A teacher has an hour to prepare a lesson. A generative AI system offers a polished plan in seconds, complete with examples and a quiz. The plan may save time. It may also invent a source, exclude a local language, expose student data if pasted into the prompt, or assess something other than the ability the teacher intended to teach. Speed alone cannot decide whether the lesson is better.
UNESCO’s 2023 Guidance for Generative AI in Education and Research, led by Fengchun Miao and Wayne Holmes, offers a human-centred framework for these decisions. It addresses governments, institutions, providers, teachers, learners and researchers. This guide reads it as a policy and design document, not as an experiment proving that a particular AI tool raises learning outcomes. It also distinguishes the 2023 guidance from UNESCO’s separate 2024 competency frameworks for teachers and students.
Why UNESCO wrote the guidance
Generative AI tools became widely available faster than many education systems could establish privacy, procurement, assessment and training rules. UNESCO’s concern was not only whether text generation worked. It was whether institutions could evaluate tools before placing them into classrooms and research workflows, and whether those tools would strengthen or displace human capacities.
The guidance builds on UNESCO’s 2021 Recommendation on the Ethics of Artificial Intelligence. That earlier recommendation supplies a broader normative setting: human rights, dignity, fairness, inclusion, environmental concerns and accountability. The 2023 document applies those concerns to education and research, where outputs can affect a child’s opportunities, a teacher’s professional judgment and the integrity of published knowledge.
Its scope is deliberately wider than a classroom prompt. UNESCO discusses data protection, age-appropriate access, provider responsibilities, institutional validation, curriculum design, teacher development and the long-term meaning of learning. A policy that says only “allow” or “ban” misses most of those questions. The guidance asks decision makers to specify what problem they are trying to solve and who bears the risk if the system fails.
Guidance is not an outcome study
The publication makes proposals and supplies examples. Those examples should not be read as demonstrated effects of an intervention. A sample use case for feedback on writing shows how a teacher might structure human interaction with an AI tool. It does not prove that every school using that workflow will produce stronger writing or more equitable assessment.
UNESCO itself warns that generative AI is unlikely to solve education’s fundamental problems by itself. A shortage of time, trained staff or inclusive materials can motivate experimentation, but technology cannot create the missing relationships and institutions simply by generating text. The document treats human capacity and collective action as central to educational change.
This distinction matters when an institution invokes UNESCO as an endorsement of a product. A vendor can claim to support teachers, yet still fail a local privacy review or provide inaccurate outputs in the language students use. UNESCO’s framework calls for local validation and continued evaluation. It is a reason to ask harder questions of a product, not a certificate that any named product has passed them.
Human agency in a learning task
The guidance argues that AI should support human thinking rather than replace it. This principle becomes practical when a learning objective is explicit. If students are learning to identify a source’s argument, having a system produce the whole argument before they read the source may remove the practice the task was designed to provide. If the goal is to compare competing explanations, an AI-generated alternative could be useful if students test it against evidence and can explain their own judgment.
UNESCO asks institutions to preserve opportunities for observation, experiment, dialogue and independent reasoning. Those activities are not nostalgic additions to education; they help learners notice errors and build a basis for deciding when an automated answer is plausible. A learner who has never attempted an unaided explanation may struggle to judge a fluent but false one.
Agency also belongs to teachers. A teacher should be able to choose, revise or reject a generated plan. Consultation with teachers and learners before institution-wide deployment matters because they understand the actual workflow and its failure points. Human accountability should remain visible in high-stakes decisions. An opaque model output is not a sufficient explanation for a consequential grade or access decision.
The principle does not require avoiding every tool. It requires identifying what the human should continue to understand and control. In one lesson that may be the selection of evidence; in another it may be the decision about an individual learner’s support. Clear responsibility is especially important when an output looks authoritative because it is well written.
Privacy before convenience
Generative AI prompts can carry names, assignments, health information, unpublished research and other sensitive material. UNESCO places data protection near the start of its proposed governmental steps and asks institutions to understand what providers collect and how the data are used. A teacher who pastes a student’s full case history into a public tool may create a risk that a beautifully worded answer cannot offset.
