Reimagining Teaching in an Accelerating World

3. Embracing the opportunities of artificial intelligence and educational technology

The AI moment in education

For years, AI in education hummed quietly in the background. It powered grading systems, adaptive tests and learning analytics but few outside the edtech field ever noticed. Then came generative AI, and everything shifted. Unlike earlier tools, GenAI didn’t arrive through ministries or school procurement plans. It landed straight on students’ phones and teachers’ laptops. That changed everything.

General-purpose GenAI tools are not designed to help students learn; they are built to do things for us. They write essays, solve problems, translate texts and generate code. Unsurprisingly, the first reaction from education systems was concern - about cheating, academic integrity and whether grading could still be trusted. But that response revealed something deeper: education’s long-standing dependence on credentials as proof of learning, rather than learning itself.

And yet, the genie is not going back into the bottle. Students already use GenAI widely outside classrooms, more so as they grow older. Parents use it at work and at home. Teachers use it to draft lesson plans, quizzes and feedback. Researchers use it to refine language, explore data, and solve problems that once took months or years. GenAI is no longer an experiment; it is part of education’s lived reality.

Its capabilities are undeniable. General-purpose GenAI can engage in dialogue, adapt to context, and synthesise information at a scale no human can match. In science, its potential is already transformative. AlphaFold, a generative AI system, predicted the structures of nearly all known proteins in a single year - a task that once consumed entire PhDs. That breakthrough has reshaped medicine and refocused human effort on harder questions. GenAI doesn’t just retrieve knowledge; it accelerates its production.

But this power comes with real risks. GenAI systems “hallucinate”, producing plausible but false answers. They are not fully consistent because they are shaped by probabilities rather than understanding, reflecting the perspectives of their training data. And even when the answers sound authoritative, they do not understand in any human sense. Their outputs demand scrutiny - often more, not less, than traditional AI systems.

For education, the stakes are high. When GenAI performs tasks for students, it risks pulling them out of the zone where real learning happens, when struggle, effort and growth take place. The same risk applies to teachers and leaders who outsource feedback, assessment, or even strategic thinking to machines, hollowing out processes whose value lies in process rather than output.

Yet used wisely, GenAI offers real opportunity. It can personalise the cognitive struggle of learning without sacrificing social cohesion, support students through conversational tutoring, and provide explanations when human help is unavailable. It can help teachers adapt instruction, analyse learning data through natural language, and focus their energy where it matters most: relationships, judgement and inspiration. It can expand access for second-language learners, rural students and those historically underserved.

The choice facing education systems is not really whether GenAI should be allowed, but how to ensure its pedagogical intent and shape its most effective use. GenAI will remain part of our societies. Education’s task is to ensure it strengthens learning, rather than quietly replacing it.

In 2024, when the TALIS data were collected, about a third of teachers were already using AI for work, mostly for planning lessons and learning about teaching topics. The uptake has probably grown since then. Out of those teachers using AI in 2024, TALIS data also reveal a quarter employ it to assess or mark students’ work. This represents a fundamental shift in how students are evaluated and raises a fundamental question: is it a good development or not?

The attraction of AI is obvious as teachers regularly say they are short on time. Between planning lessons, teaching and managing classrooms, and marking assessments, the workload can be relentless. So, if AI software claims it can assess hundreds of maths tests in seconds, it is no surprise that some teachers jump at the chance. AI is simply better than humans at doing certain things, particularly tasks requiring repetitive precision. But outsourcing assessment risks that teachers lose the connection to their students and the understanding of what their students can do. Out of teachers who use AI, at least 50% in Azerbaijan, Kazakhstan, North Macedonia, South Africa, Türkiye and Viet Nam mark schoolwork with it, according to TALIS data. In Uzbekistan, the share reaches an astonishing 85%. That is fast take-up given that ChatGPT was only launched at the end of 2022. It is also a major change that, in most cases, has occurred without formal policies or consistent training.

For critics, the speed at which AI has been adopted is a risk, with technology companies competing for a global market with little regard for oversight. But for supporters, AI is the solution to numerous issues that teachers contend with on a daily basis. Across the OECD, 40% of teachers report that too much marking is a source of stress. By instantly assessing grammar, coherence and structure, proponents of AI argue it allows teachers to spend more time on lesson planning and mentoring students, freeing them from “never ending” marking.

But there is a lack of robust, large-scale independent evidence that proves AI enhances student learning. AI may struggle to capture subjective elements of assessment, such as creativity or originality and it may also have difficulty understanding context or cultural influences. In addition, there are privacy and data security concerns. And left unchecked, AI may spread misinformation, and algorithmic grading may amplify racial, gender and socio-economic biases rather than erase them.

These issues raise concerns about what checks have been carried out before implementing AI in the classroom. The effectiveness of algorithms varies considerably. Does the AI grade fairly across different student demographics? Does its feedback genuinely help students improve? Education leaders and schools should know the answers before deploying the software, particularly as some surveys suggest that nearly two-thirds of adults oppose the use of technology for marking.

Another profound shift to consider is the relationship between teacher and student. Traditionally, marking has been about more than just correcting errors. It is a form of dialogue with teachers’ comments offering insight and encouragement to students. If replaced by AI, does it weaken the personalised relationship between teacher and student? Could some students make less effort if they realise AI, and not teachers, are marking their work?

Figure 22. AI is useful for some tasks more than others

Of teachers who use AI, the share who report using it to do the following tasks

Source: OECD, TALIS 2024 Database, Table 1.60.

These are questions that policymakers should urgently consider answering, particularly in countries where the use of AI in teaching has really taken off in ways other than marking. For example, around three-quarters of teachers in Singapore and the United Arab Emirates report using AI in their general work, according to TALIS data. Training on AI is highest in Singapore (76%) and lowest in France (9%). And among teachers who use AI, some 73% report leveraging it to efficiently learn about and summarise topics, and 69% use it to generate lesson plans, on average, according to TALIS. Given the speed at which AI is racing through classrooms, it is no surprise that policymakers are scrambling to keep up.

