When UNESCO published its Guidance for Generative AI in Education and Research in 2023, ChatGPT had just crossed 100 million monthly users and exactly one country on the planet, China, had released any kind of regulation on generative AI. That gap between adoption and governance is the opening premise of the document, and it’s the reason Miao and Holmes wrote it.
The guidance builds on UNESCO’s 2021 Recommendation on the Ethics of Artificial Intelligence and takes a position that’s both clear and deliberately cautious: generative AI in education should serve human capabilities. It should not replace them.
Two years later, I think the document holds up better than most policy writing from that era. Not all of it. Some parts feel dated, and some of the recommendations are too broad to be useful without local adaptation. But the core principles are solid, and a few of the warnings have aged particularly well.
Miao and Holmes start with a foundational claim that’s easy to overlook because it sounds obvious: GenAI doesn’t understand anything. It produces new content, text, images, video, music, code, by analyzing statistical patterns in massive datasets scraped from the internet. It predicts what words or pixels are likely to come next based on probability. That’s it. The authors put it plainly: “While GenAI can produce new content, it cannot generate new ideas or solutions to real-world challenges, as it does not understand real-world objects or social relations that underpin language” (p. 8).
The guidance identifies eight controversies around generative AI in education, and most of them still resonate. The list includes widening digital poverty, the speed of new tools outpacing regulation, use of content without consent, the black-box problem of unexplainable models, AI-generated content polluting the internet with fabricated information, the lack of genuine understanding behind outputs, the narrowing of diverse opinions, and the growing threat of deepfakes.
What I find most useful about this list is the equity argument running through it. As Miao and Holmes stated: “the rapid pervasion of GenAI in technologically advanced countries and regions has accelerated exponentially the generation and processing of data, and has simultaneously intensified the concentration of AI wealth in the Global North” (p. 14). The training data reflects Global North values and norms. Marginalized communities with minimal digital presence, both in the Global South and within wealthy nations, are doubly harmed: their voices aren’t in the training data, and they lack access to the tools built on it.
That concern about whose worldview GenAI amplifies connects to something I’ve covered in a post on the digital plastic metaphor by Roe, Furze, and Perkins (2025), who argue that AI-generated content accumulates like plastic, persistent, culturally contaminating, and unevenly distributed. Miao and Holmes flag a related risk that’s gotten worse since 2023: as GenAI content gets posted online, future models will train on AI-generated text that already contains biases and errors. The feedback loop accelerates.
The regulatory roadmap in the guidance lays out seven steps for governments, from endorsing data protection regulations to building GenAI-specific frameworks and reflecting on long-term implications for how knowledge is understood and learning is defined. The most concrete recommendation is on age restrictions: a minimum of 13 for independent use of GenAI platforms, with acknowledgment that the EU’s GDPR sets the threshold at 16.
Miao and Holmes call for countries to move beyond self-reported age verification, which is essentially no verification at all. They also raise a data privacy problem that hasn’t gotten enough attention: once personal information is embedded in a trained model, it’s technically impossible to remove it. That creates a fundamental conflict with GDPR’s right to be forgotten.
For education systems specifically, the document outlines eight policy measures, and they’re solid in principle: promote inclusion and linguistic diversity, protect human agency, monitor and validate GenAI systems, develop AI competencies for learners, build teacher and researcher capacity, promote plural opinions, test locally relevant applications with evidence, and review long-term implications.
The teacher capacity piece is one I want to flag. Miao and Holmes report that only seven countries (China, Finland, Georgia, Qatar, Spain, Thailand, and Türkiye) had developed or were developing teacher training frameworks on AI as of 2023. That number tells you everything about the gap between policy ambition and ground-level readiness. I’ve covered UNESCO’s 2024 AI competency framework for students in a previous post, and the 2023 guidance is the foundation it builds on. The progression from general principles to structured competencies shows how policy can evolve when the starting framework is strong.
The co-design framework for classroom use is probably the most practical section. Miao and Holmes map six dimensions across five use cases: research support, teaching facilitation, one-on-one coaching for foundational skills, inquiry and project-based learning, and support for learners with special needs. Each use case considers appropriate knowledge domains, expected outcomes, suitable GenAI tools, user requirements, pedagogical methods with example prompts, and ethical risks. It’s a template, not a recipe. And that’s the right approach. Classrooms vary too much for prescriptive solutions, but they benefit from structured thinking about what GenAI can and can’t do within specific learning contexts.
AI and Assessment
The section on assessment deserves attention as well. Miao and Holmes argue that GenAI’s ability to produce competent essays and pass certain exams forces a rethinking of what education should measure. They propose four categories of learning outcomes: values (human-centered design and use of technology), foundational knowledge and skills (literacy, numeracy, basic science, still essential), higher-order thinking skills (critical evaluation of AI outputs, distinguishing factual from conceptual knowledge), and vocational skills for working alongside AI.
That framework hasn’t been widely adopted, but the logic behind it is sound. If GenAI can generate a passable essay, then the essay alone can’t be the measurement. What matters is what the student understands about the ideas in it, and whether they can evaluate what the AI gave them. That argument runs parallel to what Shaw and Nave (2026) later called cognitive surrender, the tendency for people to accept AI-generated outputs without critically processing them.
Miao and Holmes saw it coming. They write: “While GenAI may be used to challenge and extend human thinking, it should not be allowed to usurp human thinking” (p. 24). And the concluding section drives the point home with a phrase that’s stuck with me: “Some experts have characterized the use of GenAI to generate text in this way as ‘writing without thinking'” (p. 38).
Limitations
I do have criticisms. The guidance is cautious to the point of vagueness in places. “Build teacher capacity” is a direction, not a plan. “Promote inclusion” is a value, not a strategy. The environmental costs of training large models get mentioned but not explored. The document calls for analyzing those costs and developing sustainable targets without offering specifics or benchmarks. And the gender bias discussion, while present, doesn’t go deep enough. The training data perpetuates stereotypical gender roles, the recursive loop of AI learning from biased AI output makes it worse, and the guidance acknowledges both without proposing concrete mechanisms for accountability.
But these are limitations of scope, not failures of vision. UNESCO published this guidance when most governments hadn’t even started thinking about GenAI policy for education. The human-centered stance, the emphasis on equity and cultural diversity, the warning about cognitive erosion, the call for assessment redesign: all of it has aged well. The document doesn’t pretend that GenAI is either a savior or a threat. It treats it as a powerful technology that needs governance, pedagogy, and intentional design to benefit learners. That’s the right starting position, and it’s one I’ve been arguing for on this blog from the beginning.
References
Miao, F., & Holmes, W. (2023). Guidance for generative AI in education and research. UNESCO. https://doi.org/10.54675/EWZM9535
Roe, J., Furze, L., & Perkins, M. (2025). Digital plastic: A metaphorical framework for Critical AI Literacy in the multiliteracies era. Pedagogies: An International Journal. Advance online publication. https://doi.org/10.1080/1554480X.2025.2557491
Shaw, S. D., & Nave, G. (2026). Thinking fast, slow, and artificial: How AI is reshaping human reasoning and the rise of cognitive surrender. Working paper, The Wharton School, University of Pennsylvania. https://papers.ssrn.com/sol3/papers.cfm?abstract_id=6097646
UNESCO. (2024). AI competency framework for students. United Nations Educational, Scientific and Cultural Organization. https://doi.org/10.54675/JKJB9835