogical agents. Li and Hu (2026) show how LLMs’ capacity to dynamically generate fluent, contextually appropriate dialogue brings new opportunities to intelligent tutoring systems in terms of adaptability to learner profiles and ability to be applied to different subject matters. They allow for a more flexible tutoring experience, capable of addressing unforeseen questions or novel problem scenarios in real time. GenAI tutors can produce human-like explanations, ask clarification questions, and scaffold student thinking through multi- turn dialogue. Using techniques such as retrieval-augmented generation (RAG) or fine-tuning, they can incorporate up-to-date factual information into their tutoring.

Still, the big challenge is ensuring GenAI tutors are configured or fine-tuned to have interactions that are pedagogically sound, as was the case for effective intelligent tutoring systems not based on GenAI. Through mechanisms like conversation history or explicit memory modules, GenAI tutors can iteratively refine learners’ profiles and adapt tutoring sessions accordingly, for example by adjusting difficulty or revisiting past misconceptions. Keeping the challenge level appropriate for learners is one of the key insights of learning science. GenAI tutors also have the potential to play different pedagogical roles, and to shift between these different roles according to circumstances. They can play the role of: mentors, providing academic guidance; coaches, providing motivational support; peers, with less formal interactions that resemble peer learning. Researchers are also exploring how they could become companions and support learners across a wide range of learning and subjects.

Often drawing on Socratic questioning and related strategies, GenAI tutors can provide scaffolded dialogue: guiding learners to develop their knowledge through carefully sequenced questions rather than delivering answers outright. This approach is rooted in Vygotskian “scaffolding” and “zone of proximal development” theory, where support is provided just beyond the learner’s current ability and gradually withdrawn as competence grows (Vygotsky, 1978). GenAI tutors are particularly well-suited to implementing Socratic questioning as they can generate an extensive range of questions and follow-ups and flexibly rephrase or adjust the difficulty of questions based on learner responses. Li and Hu (2026) take the Socratic Playground as a case study to illustrate the possibilities of GenAI tutors, as well as presenting the underlying AI models that make it possible.

How can GenAI support collaborative learning?

Improved learning outcomes come from sound pedagogical principles. When it comes to collaborative learning, Strauß and Rummel (2026) emphasise that general purpose GenAI systems are unlikely to support effective collaborative learning, just as random human beings, even if knowledgeable, would not automatically make collaboration successful. To be effective, GenAI tutors should play different roles and target group interactions differently. Research identified several roles GenAI played when integrated into collaborative learning environments: 1) it served as a repository of information for the group; 2) it collected information about the group, its collaboration, or its results; 3) it generated additional learning material for the group to use in their reflection, for example a contrasting case; 4) it encouraged active participation as a facilitator positioned as “outside of the group”; 5) it aimed at developing domain-specific knowledge as a dialogue partner, in the Socratic spirit mentioned above; 6) it brought a specific expertise to the group as an artificial group member. By playing these possible formal roles, GenAI chatbots can target different aspects of the collaboration: the cognitive part, by providing knowledge and expertise; the social part, by making sure that students contribute equally and that all voices are heard; the metacognitive part, to encourage students to reflect on the collaborative process or encourage them to make some steps in their reflections.

The small number of studies on GenAI collaborative learning show small- to medium-sized positive results, but importantly, GenAI does not orchestrate collaborative learning on its own – researchers and practitioners do. All the possible roles mentioned above must be assigned by researchers or developers by configuring or fine-tuning the GenAI tools used, building on the accumulated knowledge of computer-assisted collaborative learning. One cannot assume that, by itself, general-purpose LLMs can spontaneously take on different roles to make collaborative learning successful. However, the rise of GenAI is opening new avenues for computer-assisted collaborative learning that show promise if they remain aligned with existing pedagogical knowledge and research.

How can GenAI support teachers and preserve their agency?

As mentioned earlier, teachers often use general-purpose GenAI to support their work, for example the generation of lesson plans and learning materials. Some studies have shown that these activities can bring productivity gains. For example, a randomised controlled trial examined the use of GenAI among 259 teachers across 68 secondary schools in England and found that teachers who received practical guidance to use GenAI reduced their lesson and resource planning time by an average of 31 percent, reducing their weekly average planning time from 81.5 to 56.2 minutes, without compromising the quality of their lesson plans and resources (Roy et al., 2024). However, such uses may come with risks similar to those observed with students, of “cognitive offloading” or “metacognitive laziness”. Teachers and education systems will have to explore and define how, when and in which cases it is appropriate to use these tools. If teachers offload too many of their tasks to GenAI they may stop their professional development and harm their relationship with students. For example, evidence shows that students prefer to receive feedback from human teachers, sometimes even if they rated the feedback from GenAI tools as superior (Gašević and Yan, 2026; Cukurova, 2026). A related question is what design features ensure educational GenAI tools for teachers improve teaching quality, while maintaining teacher agency and autonomy?