It is also important to note that many teachers are less than enthusiastic about AI. TALIS data show that about half of teachers believe AI should not be used in teaching at all. And one in ten report that their schools have policies banning the use of AI in teaching. The data reveal major divides between schools and teachers on the use of AI tools in classrooms. This is reflected by countries policies too, with some nations embracing AI and others urging caution.

Figure 23. Lack of knowledge and skills is the most common barrier to AI use

Note: Out of teachers who don’t use AI, the share who report the following barriers to using AI to teach

Source: OECD, TALIS 2024 Database, Table 1.63.

In the end, AI is not a magic power. It is just an amazing accelerator and an incredible amplifier: It will amplify good ideas and good educational practice in the same way it amplifies bad ideas and bad practice. AI can advance education to create a truly level playing field that embraces the diversity of learners so that every learner will succeed. Or it can superempower those with the right resources and dispositions and leave others further behind. AI can empower teachers to become creative designers of innovative learning experiences. Or it can disempower them to become slaves of scripted lessons plans or algorithms they no longer understand. AI can help reduce human bias through better data, but it can also amplify and entrench bias. AI can connect people across geographic, linguistic or cultural boundaries, but it can also sort us into echo-chambers that amplify certain views and insulate people from divergent thinking. AI is in principle neutral, but it is always in the hands of people who are not neutral – developers and users.

It is clear that AI’s role in education is anything but settled, but education leaders must not avoid it or be afraid of it. Instead, schools need to ensure that AI is applied in contexts where it truly enhances learning. Ultimately, students and teachers must understand how AI systems operate in educational contexts, including how they evaluate work, generate feedback, personalise learning pathways and support accessibility. Schools must ensure fairness, so the AI avoids bias and ensures equitable treatment. And while AI can assist with grading, human judgement remains crucial, particularly for creative or subjective work – as well as for motivational purposes. Importantly, countries where teachers are more engaged in professional learning about AI are also the countries where teachers are using it more in their classrooms. This suggests that public policy can make a significant contribution to helping teachers feel prepared and comfortable with these technologies.

Guidance to seize GenAI Opportunities

The pace at which AI technologies are developing makes it increasingly difficult to say, with confidence, what works, in which context, and with what effect. For teachers and policymakers alike, this is unfamiliar terrain. AI is evolving faster than our typical traditional evidence-gathering cycles can accommodate.

As a result, we must look outward across systems, countries and experiences, to understand what may work best in the future. In this context, comparative insight becomes a necessity, not a luxury.

To that end, the OECD Secretariat has undertaken extensive cross-country analysis of how GenAI is already being adopted and used in education systems around the world, through its Smart Data and Digital Technology in Education project.

The insights that follow reflect the OECD Secretariat’s current thinking on how GenAI can be developed, deployed and assessed in ways that serve education’s core mission: to support pedagogy, deepen meaningful learning, foster equal opportunities, and promote human knowledge, skills and well-being.

These insights, drawn from the OECD Digital Education Outlook 2026: Exploring Effective Uses of Generative AI in Education, are organised into five sections, each addressing a key dimension of how GenAI can be harnessed responsibly and effectively in education systems today.

• Human-centred teaching and learning with GenAI;

• Investment in GenAI-related educational R&D;

• Enabling policy environment for trustworthy GenAI in education;

• Effective GenAI infrastructure for all;

• Human agency and human in the loop.

1. Human-centred teaching, learning and assessment with GenAI

Learning and teaching primarily aim to develop valued human knowledge and skills such as independent thinking and foundational skills across subjects, while fostering well-being and student agency. This may start without GenAI, then with educational GenAI, before incorporating general-purpose GenAI. Used selectively and purposefully for pedagogical reasons, GenAI can enrich learning and not replace cognitive effort or weaken the human relationships at the heart of education.

For teachers, GenAI offers a powerful set of supports. It can help prepare lessons, personalise curricula, design assignments and exams, draft feedback, and even assist with grading. It can also support administrative work such as documentation and routine reporting. In the best-case scenario, GenAI buys teachers time to build relationships, give richer feedback, and focus on the human judgement and care that no machine can replicate. In some contexts, it may even support less experienced tutors in real time, raising the overall quality of instruction.

Students are already using GenAI as a cognitive companion. Large language models have become a kind of supercharged search engine. They explain concepts, answer follow-up questions, brainstorm ideas, polish drafts, debug code and suggest solutions. When designed specifically for learning, educational GenAI tools can scaffold understanding much like intelligent tutoring systems, guiding students step by step rather than simply delivering finished answers. At the system level, GenAI can help develop assessments, expand automated evaluation to oral tasks, tag learning resources, and provide administrative support across institutions.

But here’s the catch: Learning does not happen when the machine does the thinking for you. Writing, problem-solving, and teaching are not just tasks; they are cognitive workouts. Outsource them to GenAI, you may get a polished product, but you lose the learning that comes from struggle. This is the risk of skill atrophy: when tools designed to help us think end up thinking for us instead. The danger is especially acute with general-purpose GenAI, which is built to perform tasks efficiently, not to support learning.

This raises thorny questions. When is it appropriate to use GenAI, and for what? Should young children use general-purpose GenAI for homework? Probably not, unless teachers expect this use. But educational AI tools are a different story. They can be purpose-built, safe and effective, such as tutoring systems for math or reading, which have long been used successfully in primary education and are now becoming more conversational through GenAI interfaces. However, to learn with GenAI, students need enough foundational knowledge to judge its outputs critically. Without that, the tool can lead to be harmful.

There is another risk that matters just as much: the erosion of human relationships. Education is not only about acquiring knowledge; it is about learning how to interact with others, such as receiving feedback, earning approval and navigating disagreement. If teachers stop giving personal feedback, or if students increasingly interact with bots instead of peers and adults, education risks becoming efficient academically but meaningless.

The lesson is simple and urgent. GenAI can elevate human learning, but not act as a replacement. It should strengthen human connection, judgement and understanding - not quietly take their place. If we use GenAI with intention and care, it can become a tool that reinforces what makes education profoundly human rather than eroding it.

