Artificial intelligence has reached an awkward stage in education. It is no longer sufficiently new for adoption itself to count as progress, yet it is spreading faster than schools, universities and governments can agree on how success should be measured.
Teachers are using generative systems to prepare lessons. Students can translate, summarize and interrogate material in seconds. Universities are experimenting with tutors, assessment tools and personalized learning environments. Governments are considering AI as part of national education strategies.
The capabilities are real. So are the risks of confusing capability with educational value.
A major global report published this year by UNICEF and EdTech Hub arrives at precisely this moment. Drawing on international evidence and country experiences, it warns against another cycle in which education embraces technology first and asks whether it improved learning later. Its central premise can be expressed in a remarkably simple question: what problem are we actually trying to solve?
That question should now sit at the centre of responsible AI in education.
For the next stage of educational technology, the meaningful measures are no longer the number of devices distributed, licences purchased or AI features added to a platform. They are whether students learn more, teachers gain useful capacity, inequality narrows rather than widens, and institutions remain capable of governing the systems they introduce.
Technology arrived. The learning crisis remained.
The timing matters because AI is entering education during a period of profound educational inequality. UNICEF reports that 273 million children are out of school, while large numbers of children who do attend school still fail to acquire foundational literacy. Conflict, poverty, disability, climate disruption and chronic underfunding continue to interrupt learning around the world.
AI did not create those problems, and there is little reason to believe that simply deploying more technology will solve them.
The history of EdTech should make policymakers cautious. Technologies that perform impressively in demonstrations can struggle when they encounter unreliable connectivity, limited teacher preparation, language barriers, weak institutional capacity or costs that become unsustainable after a pilot ends.
The UNICEF–EdTech Hub work makes this point forcefully. Digital and AI tools can accelerate learning and widen access, but poorly designed adoption can also waste public resources, deepen vendor dependence and widen existing inequalities.
That is an important distinction. The argument is not against technology. It is an argument for taking technology seriously enough to demand evidence from it.
The learner, not the machine, should define the problem
For much of the current AI boom, the conversation has begun with what the technology can do. Education requires the opposite sequence.
A learner may struggle because the available material is not in a language they understand well. A teacher may have access to thousands of resources but no efficient way to locate what is relevant. A student with a disability may encounter content that is technically online but practically inaccessible. Another learner may be working, caring for family members or living with inconsistent connectivity.
Each situation is an educational problem before it is a technological one.
Only after the problem is understood does AI become useful: Can translation reduce a language barrier? Can searchable video help a learner locate a specific concept inside an hour-long lecture? Can intelligent indexing make a university’s existing knowledge more useful? Can flexible delivery help someone continue learning when conventional schedules do not fit their circumstances?
This is also why the distinction between access and usable access matters.
Putting education online removes one barrier. It does not remove all of them.
At ICARUS, much of our work has been built around this gap. UNESCO’s SDG 4 Knowledge Hub has documented the ICARUS practice of using AI-enabled multilingual search, transcription, translation and content indexing to make educational resources easier to discover, navigate and reuse.
The technology matters because of the friction it can remove. The educational value lies in what the learner can do afterward.
Teachers are not the obstacle to scale
AI also forces education to reconsider what it means by efficiency.
It is tempting to look at teaching as a collection of tasks that can progressively be automated. Some tasks certainly can be. Educators spend enormous amounts of time searching for material, preparing resources, reorganizing content and performing repetitive administrative work. AI can assist with many of them.
But education is not simply the transfer of information from a source to a recipient.
Teachers interpret. They notice confusion that is not expressed in a prompt. They understand context. They motivate students who have lost confidence. They challenge assumptions and help learners distinguish understanding from memorization.
The UNICEF–EdTech Hub report places teachers at the centre of successful digital transformation and argues that they should participate in the design and selection of the technologies they are expected to use. UNESCO’s AI Competency Framework for Teachers takes a similar position, combining technical competence with ethics, pedagogy and a human-centred approach.
This suggests a more useful way of thinking about productivity in education.
The relevant question is not how many teachers AI might replace. It is how much educational capacity responsible AI can return to teachers.
A good system should allow an educator to spend less time navigating information and more time exercising the judgment for which educators are actually needed.
Personalization can help learning — or quietly isolate it
Personalization is one of AI’s most attractive educational promises. Learners do not progress at identical speeds, begin with identical knowledge or need identical explanations. Technology capable of responding to those differences can be genuinely valuable.
Yet there is an unresolved tension.
An education system can become more personalized while becoming less human.
A learner guided continuously by an algorithm may receive material calibrated perfectly to an individual profile and still lose something important: intellectual friction with other people, disagreement, collaboration, mentorship and the experience of discovering that another person sees the same problem differently.
