On the first day of the OECD Digital Education Outlook 2026 conference on “Exploring Effective Uses of Generative AI in Education,” we saw what happens when education system stakeholders ask not whether AI can improve performance, but whether it helps human beings learn and grow.
The focus was K‑12, but the implications reach far beyond schools, into every place where human beings need to learn together in order to act, including the humanitarian and health worker networks that I work with at The Geneva Learning Foundation.
The sessions did not offer comfort.
They documented performance gains alongside learning losses, gains in efficiency with flat or declining human competence, and the emergence of what Dragan Gašević called “metacognitive laziness,” a reduction in the productive struggle that genuine learning requires.
At the same time, the OECD Digital Education Outlook conference offered glimpses of a different future, in which AI is used to strengthen human capability, amplifying the social, situated, reflective learning that we already see at scale in peer networks of health workers and humanitarian practitioners.
What the first day’s sessions at OECD Digital Education Outlook 2026 taught us about AI and learning
Several threads ran through the eight sessions of the first day.
Pedagogy matters, otherwise AI may improve task performance but undermines learning
First, the evidence is now clear that generative AI can dramatically improve task performance while undermining learning if used without careful pedagogical intent.
Andreas Schleicher illustrated this with studies that should make every education leader pause: in one United States study, students who used large language models wrote better essays, but 80% could not remember what they had written about afterwards.
In a Turkish study, students using ChatGPT for mathematics did better on exercises, yet performed worse when their mathematical thinking skills were assessed.
The lesson is blunt.
Doing something with AI does not mean learning from it.
Metacognitive laziness
Second, Dragan Gašević’s keynote took us inside the dynamics that explain these findings.
When learners delegate too much of the cognitive work to AI, they think less about their own thinking.
He described how generative AI reduces the “productive friction” that is essential to learning and coined a phrase that captures the risk: metacognitive laziness.
His meta-analyses show that when we look only at performance metrics, AI interventions appear to have large positive effects.
When we isolate genuine learning outcomes, the effects are small and sometimes negative, especially in languages other than English and Chinese.
In other words, AI can make us look smarter while making us less so.
The line between general-purpose tools and educational AI
Third, several sessions drew a sharp line between general-purpose tools and educational AI.
Educational-purpose tools, designed around learning science and pedagogy, perform better and more consistently than generic chatbots.
Ryan Baker’s GPTA system is one example.
It uses retrieval-augmented generation (known as “RAG”) to ground feedback in course syllabi, rubrics, and instructor comments, then delivers rubric-aligned feedback on essays and projects within minutes.
Students revise immediately and substantively.
In studies across five universities and sixteen courses, only about 20% of feedback issues reappeared in subsequent assignments or courses, suggesting that students actually internalize and reuse what they learn, rather than gaming the system for a better score.
Here, AI is helping students learn, not only perform.
How we currently benchmark AI
Fourth, the day developed a powerful critique of how we currently benchmark AI.
Mutlu Cukurova showed that almost all technical benchmarking asks a narrow question: can the model complete a task as well as, or better than, humans.
His teacher‑AI teaming framework proposes a different benchmark.
The real question is whether the use of AI increases human capability over time.
He identified three paradigms for teacher‑AI interaction: replacement, complementarity, and synergy.
Today, most teacher use of AI sits in the replacement paradigm, automating tasks like lesson planning, which yields productivity gains but risks flattening or eroding teacher competence over time.
The more ambitious goal is synergy, where teacher and AI together achieve more than either could alone, and both become more capable through the interaction.
Governance choices matter
Fifth, the country sessions made it clear that governance choices matter.
The United Kingdom has moved from broad “expectations” to detailed product safety standards for generative AI in education, including requirements around stated purpose, cognitive development, emotional and social development, mental health, and protection against manipulation.
These standards were not written in a back room.
They incorporated deliberative dialogues with over 1,000 students, giving young people a direct voice in defining what safe and beneficial AI looks like in their classrooms.
Estonia is running a national AI leap that treats language as a strategic concern.
When AI tools do not speak good Estonian, students immediately switch to English, with real consequences for culture and identity.
France has deliberately refused to have an AI strategy, choosing instead to embed AI within a broader digital education strategy so that technology does not drive educational goals.
Each of these choices reflects an underlying philosophy about who education is for and what kind of future we are building.
AI as an amplifier
Finally, the main OECD insights session surfaced the ethical core of the debate.
Schleicher framed AI as an amplifier.
It can widen opportunity, for example for students with disabilities, or entrench inequality.
It can empower teachers as designers of learning environments, or reduce them to executors of scripted lesson plans.