Doing a task with GenAI isn’t the same as learning from it

The use of (general-purpose) GenAI may enhance performance in a study task without leading to knowledge acquisition. Learning requires the active and often social work of students’ mind. Much of the art of curriculum design and pedagogy is to create tasks that will elicit student reasoning and give them opportunities to advance it. None of this can happen if the student avoids cognitive effort or thinking. For example, translating a text into a foreign language with an LLM is unlikely to correlate with improved mastery of the foreign language. This has also been shown with the use of general-purpose GenAI as an adaptive learning system in mathematics or for improving one’s written text. What is in principle the product of learning (the proficient achievement of a task) may no longer reflect actual learning if GenAI is involved. This does not imply that one cannot learn by using general-purpose GenAI tools, or that educational GenAI tools have the same effect, but the use of GenAI tools should be guided by clear pedagogical objectives or educational purposes. Jurisdictions should consider informing and sensitising parents and caretakers to this risk of using general-purpose generative AI tools when supporting their children.

As machines get smarter, human skills matter more

Education institutions aim to develop skills that are socially valued for humans, whether these skills become GenAI capabilities or not. The OECD has developed “AI capability indicators” designed to assess and compare AI advancements against human abilities. They include creativity, critical thinking and metacognition, and show that AI capabilities are progressing in these valued human skills. Even if GenAI could outperform humans on these skills, it would not imply that humans should stop valuing and developing them. For example, GenAI can be used to support the creative process and the acquisition of creative skills. However, it has been shown to also reduce collective originality, probably because GenAI “averages” what is publicly available on a topic in its training data set and will tend to make different but similar suggestions in response to similar requests. Nurturing individual creativity with no use of GenAI may thus remain socially beneficial.

Similarly, because GenAI can generate incorrect or biased answers, and also help identify different perspectives on a subject, teachers may use GenAI outputs to help develop their students’ critical thinking. However, students must also develop critical thinking with no GenAI and learn to think like scientists and navigate uncertainty and ambiguity, make and test assumptions, and learn to weigh the strength of evidence.

Requesting students to provide personal insights on a subject, to research and work on topics they are personally interested in or to try to understand how they and others feel (empathy) can help mitigate the risk of overreliance on GenAI tools as well as nurture personal expression, independent thinking, appreciation of learning and of the world, and make educational tasks less standardised and more engaging for students.

Students must learn to think (without GenAI) before they learn to prompt

As experts using general-purpose GenAI tools know, their outputs vary from excellent to irrelevant – and the reason they know is that they are expert. Students are not experts yet, and they should reach a certain level of expertise in what they are studying to know when GenAI outputs should be discarded, improved or used. Expertise is correlated with age, especially so when students are younger, but proficiency in what one is supposed to learn trumps age when it comes to the use of GenAI. In some instances, teachers may ask students to use GenAI as a scaffold to help them with aspects of the task they do not master yet and to focus their attention on the acquisition of skills or knowledge that they are closer to mastering. Foundational knowledge that must be mastered for a given field should be assessed with no GenAI support. For aspects of the educational task that are peripheral to the key learning objectives, the use of general-purpose GenAI as a performance-enhancing tool is not necessarily a problem though. Assessment will also have to pay more attention to learning processes rather than just student products.

Assessing student knowledge with or without GenAI can be done in a controlled environment, either by having supervised paper and pencil exams or by using digital tools that block or document student use of GenAI tools. It is more difficult for tasks that are not meant to be supervised. Beyond the design of assignments that cannot easily be performed by GenAI, several models are possible, such as oral examination, the request of short videos rather than, say, short essays, a focus on conceptual understanding rather than practical performances. Alternatively, non-supervised assignments can also be used to develop students’ skills in creating or engaging with GenAI tools.

GenAI works best when teachers design the task.

It is a new responsibility for teachers to guide students in using GenAI from time to time as part of their studies. They should design some learning activities that involve the use of GenAI, whether general-purpose or educational. The younger and less proficient students are, the more appropriate educational (as opposed to general-purpose) GenAI tools are. Teachers should design tasks that require cognitive effort and work despite the use of GenAI tools. Teachers should also support students on how to use GenAI tools: they may request intermediate outputs, prompts, or have access to learning analytics documenting the student use of GenAI.

A productive collaboration with GenAI tools can take different forms. It could be about receiving feedback on one’s work, but also about evaluating GenAI tools’ outputs in a mindful and critical way. While education systems value students’ ability to answer questions and provide solutions, they may now emphasise more students’ ability to ask good questions and elicit good answers, by refining their questioning, identifying incorrect or misleading information, knowledge gaps or missing perspectives in answers, so as to get to balanced and satisfactory answers from GenAI tools (and also humans). They may also put more emphasis on the metacognitive skills they are developing.

While the appropriate use of GenAI tools may support the learning of academic knowledge and skills in any subject, students also need to develop a certain level of literacy in GenAI itself as part of their digital literacy. The OECD AI literacy framework describes AI literacy as four competences: engaging, creating, managing and designing AI. This literacy includes engagement with a variety of GenAI tools, including for non-academic purposes, but also encompasses understanding the basics of how GenAI works.

Preparing students to use GenAI does of course not imply that they should use GenAI all the time. As for any other tool for learning, teachers should provide students with a variety of learning activities, experiences and scenarios, including some with no use of GenAI tools (or technology), and ensure diversity and balance in learning activities. This is essential for student wellbeing and engagement.

No GenAI lesson plan replaces professional judgement

Like their students, teachers should exercise their evaluative judgement and critical thinking on what GenAI tools produce and ensure the produced contents are appropriate for their students in their given context. These time-saving tools are useful to provide teaching ideas, to contextualise courses or personalise them to specific classes, but they may also include inaccurate knowledge, misunderstandings of the specific learning contexts, etc. For example, while teachers should mobilise their knowledge of how students have and have not been progressing in the classroom to decide which topics need revisiting as a priority and which do not, they could use a GenAI tool to generate warm-up activities bringing these topics together. Ideally, if available, teachers should use educational GenAI tools that are aligned with their course contents and curriculum rather than general-purpose GenAI tools.