Practical validation asks where data are stored, whether prompts are retained or reused, who can access them, what contracts apply, and whether the institution can correct or delete records. The answers depend on the provider and jurisdiction. The 2023 UNESCO guidance is not a substitute for current local law or a data-protection assessment. It supplies the questions and the human-rights rationale for asking them.
Research has parallel concerns. An unpublished manuscript, interview transcript or confidential dataset may be inappropriate to upload even when a generated summary would be convenient. A researcher also needs to know whether the tool fabricates references or transforms a source’s meaning. The more consequential the material, the more important it is to verify both the input conditions and the output.
Age and developmental appropriateness
UNESCO calls for considering age limits on independent use of generative AI platforms and for age-appropriate educational design. That is not the same as a universal law created by UNESCO. The guidance invites regulators and institutions to decide how children can use tools safely, with attention to privacy, developmental needs and supervision.
A young learner may benefit from a well-designed explanation or translation while still needing direct instruction, conversation and practice. A platform that responds persuasively to any question can make mistakes that a child is less equipped to detect. The appropriate safeguards include the task design, teacher mediation and the provider’s behavior, not only a number on an age gate.
Policy should distinguish assisted classroom use from independent conversation with a general-purpose platform. It should also consider accessibility and language. A blanket prohibition may remove support for a student who needs it; unrestricted use may expose students to harms that adults have not assessed. UNESCO’s framework asks for proportionate, locally validated arrangements rather than assuming one setting fits every learner.
Inclusion, language and cultural diversity
A model’s apparent fluency in a dominant language can hide poor performance in a minority language or local curriculum context. UNESCO makes linguistic and cultural diversity an explicit part of a human-centred policy. If a generated explanation consistently uses unfamiliar examples, erases a local perspective or treats one tradition as universal, access to the tool is not the same as equitable education.
Evaluation must therefore include the languages, dialects and knowledge practices of the intended learners. Local teachers and researchers can test whether a tool represents their subject accurately. They can also identify where it needs a human supplement or should not be used. A global benchmark score alone cannot show that a tool serves a particular community well.
Inclusion is not limited to output language. Devices, connectivity, disability access, staff training and paid subscriptions shape who receives a benefit. A school may distribute a tool equally in name while learners with less reliable access or less support are left behind. UNESCO’s call for equity points toward testing outcomes across groups and repairing the conditions that make the tool unevenly useful.
Validation before scale
The guidance asks educational institutions to validate the ethical and pedagogical suitability of generative AI systems. This is more demanding than checking that a chatbot can answer sample questions. The institution needs to specify the learning goal, the expected advantage over a no-tech or other available approach, the risks and the evidence that the proposed use delivers the intended result.
Consider an AI writing coach. A pilot could compare student drafts and later unaided revisions, check whether feedback is accurate and understandable, examine effects across language groups, and ask teachers whether the tool saves or adds work. It should record incidents and give users a way to contest errors. That design would test a defined use rather than announcing that “AI improves writing.”
UNESCO’s examples of co-design ask for the domain of knowledge, expected outcomes, appropriate tools, user requirements, human pedagogical methods and ethical risks. Teachers, learners and researchers should help shape those choices. A vendor’s demonstration is useful as a starting point but cannot replace classroom evidence.
Evaluation also has a time dimension. A system can produce a good answer today while weakening the opportunity to practice an ability needed later. Conversely, it can free time for deeper work. To distinguish these possibilities, a pilot needs measures of independent understanding, supported performance and longer-term learning. The 2023 guidance raises the question; local studies must supply the answer.
Assessment and the meaning of authorship
Generative AI changes what a polished assignment can show. If an assessment aims to measure a student’s ability to reason through evidence, an essay composed largely by a tool may not demonstrate that ability. The response need not be a surveillance system or an impossible demand to prove every sentence’s origin. It can be a better assessment design.