David Edwards of Education International asked a question that still echoes: will we give poor children chatbots while rich children get AI‑enhanced human tutors with PhDs?
Young participants like Khadija Taufiq reminded us that we are placing a heavy burden on learners, expecting them to police their own cognitive offloading and long‑term interests in environments optimised for short‑term convenience.
If there is one thing that the first day established, it is that we cannot outsource educational ethics to platform design or market forces.
Beyond K‑12: What this means for open learning in humanitarian and health systems
Although the OECD event focused largely on schools, much of what we heard speaks directly to the work that thousands of humanitarian and health workers are doing in local communities, often far from formal classrooms.
For the last decade, The Geneva Learning Foundation has been building networked peer learning systems in which frontline practitioners teach and support each other at scale, across more than 130 countries.
In these networks, health workers co‑create knowledge, adapt global guidance to local realities, and hold each other accountable for implementation.
When we began to experiment with AI in 2019, and later with agentic AI that acts as an active co‑worker, our concern was not primarily how to automate content delivery.
It was how to protect and expand what makes these networks powerful: local agency, collective intelligence, and the ability to learn from lived experience in complex, uncertain environments.
From that vantage point, three implications of the OECD discussions stand out.
1. The performance-learning gap is even more dangerous in crisis settings
In K‑12, a widening gap between performance and learning is troubling because it undermines young people’s capacity for lifelong learning.
In humanitarian response and health work, that same gap can cost lives.
If a health worker uses a general-purpose chatbot to draft a vaccination micro‑plan or an outbreak response protocol, the output can look impressive, coherent, and authoritative.
The OECD evidence warns us that this may say more about the model’s fluency than about the worker’s understanding.
Gašević’s findings on metacognitive laziness suggest a real risk that practitioners who lean heavily on AI for technical tasks will spend less time reflecting on their own reasoning and less time integrating new knowledge with their hard‑won local experience.
In my own work, I have seen how quickly AI can replace certain knowledge‑worker tasks in education and global health, and how tempting it is to accept the surface gains while ignoring the erosion of human capability underneath.
In crisis contexts, that erosion is not an abstraction.
It directly affects the ability of local teams to anticipate and response, to detect when guidance is wrong for their context, and to challenge flawed assumptions coming from distant experts.
The take-away from these OECD sessions is clear: if we use AI to optimise for speed and performance alone, we may strip away exactly the reflective, situated learning that keeps communities safe.
2. Human–AI teaming must start from human networks, not tools
Cukurova’s framework for teacher–AI teaming translates directly into the world of health and humanitarian work, where human-led peer learning networks coordinated by TGLF already function as large‑scale learning systems.
In these networks, human peers review each other’s plans, share stories of failure and adaptation, and build what I have called a “scholarship of practice”. When we introduce AI into these systems, the question is not only what tasks it can automate, but what kind of human–AI teaming we are enabling.
Most of the early AI experiments in global health and humanitarian training sit firmly in the replacement paradigm.
They automate translation, summarisation, or even the facilitation of online courses.
As in schools, this can yield real productivity gains.
It allows a smaller core team to reach more learners with fewer resources, and I have openly described how, in one programme, we replaced key functions initially performed by humans and refocused a smaller team on work that we could not or did not want to automate.
However, the OECD sessions make it painfully obvious that if we stop there, we risk freezing or degrading the collective capabilities of the network.
The complementarity and synergy paradigms that Cukurova outlined offer a better direction.
In a global immunisation network, for example, an AI agent should not be the one designing local strategies.
Instead, it should help health workers surface patterns across thousands of peer‑generated plans, highlight outliers, or suggest overlooked evidence, while leaving final judgment with the practitioners who know their communities.
In such a configuration, human–AI teaming can increase the capability of both the network and the individual practitioners, rather than hollowing them out.
3. Equity, access, and language gaps are structural, not peripheral
My own writing about AI has stressed that equity is not an add‑on question for AI in health and humanitarian work.
It is the main question.
The OECD discussions about geographies and languages brought this into sharper focus.
Gašević’s meta‑analysis shows that studies in English and Chinese tend to find small positive effects of GenAI on learning, while studies in other languages show considerably negative effects.
Estonia’s experience with students switching to English when models fail to handle Estonian is a vivid reminder that language is not just a technical parameter.
It is culture, identity, and power.
In many of the countries where we work, access to advanced models is limited by geofencing, pricing, infrastructure constraints, or risk aversion by international organisations worried about compliance.
In these contexts someone who uses AI is often more likely to be punished than rewarded, even if the outcomes are better and the costs are lower.