Algorithms may suggest; teachers must decide. Teachers should not outsource assessment to AI

Providing feedback and assessing the quality of student work is time-consuming and sometimes considered a “burden”. However, assessment is not an administrative task but an essential part of teaching and learning. While technological support can help, especially when teaching many students, teachers’ ability to provide personal feedback and motivate their students is bound to their personal knowledge of students’ work, their strengths and weaknesses, and a good understanding of their classes.

GenAI tools open new perspectives for student work assessment. For example, they ease the assessment of assignments and exams with rubrics that include multiple explicit assessment dimensions and proficiency descriptions. They also allow for the generation of more detailed feedback, which is sometimes found more useful than human feedback (at least when the recipient does not know the source). While teachers can take advantage of those new possibilities, and occasionally automate assessment tasks or subtasks, they should continue to regularly and primarily assess their student work themselves and always check and take responsibility for any mark that could have a stake for students’ future. The design of GenAI assessment tools could for example ensure that feedback or marks are actively validated and thus facilitate teacher agency. GenAI can also be used by students to receive immediate feedback on their work when instructors are unavailable, but this cannot replace the professional human feedback of teachers. The social dimension of feedback, that is, the knowledge that another person is engaging with the student work, may change the way students react, feel, and behave when they produce it. The authenticity of feedback is not only important from the student perspective, it is also a professional and moral imperative for teachers.

2. Investment in GenAI-related educational R&D

As general-purpose GenAI tools are not designed specifically for learning, education systems may consider supporting the development of tools that aim to enhance teaching and learning. Investing in educational GenAI grounded in educational research, co-created with teachers and students, and supported by rigorous research on their effectiveness can open new possibilities for educational improvement.

The OECD AI Principles promote use of AI that is innovative and trustworthy and that respects human rights and democratic values. Adopted in May 2019, and updated in May 2022, they set standards for AI that are practical and flexible enough to stand the test of time. If there is one lesson the OECD AI Principles made unmistakably clear, it is this: you don’t stumble into good AI by accident. You build it deliberately, collaboratively and with the long view in mind. It is the same lesson for GenAI. To make GenAI work for education rather than merely in education, the people who live in classrooms every day must be involved from the start.

Too often, teachers and students are invited into the process only at the end - when the code is written, the architecture fixed, and the purpose already decided. They are asked to “pilot” or “test” tools whose fundamental design choices are no longer up for debate. The same pattern is now repeating itself with generative AI.

Yet teachers and students are not just users; they are repositories of practical wisdom. Their insights rarely appear in academic papers or engineering specifications, but they are precisely what will determine whether a GenAI tool becomes transformative or forgotten. That is why genuine co-creation matters. Not consultation as an afterthought, but partnership at the design table, when goals, safeguards and pedagogical assumptions are still fluid.

This is also where responsibility enters the picture. As GenAI becomes more powerful and more embedded in education, EdTech companies are no longer just software vendors; they are shaping learning environments. With that influence comes an obligation: to design tools that respect the complexity of teaching and learning, that support human judgement rather than replace it, and that align with educational values.

At the same time, there are still major gaps in educational research on GenAI, including about what works, under what conditions and at what cost. These are not questions any one country, company or discipline can answer alone. Instead, they call for international co-operation, shared evidence and a genuine willingness to learn across systems. Only then can we be certain that GenAI will strengthen education rather than distort it.

EdTech should build GenAI for classrooms, not just consumer markets

Because general-purpose GenAI tools are performance- rather than learning-enhancers, GenAI developers and technology companies should develop GenAI tools meant for student learning. Those tools can be designed to be used by students, by teachers, or other education stakeholders. Their big difference when compared to general-purpose GenAI tools is that they are designed to improve specific aspects of learning. Student-facing learning tools should start from and incorporate insights from educational research and pedagogical knowledge while those for teachers should allow them to use them to support their pedagogical choices, either before or during teaching. For example, teachers may be allowed to easily increase the level of hallucinations of a GenAI system used by their students to support their critical thinking. The educational orientation of GenAI cannot be limited to educational intentions or be merely declarative though: ideally, it should undergo impact (or other) evaluations demonstrating their educational value. International co-operation on educational efficacy would help advance this agenda. Governments can support companies by making some datasets based on their curriculum and education research available, as is the case in England (United Kingdom) with the development of a “content store” allowing companies to finetune their GenAI models.

If it isn’t tested for safety, bias, and impact, it isn’t ready for schools

Because of their generative nature, GenAI tools bring more safety risks when it comes to age-inappropriate content, content not aligned with the curriculum, with educational research or pedagogical knowledge, etc. As key actors in the advancement of trustworthy GenAI, developers, and EdTech companies are responsible for the quality, safety, and law-enforcement of their tools. While existing datasets reflect past and current societal perspectives and do not reflect all different voices across the globe, technology companies and governments should collaborate to remedy to the extent possible this legacy and curate AI training datasets that include a greater variety of perspectives and cultures. This was, for example, the ambition of Apertus, an open-source language model developed in Switzerland. Relevant communities should be involved as appropriate in how their languages and cultures are represented in these new datasets. As some forms of algorithmic bias are impossible to remove, a regular audit and benchmark of educational LLMs could help identify and mitigate those biases.

The strongest GenAI will be co-designed with teachers and learners

A few countries have launched new initiatives to get end users involved before the late stages of the AI tool development process. The objective is to ensure the utility of these tools for education and their subsequent adoption. In the Netherlands, the government-funded National Laboratory on AI (NOLAI), which brings together teachers and school representatives with academics, EdTech companies and policy makers to design AI tools in a co-creation process. The French ministry of education also made it a requirement of its multi-step public procurement process of AI tools that companies co-create their products with teachers and students through its AI Innovation partnership (P2IA).