A teacher might include a short explanation of source choice, an in-class discussion, drafts with feedback, or a task in which students critique an AI-generated answer. The right design depends on the objective. When the objective is competent use of a tool, responsible prompting, verification and disclosure may be part of the assessed skill. When the objective is unaided fluency, a different condition is needed.
Research authorship has related but distinct obligations. A generated paragraph does not take responsibility for the accuracy of a citation or the consent of a participant. The researcher and institution remain accountable for methods, data and claims. Policies should make permitted assistance and disclosure clear, while recognizing that journal and institutional rules may differ and change.
The 2024 competency frameworks are separate resources
After the 2023 guidance, UNESCO published AI competency frameworks for students and teachers in 2024. They should not be folded back into the earlier document as if its recommendations had been rewritten. The student framework describes twelve competencies across a human-centred mindset, ethics of AI, AI techniques and applications, and AI system design. It can help educators plan what learners should understand and be able to do.
The teacher framework describes fifteen competencies across five dimensions, including a human-centred mindset, ethics, AI foundations and applications, pedagogy, and professional development. It recognizes that teachers need more than a list of prompts. They need judgment about when a tool serves a learning goal, how to protect learners, how to evaluate an output and how to keep developing their practice as tools change.
These frameworks can complement the 2023 guidance: the guidance frames governance and educational design, while the later documents provide more explicit competency structures. Neither turns a tool into a curriculum on its own. A local school or university still needs to choose age-appropriate goals, allocate training time and assess whether learners and staff can use the chosen system responsibly.
What an institution can decide first
Begin with the educational problem, not the product. Name the learner group, task and capability at stake. Ask whether an existing non-AI method already addresses the problem well. If a generative system offers a plausible advantage, identify the data it requires, the errors it can make and who will review them. This follows UNESCO’s emphasis on comparative advantage and proportionate use.
Next, define a limited pilot. Choose outcomes that include both immediate assisted performance and later independent understanding where relevant. Include diverse users in testing. Document the provider’s data practices, accessibility, language performance and complaint route. Make a teacher or researcher responsible for the decision that the system supports.
Then decide what evidence would justify expansion, revision or withdrawal. A policy should be revisited when models, terms of service, curricula or local law change. A tool that passed one task’s evaluation has not automatically passed a different age group, subject or country. Continuous review is part of the design, not an admission that the initial decision failed.
What the evidence does not show
UNESCO does not claim that using generative AI always improves learning, nor that withholding it always protects learning. Its 2023 publication is a normative and practical guide written during rapid technological change. The examples point to possible uses that need local testing. Its principles are broad enough to travel, but application depends on institutions, learners and current tools.
It also does not create a binding global age law, certify an AI vendor or settle academic authorship rules for every institution. Readers should consult current local requirements when making operational decisions. The guidance’s durable contribution is the structure of the questions: human agency, privacy, inclusion, validation, pedagogy and accountability.
The teacher’s lesson plan from the opening may be useful after careful checking and adaptation. It may also be rejected because its examples are false or its data practices unacceptable. What matters is that a human educator can make that decision in light of a clear learning aim, rather than allowing the fluency of an output to make it for them.
Return to the classroom
The teacher can use the saved hour to discuss a difficult idea with students, if the generated material is sound and the workflow protects them. The same hour can be lost correcting errors or managing a tool that does not fit the class. UNESCO’s guidance asks schools to learn which outcome they are actually producing.
A human-centred education system does not measure progress by the number of generated answers. It asks whether learners can understand, question and act; whether teachers retain professional judgment; and whether research remains accountable to evidence. Generative AI may support those aims, but only through decisions that keep the aims in view.
Research references
- Miao, F., & Holmes, W. (2023). Guidance for generative AI in education and research. UNESCO. View source
- UNESCO. (2021). Recommendation on the Ethics of Artificial Intelligence. View source
- UNESCO. (2024). AI competency framework for students. View source
- UNESCO. (2024). AI competency framework for teachers. View source
- UNESCO. (2023). Global education monitoring report: Technology in education, a tool on whose terms? View source