The OECD’s insistence that black‑box AI has no place in education applies with equal force to AI that will influence health decisions, emergency response, or the allocation of scarce resources.
Transparency in models and governance is not a luxury.
It is a condition for legitimate use.
There is another form of inequity that was less explicit in the OECD discussions but is central in humanitarian work.
AI systems are being deployed and marketed in ways that privilege data‑rich environments, high‑resource languages, and actors with legal and technical capacity to negotiate contracts and standards.
The result is a growing disconnect between where AI innovation happens and where the most severe global health and humanitarian challenges are located.
The Swiss PanoramAI conference that I spoke at last year had limited engagement with Africa, Asia, and Latin America, even though those regions are likely to be the future markets for AI and are certainly where the highest stakes are.
If we take seriously the OECD’s emphasis on listening to students when setting K‑12 AI standards, then by analogy we must take equally seriously the voices of local health workers and community responders when shaping AI for global health.
4. A different path: Networked peer learning with AI as co‑worker
The first day of the OECD conference repeatedly returned to the idea that AI is an amplifier that can make good practice better and poor practice worse.
In my own work, I have seen a glimpse of what it looks like when AI amplifies a very different kind of practice from what we see in traditional K‑12 systems.
In programmes like Teach to Reach, thousands of health and humanitarian professionals from around the world develop context‑specific projects, exchange structured peer feedback, and participate in dialogues that help them reflect on their own practice.
The architecture is deliberately designed to foreground peer interaction, local adaptation, and metacognition.
AI enters not as an all‑knowing tutor, but as a co‑worker that helps with tasks that peers have neither time nor bandwidth to perform at scale: matching peers with complementary experience, surfacing emerging patterns across thousands of projects, or highlighting exemplary practices that might otherwise remain invisible.
The guiding principle is simple.
AI should augment human connection, not replace it.
The OECD sessions on GPTA and other educational AI tools show that when we design with this principle in mind, AI can indeed help learners internalize feedback, strengthen their own capabilities, and transfer learning across contexts.
When we strip away pedagogy and treat AI as a general‑purpose shortcut, we get the worst of both worlds: more polished outputs with shallower understanding.
For humanitarian and health systems that urgently need both speed and depth, this is not a trade‑off we can afford.
A leadership agenda for learning in the age of AI
What, then, should education and system leaders take from the first day of the OECD conference, if we look beyond the boundaries of K‑12?
First, leaders need to treat the performance–learning distinction as non‑negotiable. In schools, universities, public health institutes, and humanitarian agencies, any AI deployment that boasts efficiency gains without a clear strategy to protect and deepen human capability should be treated as a risk, not a success.
Second, leaders should resist the temptation to adopt AI tools that come without a pedagogy or a theory of learning. The most promising work showcased at the OECD event, whether Baker’s GPTA or the Dutch work on transforming general‑purpose AI into educational AI, starts from a clear understanding of how learning happens and designs the technology around that.In humanitarian and health systems, we already have emerging models of networked peer learning that are producing measurable results at scale. AI should be designed to fit those models, not the other way around.
Third, equity and language must be part of AI strategy from the outset, not retrofitted as a set of mitigation measures. Estonia’s insistence that AI tools speak Estonian as well as English is not only about national pride. It is about ensuring that AI does not further marginalise those who already face educational and economic barriers.In global health, that same logic requires us to confront geofencing, pricing models, and risk frameworks that lock out the very practitioners who most need support.
Fourth, leaders should make room for the kind of uncomfortable honesty that the OECD event and my own experience both point to. It is no longer credible to tell professionals that AI is not coming for their jobs, when in our own organisations we have already replaced certain knowledge‑worker functions with AI and will likely continue to do so. What is left for humans to do is not a rhetorical flourish. It is a design question about the future division of labour between humans and machines. In my view, the most irreplaceable human work lies in facilitating learning, building trust, holding space for difficult conversations, and making value‑laden decisions in the face of uncertainty. AI can augment this work, but it cannot substitute for it.
Finally, we need leaders who are willing to look at education not only as the production of individual human capital, but as the cultivation of collective intelligence in networks that can adapt to complex, rapidly changing challenges.
The OECD conference provided evidence that when AI is used uncritically, it can undermine exactly the metacognition, collaboration, and critical thinking that such networks require.
The peer learning architectures emerging in global health show that a different outcome is possible.
In those systems, AI is not a teacher.
It is a co‑worker in a community of practice that knows how to learn.
The first day of the OECD conference was a reminder that learning in the Age of artificial intelligence is not a technical problem to be solved by better models.
It is a political and ethical choice about what kind of humans we want to become and what kind of systems we want to build together.