We need global yardsticks - and serious research - to measure what GenAI actually changes in learning

On the research side, there is still a dearth of information about what students and teachers do with GenAI tools for educational purposes, what types of tools they use, how they feel when using them, as well as what jurisdictions and schools make available to them. Governments should co-operate to develop international metrics of availability and use of these tools. Research on the effectiveness of GenAI for learning and for productivity gain is emerging, but most available experimental research currently lacks statistical power and is focused on short interventions that are hardly generalisable to a repeated use of GenAI. Moreover, many research studies on GenAI tools do not provide enough specificity about the pedagogical purpose and context of their use, making it difficult to replicate or scale studies showing promising effects. While educational GenAI tools are being developed, robust research on their effectiveness is needed. All these efforts should contribute to the international knowledge base and provide enough information about context. Internationally coordinated educational research studies may also allow one to better understand how to adapt the use of GenAI to educational contexts.

Emerging research on the use of (mainly general-purpose) GenAI tools points to the need to change learning and assessment tasks so they are adapted to the availability of these tools. Educational GenAI tools may mitigate those risks and enhance learning, but still need to be meaningfully integrated in teachers’ pedagogical practices. Teachers should be encouraged and given the possibility to initiate or participate in research projects about the educational use of GenAI tools. This does not imply that all research projects have to be practice-engaged though. International sharing of educational practices and evidence on how to adapt educational tasks and pedagogies so that their use enhances (or at least does not hamper) learning will help to more quickly understand and adjust to the digital transformation and seize the opportunities brought by GenAI in education.

While research access to educational administrative datasets is possible, it entails privacy risks. The creation of synthetic datasets keeping the characteristics of the real datasets without retaining any personal information has now become possible, largely thanks to GenAI. This possibility should be more systematically exploited by educational jurisdictions to broaden datasets for education research. Initiatives from other sectors could be examined for inspiration. The creation of these privacy-preserving synthetic datasets should not be limited to administrative data but also encompass learning data that are typically collected by private companies, at least partly through public contracts.

Universities should train GenAI builders, not just GenAI users

GenAI is also transforming scientific research: specialised GenAI tools are increasingly available and used to support almost all the stages of research, from literature reviews to experimentation through hypothesis generation. Higher education institutions should prepare future researchers to use these specialised scientific GenAI tools. Governments are expected to also encourage the development of GenAI tools that support educational research specifically and explore the new avenues they may open for education research.

3. Enabling policy environment for trustworthy GenAI in education

Policy and regulatory frameworks are designed to protect students and support learning while enabling innovation. Combined with ongoing stakeholder dialogue, clear expectations on privacy, safety, bias testing, age-appropriateness, transparency and alignment with educational objectives can create an enabling environment for the trustworthy and meaningful use of GenAI in education.

The challenges GenAI poses are largely the same ones we already know from non-generative AI, and they line up squarely with the OECD AI Principles: transparency and explainability so systems don’t become black boxes; robustness, security, and safety so they don’t fail at scale; and accountability so responsibility never evaporates into the cloud. Add to that the OECD’s long-standing standards on privacy, data protection, and digital security risk management, and the foundations are already in place.

But principles alone do not govern technologies - policies do. That is why the OECD AI Principles call on governments to do two things at once that are often seen as opposites: to shape an enabling environment for AI and to continually review and adapt regulatory frameworks so innovation and competition can flourish, without compromising trust. In other words, the goal is not to slow AI down, but to steer it in the right direction.

All the guardrails laid out in the 2023 OECD-EI Guidelines (Opportunities, guidelines and guardrails for effective and equitable use of AI in education) remain fully relevant in the age of GenAI: ethics, safety, privacy, transparency, explainability, and the capacity for human negotiation and oversight. What differs is how countries choose to enforce them. Some embed expectations in public procurement rules, while others issue guidance, standards or regulatory requirements.

Across the OECD, most countries have already built strong data protection and privacy regimes into their digital education governance, often with heightened safeguards for minors and age-specific thresholds for different kinds of data. In the European Union, new legislation now sets clear obligations for the development and deployment of AI systems, including those used in education. Developers of educational GenAI tools are no longer operating in a regulatory vacuum.

Yet education is not just another market. Learning objectives must come first. Regulatory frameworks need to protect learners while also creating the right incentives for EdTech companies to build tools that are affordable, effective and equitable, rather than flashy and exclusive. Get this balance wrong, and you either stifle innovation or open the door to systems that undermine trust and widen gaps.

This is where international co-operation becomes a force multiplier. A more harmonised global approach can reduce market fragmentation, lower costs and give educators and learners access to better tools. This approach also ensures that developers are not forced to navigate a maze of incompatible rules.

Black-box AI has no place in education

The first, basic level of transparency is for users to know that they interact with a GenAI tool rather than with a human, unless it is self-evident. A second aspect lies in how they protect the privacy, data and intellectual property of their users. While GenAI tools do usually protect the privacy of the data provided by users in their queries when used for a fee, this is often not the case for their free versions which may typically be used by students, teachers, and education staff (unless GenAI tools are provided by their educational institution). GenAI providers should ensure transparent information on whether the user data and queries are collected or used to further train their models. They should also aim to be as transparent as possible about how the tools operate, the criteria they use to generate their content as well as the datasets on which they were trained. Educational institutions or jurisdictions running privacy audits on cloud applications for education could consider doing the same for GenAI on cloud.

Jurisdictions and educational institutions that provide GenAI tools to education stakeholders should also have the same level of transparency with the end users and be transparent about their own uses. To the extent possible, if they use them to make or inform decisions, they should explain how the tools are used, and also the methods and criteria used to inform their decision. Transparency may also require social dialogue and negotiation of those criteria with teachers and their representative organisations, students (and their representative organisations when they exist), parents and families, and all other stakeholders. Transparency alone does indeed not guarantee social acceptance of the social tools.