Cukurova (2026) proposes a conceptual framework for human-AI interactions and defines three different types: replacement (or full automation), when the AI tool accomplishes a task or sub-task for the teacher; complementarity, when the AI tool amplifies a teacher’s capabilities while the teacher remains actively involved; augmentation, when the human-AI system accomplishes a task with an improved output that outperforms what either the human or the AI system could have performed alone. While replacement may boost teachers’ productivity by saving time, it may also come at a cost in terms of teacher professional development and autonomy but also impoverish human relationships in teaching and learning. Which tasks one would want to automate requires thorough reflection. Complementarity gives a boost to teachers’ productivity while the teacher is still in control, but typically without enhancing teachers’ competence. Augmentation requires interactions during which both teachers and AI evaluate and critique each other’s suggestions and propositions to move towards a shared understanding and mutual development while solving a problem.

Cukurova (2026) presents different examples of human-AI interactions and argues that GenAI tools provide new possibilities for augmenting teachers’ professional abilities. For example, Reza et al. (2025) developed a GenAI tool through a co-creation process with ten mathematics teachers who worked on content creation for adaptive learning platforms. Using a prompting tool, teachers could quickly see how small changes affected the feedback being provided to students. In this case, the human-AI iterations were iterative and evaluative, but reduced the teachers’ perceived workload by 50% and shortened the content development process from several months to a few hours. While GenAI holds promise for more reciprocal exchanges thanks to its dialogic functionalities, GenAI tools with the capacity to push back against inappropriate or suboptimal pedagogical decisions, to question teachers’ assumptions, or to propose alternative perspectives grounded in educational theory and evidence remain to be developed. It will require advances in both the cognitive modelling of teaching expertise and the design of teacher-AI interaction interfaces.

The process to design educational GenAI systems also matters. For example, Topali, Ortega-Arranz and Molenaar (2026) provide an example of the different steps of the human-centred design approach for educational GenAI tools for teachers. By involving teachers and students in the design from the start, first by eliciting teachers’ and students’ uses and expressed needs and suggestions, the developers involved them at different stages of creating a prototype. This involvement ensured the development of tools that align with teacher and student needs but also recognised their autonomy and agency, both during the design and use stages of the GenAI tool.

The exemplar prototype co-designed through this process allows teachers to monitor student-AI interactions and to set the GenAI behaviour to some extent. Teachers have the possibility to easily set the “percentage” of hallucinations of the tool, depending on how much critical thinking they want students to exercise while interacting with it. From the student perspective, this is a general-purpose GenAI chatbot that they use for education. Teachers valued the potential of the tool to increase their insights into student progress and to personalise their feedback, while maintaining pedagogical control. However, some pointed to the risk of added complexity and workload. Meaningful AI integration in teaching requires intentional design of autonomy, where teachers define pedagogical approaches, have settings allowing them to define AI behaviour, and retain responsibility for educational interpretation. GenAI then serves as an assistant within teacher-defined parameters.

Some educational GenAI tools that meet these requirements are already developed and deployed to support teachers, teacher assistants and students in higher education. Baker et al. (2026) provide several examples in Czechia and India, and present an AI teaching assistant that they developed in the United States as a case study: JeepyTA. JeepyTA supports teaching assistants, students and teachers at various tasks: 1) answering logistics questions related to the course (admission, requirements, dates, etc.); 2) providing feedback to student essays based on specific pedagogical goals and rubrics designed by the instructors; 3) responding to student reflections and questions on the course readings and lectures, offering additional clarification, prompting further thinking, and connecting ideas across course materials; 4) providing debugging support to programming code (in courses where programming is not the learning goal); 5) generating a discussion prompt to start weekly discussions among the class and summarising conversations on the discussion forum, sometimes to provide information to teachers, sometimes to make it visible to learners; 6) suggesting ideas for essays and supporting student brainstorming; and acting as different personas during a course.

In order to maintain teachers’ autonomy and control of the tool, instructors or teaching assistants can modify the tool settings so that responses are automated (and go directly to students) or reviewed by them first. This helps prevent the provision of incorrect or misleading information, a particular issue in subject areas where misconceptions are highly present on the web and therefore also in the LLM knowledge base. Instructors or teaching assistants can review flagged responses. If a response is inaccurate, they can discard it and reply directly. If the response is mostly correct but needs refinement, they can edit it before posting. When a response is accurate and well-structured, it can be approved with no changes. This additional layer allows JeepyTA to provide timely support while making sure students receive information that is accurate, relevant, and aligned with the course objectives.

While researching the efficacy of JeepyTA and similar GenAI use cases is still underway, existing studies show that JeepyTA reduced median response times to students from around 7 hours to approximately 2 hours, while keeping human oversight in place. It brought some productivity gains leading to a better student experience. Students rated JeepyTA as comparable to human teacher assistants in clarity, accuracy and professionalism, though it was weaker in motivating students or offering higher-level developmental guidance. When used for feedback on essays, JeepyTA raised the proportion of students achieving top grades on essay assignments from roughly 64% to 95%, illustrating that careful prompt design and alignment with instructor expectations can significantly improve revision quality. However, risks include homogenisation of ideas (seen in brainstorming tasks), overreliance, and the temptation for institutions to reduce human teaching agency.

Together, these empirical findings confirm the potential of GenAI to amplify teacher productivity, improve instructional quality, while keeping human oversight and agency central to system design.