Education has always been partly social.
The UNICEF–EdTech Hub evidence warns against models in which personalization produces isolation, particularly when lower-cost automated provision becomes the default for learners who already have fewer educational opportunities while wealthier students retain access to strong human support.
That possibility deserves much more attention.
AI should make high-quality education more available, not create a two-tier future in which some students receive teachers, mentorship and technology while others receive primarily technology.
AI access is not the same as AI agency
There is another inequality emerging that will be harder to measure than access to devices.
Two students can use the same AI system and derive very different value from it.
One may know how to challenge a response, verify evidence, recognize uncertainty, compare competing explanations and understand when specialist knowledge is required. The other may accept an answer because it is fluent and immediate.
Both have AI access.
Only one has meaningful AI agency.
At the United Nations Global Dialogue on AI Governance in Geneva this year, our ICARUS side event focused on precisely this distinction. We argued that education must become part of AI governance because societies cannot govern powerful systems successfully when the people using them lack the capability to question them.
This changes what AI literacy should mean.
Knowing which button to press is not literacy. Neither is knowing how to generate a polished document quickly.
AI literacy increasingly involves judgment: when to trust, when to verify, when to disagree and when not to use the system at all.
That capability will matter in universities, workplaces and democratic institutions long after today’s specific models have disappeared.
Safety and public trust cannot be added later
Education also gives AI access to unusually sensitive environments.
Students generate information about academic performance, behaviour, interests, difficulties and, sometimes, highly personal circumstances. Children deserve an especially high level of protection.
UNICEF and EdTech Hub therefore argue for privacy, safeguarding and child-rights protections before technologies enter schools, rather than after problems emerge. They also warn about commercial dependence when governments adopt platforms without fully understanding long-term costs, data practices or the consequences of becoming locked into a particular provider.
This is where responsible AI becomes institutional rather than merely technical.
A system can perform extremely well and still be inappropriate for education if its incentives are wrong, its data practices cannot be explained, or the institution using it is unable to challenge the decisions it makes.
Trust depends on institutions knowing what they are deploying.
It also requires the courage to stop deployments that do not produce sufficient educational value.
Evidence should come before scale
AI companies are rewarded for speed, while education systems should be more careful.
Experimentation is necessary, and waiting for perfect certainty would also be a mistake. Schools and universities need opportunities to test new approaches, learn from failure and identify where AI creates genuine value.
But pilots and system-wide implementation are different decisions.
A tool can work brilliantly with a few hundred students, committed teachers and intensive technical support, then perform very differently when introduced across thousands of classrooms.
That is why evidence, cost and institutional readiness should determine scale.
The UNICEF–EdTech Hub report calls for policymakers to ask what works, for whom and under what conditions. It also emphasizes sustainability, interoperability and the full costs of adoption rather than only the initial price of software or equipment.
This is not bureaucracy obstructing innovation.
It is what responsible innovation looks like when the consequences affect millions of learners.
The next phase of AI in education
At ICARUS, the direction this points toward is increasingly clear.
We believe AI should make knowledge easier to reach across languages, help educators use their time more effectively, connect learning more clearly with competence and give learners greater capacity to navigate information independently.
But those ambitions would only matter if they can be demonstrated in final outcomes.
The next phase of AI in education should therefore be less impressed by the presence of AI itself.
We should ask whether learners understand more. Whether educators can teach better, people previously excluded from knowledge can now participate, students become more independent thinkers rather than more dependent users. Whether institutions can govern the systems they adopt and sustain them without surrendering control.
Google, UNICEF, UNESCO, universities, governments and technology companies will use different language for this transition. The underlying principle is simpler.
Technology adoption is an input. Education is the outcome.
Twenty-five years of digital education have already shown how dangerous it can be to confuse the two. Generative AI gives us an opportunity to avoid repeating that mistake at a far greater scale.
The most important question about AI in education is therefore no longer what the technology can do.
It is what the learner can do because the technology was there.
Sources:
- UNICEF & EdTech Hub — Shaping the Future of Digital Education – UNICEF / EdTech Hub global report
- UNICEF – Digital Education – UNICEF Digital Education
- UNICEF – Education Strategy 2026: Shaping the Future of Learning – UNICEF Education Strategy
- UNICEF – EdTech for Good Framework – UNICEF EdTech for Good
- UNESCO – AI Competency Framework for Teachers – UNESCO AI Competency Framework for Teachers
- UNESCO SDG 4 Knowledge Hub – ICARUS AI-enabled Multilingual Learning Infrastructure – UNESCO ICARUS good practice
- ICARUS – AI for Every Mind: Education, Human Agency & Governance – ICARUS UN Global Dialogue report