Explainability is difficult to fully achieve with GenAI, even more so than with non-generative AI. Most GenAI systems are proprietary and trained on datasets that are not disclosed, making it challenging to understand the factors that shape their outputs. Educational GenAI tools are built on these underlying models and therefore inherit this opacity. Nevertheless, explainability remains possible at certain levels: some educational GenAI tools are designed to provide information about why and how specific content was generated, including the features considered or the steps taken. Such approaches can help strengthen trust and support responsible use, even if they only offer partial explainability.

AI safety and standards demand global cooperation

Governments share a series of common challenges to ensure the quality, safety and education quality of GenAI tools: safety, data protection, alignment with pedagogy, bias, etc. They could for example work internationally, in collaboration with GenAI companies, to develop international safety expectations for GenAI tools used in education, as is for example the case in England, partially in EU countries through the AI act, or in other countries through their procurement process. They may ask GenAI companies to comply or at least be transparent about their degree of compliance with these expectations.

Used for educational purposes by education institutions, GenAI tools should generate safe and age-appropriate content. This risk is usually much lower with non-generative AI given its “rules-based” nature. Beyond GenAI tools used in or for education, non-educational GenAI tools can pose safety risks for students and other stakeholders, for example through new forms of cyber-bullying using their generative affordances in offensive ways. Jurisdictions and education institutions should monitor those risks and adapt education systems’ and institutions’ regulations and codes of ethics accordingly.

GenAI tools also pose greater risks of algorithmic bias. For non-generative AI, the risks often materialise as AI tools not performing equally well for different subpopulations (and thus discriminating against some). GenAI tools amplify the risks of presentational bias: images, texts or other content that reproduce views now considered as inaccurate or inappropriate (or locally considered as such). Governments and companies may work together to identify different layers of agreement on these so that educational GenAI is aligned with their curriculum and what they consider acceptable views within the education sphere. Higher education institutions may fine-tune GenAI systems so they align with the course content and knowledge base suggested by their academic staff.

Similarly, while all countries may not agree on what “good pedagogy” is, given the variety of contexts, they may still agree on some basics anchored in educational research, including learning science and all its other sub-disciplines, and develop and share expectations for GenAI tools. This could lead to a multi-layered set of expectations that different countries and jurisdictions could choose from.

Responsible GenAI use requires continuous guidance

Students and teachers need to receive clear guidelines about how GenAI tools can and should be used in the learning process. Jurisdictions and educational institutions should consider having a public position about the appropriate and inappropriate uses of GenAI, following a dialogue with their students, teachers, staff and their representative organisations. Education technology is a socio-technical system, and a public position based on social dialogue will also facilitate further agreement on the responsible uses of GenAI by students as well as teachers. This echoes the OECD AI Principles recommendation to “work closely with stakeholders to promote the responsible use of AI”. These public positions should be updated regularly given the evolution of these technologies.

This includes providing clear and plain-language information and supports to parents and caregivers, recognising their crucial role in guiding children’s use of GenAI outside of school. These supports may include glossaries, guidance documents, and practical examples to help families understand the possibilities and limitations of GenAI tools, the associated risks of overreliance, hallucinations or biased outputs, and the importance of balancing independent learning with (hopefully teacher-guided) GenAI-assisted tasks. Engaging parents in this dialogue can contribute to more coherent expectations across school and home environments, and support more equitable and responsible use of GenAI among students. Some countries such as India also provided parents with GenAI applications that support them in reading and inventing new stories to their children, which is typically associated with stronger learning outcomes.

Many countries provide such guidance, which should be regularly updated based on emerging evidence on the actual use of and research on GenAI in education.

Every AI decision must be open to human challenge

The principle of having a human alternative or final resort is as important for GenAI as for non-generative AI. GenAI tools should play a limited and clearly defined role in decision-making, especially in situations where their use may affect individual’s rights (for example, assessment, certification, admissions, disciplinary matters, or comparable tasks). While GenAI tools could be the first source of information for some questions or problems, providing the convenience of being available at any time of the day while humans may not be, users should be able to ask humans to provide those answers when they feel GenAI answers are unsatisfactory. In the (rare) event that GenAI tools are used to make decisions, those affected by those decisions or recommendations should be able to challenge them and get humans back in the loop, as noted by the OECD AI Principles.

4. Effective GenAI infrastructure for all

Jurisdictions typically ensure that all students and teachers can benefit from GenAI thanks to digital infrastructure and support accessible to all (devices, connectivity, digital resources and professional learning opportunities). Providing curriculum-aligned GenAI resources, alternative solutions where divides persist, and sustained professional learning enables effective and meaningful uses of GenAI in education.

The promise of generative AI in education does not begin with algorithms. It begins with plumbing. Without reliable, affordable digital infrastructure, which includes devices, connectivity, platforms and skilled teachers, GenAI will remain a luxury good rather than a learning tool.

GenAI needs high-quality connectivity with enough speed and low enough latency to allow students and teachers to interact, ask follow-up questions and explore ideas without friction. Yet in many countries, universal access to such connectivity remains more aspiration than reality. Broadband divides between urban and rural areas, rich and poor households, and connected and disconnected schools, still shape who gets to learn with this new technology.

Closing these divides is an educational imperative. Ensuring affordable, high-quality digital infrastructure everywhere is the first step toward equal opportunity in the age of GenAI. That is why the OECD Recommendation on Broadband Connectivity emphasises the importance of access to networks and the need for digital skills to make meaningful use of applications like GenAI tools.

But infrastructure alone is not enough. This new phase of AI challenges earlier assumptions in education. Earlier, non-generative educational AI tools, such as intelligent tutoring systems, were typically institution-owned and tightly controlled. General-purpose GenAI is different. It is freely available to anyone with an internet connection, while its most powerful versions are increasingly paywalled. That reality raises uncomfortable questions about equal opportunities for all, oversight and who benefits first from the most capable tools.

Education systems therefore need to think in two lanes at once. The first is supervised, intentional use: GenAI integrated into teacher-designed learning activities, aligned with educational goals and supported by professional judgement. The second is unsupervised, organic use: students and teachers engaging with GenAI outside institutional planning, often invisibly, but with real consequences for learning, assessment and equity.

To navigate both, skills matter as much as access. Students and teachers need the capacity to use GenAI thoughtfully for learning. That capacity grows through skilled teachers, knowledgeable administrators and peers who model effective practice. It also requires investment in support structures like coaches, mentors and technical staff, who can help translate powerful tools into sound pedagogy and responsible administration.

Professional learning must evolve accordingly. Teacher education, ongoing professional development and communities of practice all have a role to play in building confidence, judgement and shared norms around GenAI use. The goal is not to turn educators into engineers, but to empower them as designers of learning in a world where AI is always on.

In the end, GenAI will only be as equitable and effective as the foundations beneath it. Build the infrastructure, invest in people, and the technology can amplify learning. Skip those steps, and GenAI risks becoming just another force that widens the gap between those who do and don’t have access.

No devices, no connectivity - no AI learning

This is a pre-condition for any use of AI for educational purposes. Closing broadband connectivity divides, as well as broader digital divides is a challenge that goes beyond education, but remains an important priority to ensure the adoption and effective use of GenAI in education. The OECD Recommendation on Broadband Connectivity provides recommendations to help bridge these digital divides, promote widespread broadband deployment and foster competition and investment in broadband development. The quality and reliability of connectivity is a key element of access. In education, quality devices need to be adequate for learning applications and have sufficient performance, including long enough battery life, be long lasting and abide by sustainability standards. This does not imply that public authorities should pay for devices and connectivity for all students. Leases, loans, zero-rating of specific sites and other solutions have been successfully experienced in the past years in different countries and jurisdictions. If publicly provided, devices should be affordable, measured by the total cost of ownership. Among many possible models, Ceibal in Uruguay has developed a long experience in the provision and support of educational devices for all their public-school students.

GenAI belongs alongside other digital resources and textbooks as a core learning tool

General-purpose LLMs and other tools are usually not educational in purpose. Jurisdictions and institutions should consider providing access to GenAI tools that are aligned with their curriculum, courses, pedagogical preferences and other educational objectives. This provision is also a way to provide tools that respect privacy and data protection. Typically, they should provide a mix of general-purpose and educational GenAI tools. Depending on the task, they may consider whether non-generative educational AI tools would not be more effective or appropriate. One challenge for education lies in the cost of licensing privacy-preserving GenAI tools for students and teachers, but many countries, jurisdictions or education institutions already do so.

Some countries, jurisdictions and higher education institutions have adopted different approaches. England has developed a content store providing EdTech developers resources to align their educational GenAI models to the English curriculum and to research evidence about teaching. France is developing a “sovereign AI” for teachers and students that aligns to its values and curriculum and respects privacy. Korea has introduced AI-powered digital teaching and learning materials that blend traditional resources with pedagogically validated AI digital tools. These materials are designed for targeted use by students and teachers across specific school levels and subject areas. As part of its AI Leap programme, Estonia is working with GenAI companies and licensed general-purpose GenAI tools to teachers and educational version to their high schools students. Some higher education institutions across the world have also either developed their own GenAI tools to support their students and teachers or provided access to commercial models while ensuring their privacy, intellectual property and data protection.

Given the current reality of persistent connectivity and digital divides, jurisdictions and educational institutions can consider the “AI unplugged” model. This approach consists of using AI tools and resources that work with existing digital connectivity infrastructure (usually mobile phone or smartphones) and level of digital skills. Brazil has for example successfully used it in some areas of the country where devices and connectivity are scarcely available. In the case of GenAI, Small Language Models run on mobile phones offline and can provide students and teachers with a first access to AI language models, although with limited capabilities compared to the large ones. GenAI companies should develop small GenAI models for educational purposes, and jurisdictions should make education stakeholders aware of them where appropriate.

Teachers need ongoing AI literacy, not one-off training

Teachers’ GenAI pedagogical literacy can come from a variety of professional learning opportunities, in pre-service teacher education as well as during in-service learning opportunities. As for any other professional knowledge, it comes from learning by doing and thus from the access and use of GenAI tools for some professional purposes as well as from different types of real-time human support, which may range from “digital” or “AI” champions and professional learning communities through to innovation projects focused on the use of GenAI for teaching and student learning. Teachers should acquire enough professional knowledge to make informed decisions as to when GenAI tools should and should not be used, and how they can best support student learning.

For teachers, GenAI literacy also encompasses the ability to use GenAI tools in their teaching to support their students’ learning and gaining a basic understanding of GenAI as a technology, including its strengths and limitations, and learning how to effectively and ethically use it. The AI literacy framework published by the European Commission and the OECD could support this effort.

Jurisdictions should support teachers and administrators to keep abreast of new developments in the landscape of educational GenAI tools and inform and curate them so they can more easily select and appropriately use them. This curation should take into account non-generative educational AI tools that may be more effective or appropriate than GenAI tools for similar use cases. Teachers also need support to navigate the ethical, legal and pedagogical issues that arise when students and families use generative AI beyond the classroom.

GenAI should empower teachers, not just watch them

AI tools such as classroom analytics provide teachers with feedback about some aspects of their teaching, with no intrusion of a human observer and the possible uncomfortable feeling of being judged, but also with significant risks of unethical use by their employers. For example, these tools can record how much time teachers have spoken during class, what mix of pedagogical activities or learning strategies they have used, etc. GenAI tools can help turn the analysis of these classroom observations into suggestions and recommendations, sometimes in real-time, but most often after class. This feedback would allow teachers to reflect on and improve their practice. While research is still underway, this GenAI feedback may be more impactful when it is discussed with a coach or mentor who has not observed the class (nor made the initial suggestions). Jurisdictions and developers have to ensure these tools do serve professional learning and not surveillance, for example by limiting access to the collected information to the interested teachers only.

5. Human agency and humans in the loop

Human judgment, relationships and responsibility should not be inadvertently replaced by GenAI, even though some educational tasks can be purposefully automated. Designing educational GenAI tools and practices that preserve human oversight, and fostering dialogue on what may and may not be automated, can support more trustworthy, equitable and educationally meaningful uses of GenAI.

Both generative and non-generative AI can automate many of the routine tasks that weigh down education systems. It can help grade, schedule, draft, summarise and organise. But while AI can automate tasks, it cannot automate the teaching profession itself. Education is built on judgement, relationships, responsibility and trust - which cannot be safely outsourced to machines.

That is why the OECD-EI Guidelines insist on keeping a “human in the loop” and, whenever possible, a human alternative alongside digital tools. This is not a technical preference. It reflects the human-centred vision at the heart of the OECD AI Principles, and it runs through everything discussed in this report. This final section brings that principle into sharp focus.

As previously mentioned, used wisely, GenAI can take on time-consuming work. This creates time for teachers to focus on learning, mentoring, motivation and relationships. But that promise only holds if humans remain active agents, not passive approvers of machine-generated output.

Students must remain agents of their own learning. Learning does not happen when answers arrive friction-free; it happens when students think, struggle, revise and reflect. Teachers, too, must remain agents of their teaching, and not just consumers of AI-generated content. Otherwise, the risk is skill atrophy. Education institutions might even consider periodic “automation and agency” audits to ask questions like: What have we automated? Why? And where must human judgement remain non-negotiable?

Design matters here. Developers play a crucial role in protecting human agency. GenAI tools should be built to keep teachers in the driver’s seat. This approach will allow them to co-design workflows, customise systems, and validate or revise AI outputs. Human agency is not just the ability to override AI; it must be actively supported by design choices that invite reflection, decision-making and responsibility.

The higher the stakes, the more essential this becomes. GenAI’s fluency and versatility can create a dangerous illusion of infallibility. Even systems that perform impressively most of the time can hallucinate, make errors or behave inconsistently. When AI-generated outputs affect grades, opportunities or life trajectories, human scrutiny is not optional. It is an ethical necessity.

That is why education systems need an ongoing, honest conversation about what should, and should not, be automated. Keeping humans at the centre is not the job of any single actor. It is a shared responsibility across governments, institutions, teachers, students, researchers and the companies building these tools.

Get this balance right and GenAI can make education more meaningful, effective, equitable and trustworthy. Get it wrong, and we risk replacing a deeply human enterprise with an efficient but empty loop.

Encouraging innovation inside and outside of school

When other sectors experience flat-lining productivity, they turn to innovation. In education, however, systematic investment in innovation is still rare. Public health research budgets in OECD countries are many times larger than education research budgets. This disparity says a lot about the role we expect knowledge to play in advancing practice. Beyond the volume of innovation, its relevance and quality are also important, alongside how fast ideas translate into positive impact.

GenAI is an innovation that is already beginning to transform how we think, work and learn. But innovation can be more difficult to achieve in hierarchical structures that are geared towards rewarding compliance with rules and regulations. One policy approach to foster innovation in education has been to increase autonomy, diversity and competition among educational institutions. But evidence of the benefits of this approach remains patchy.

Innovation in the instructional system is another challenge. There is a long history of introducing new methods in education - whether it was television, video, digital whiteboards or computers in the hope of radically improving teaching and the effectiveness of schooling, only to find, at best, incremental change achieved at higher cost and complexity.

But the bigger issue is that, even where good education research and knowledge exists, many practitioners just do not believe that the problems they face can be solved by science and research. Too many teachers believe that good teaching is an individual art based on inspiration and talent, and not a set of skills you can acquire during a career. Yet it would be a mistake to blame just teachers for that. This problem often goes back to policy due to a lack of incentives and resources to codify professional knowledge and knowhow. In many countries, the room for non-teaching working time is far too limited for teachers to engage in knowledge creation. Because education has not been able to build a professional body of practice, or even a common scientific language as other professions have, practice is often not articulated, invisible, isolated and difficult to transfer. Investing in better knowledge - and disseminating that knowledge widely must become a priority; it promises to deliver huge rewards.

It is also important to create a more level playing field for innovation in schools. Governments can help strengthen professional autonomy and a collaborative culture where great ideas are refined and shared. Governments can also help with funding, and can offer incentives that raise the profile of, and demand for, what works. But governments alone can only do so much. Silicon Valley works because governments created the conditions for innovation, not because governments innovate. Similarly, governments cannot innovate in the classroom; they can only help by opening systems so that there is an innovation-friendly climate where transformative ideas can bloom. That means encouraging innovation within the system and making it open to creative ideas from outside. More of that needs to be happening.

Policymakers often view education industries as providers of goods and services to schools. They tend to underappreciate that innovation in education is also changing the very environment in which schools operate. In particular, technology-based innovations open up schools to the outside world, both the digital world and the social environment. They also bring new actors into the system, including education industries with their own ideas, views and dreams about what a brighter future could hold.

It is difficult for education systems to treat industry as a valuable partner. Fears of a perceived “marketisation” of education, or the displacement of teachers by computers, often endanger what could be a fruitful dialogue. At the same time, we should be more demanding of the education industry. Most of our children would not voluntarily play with the kinds of software that companies are still able to sell to schools. Is innovation in the education industry as dynamic as it should or could be? Can we break the cartel of a few large suppliers of educational resources who use an army of salespeople to sell their services to a fragmented market? Can we overcome the slow sales cycles, where buyers have to deal with layers and layers of people all “in charge”?

Is it possible to create a business culture for managing innovation in school systems? At the moment, it is easier for administrators to buy new tools and systems, as well as use existing staff, because this can appear to be a quicker and more straightforward option. Yet, the treatment of teacher time as a sunk cost means people see no benefit to saving this time. It is worthwhile to explore how industry can help the education sector close the productivity gap with new tools and new practices, organisations and technology.

Finally, education systems need to better identify key agents of change and champion them; and they need to find more effective ways of scaling and disseminating innovations. That is also about finding better ways to recognise, reward and celebrate success, to do whatever is possible to make it easier for innovators to take risks and encourage the emergence of new ideas.