What Will Children Thank Us For?

Wrestling with AI’s Role in Education

Convened alongside the CIES Annual Conference, EdTech Hub and the Atlassian Foundation brought together 50 leaders from education, technology, philanthropy, research, and civil society for a breakfast conversation at the Atlassian Foundation San Francisco office.

There were no panels and no presentations; the format was designed to surface honest disagreement and shared priorities through facilitated dialogue. The event was held under the Chatham House Rule. The provocation for the morning was: what decisions about AI and education will children be grateful that we made?

Across two hours, the conversation moved through three broad territories: teachers, students, and the systems that sit behind the classroom. Within each, the group grappled with questions of agency, protection, partnership, and pace. What follows is a synthesis of the ideas that emerged.

For the Atlassian Foundation and the EdTech Hub this conversation connects directly to how we work and where we are headed. We have long believed that the most lasting change in education happens when multiple players work toward shared goals together rather than in parallel. AI in education is precisely the kind of challenge that no single institution can solve alone. The policy, the technology, and the practice are held by different actors, and the stakes for underserved learners are too high to leave coordination to chance. 

We co-hosted this breakfast because we wanted honest disagreement in the room, not a showcase. What we heard reinforced our conviction that the question is not whether AI will reshape education, but whether we will build the shared infrastructure to shape it responsibly. Now we’re sharing the critical themes that need the sector’s consistent attention and collaboration if we’re going to build a future children around the world will truly benefit from. 

Teachers, the Experts Already in the Room

There was broad agreement that teachers must be at the centre of how AI in education is developed, tested, and scaled, but the conversation pushed well past the familiar call to “include teachers”. Teachers should be recognised as the sector’s R&D partners of choice. That clear proposition rang through conversations. Teachers represent the greatest force within education and the largest single investment by ministries. They carry the imagination, professional judgement, and moral responsibility of stewarding the next generation of learners. Without their active leadership in research and development, AI tools will be designed in a vacuum and deployed into a reality they are not equipped for.

The evidence supports this. EdTech Hub’s survey of more than 800 teachers across low- and middle-income countries found that 48% were already using AI, not because anyone told them to, but because they saw value in it. What they most wanted was not generic efficiency tools but targeted support: help delivering difficult STEM content and improving their day-to-day pedagogical practice. When one team co-designed a WhatsApp-based support tool with government teachers in Rwanda, they discovered the real need was just-in-time pronunciation help for English scientific terminology, something no product designer sitting in San Francisco would have guessed. When the AI initially responded in a North American accent, teachers said it was not useful. The tool was re-recorded in a locally accented voice. These are the kinds of essential details that surface when teachers lead.

The cultural risks of designing without teachers are equally instructive. AI tools trained predominantly on Global North data carry assumptions that do not fit every context. In one case, a teacher in rural Kenya using an AI tool for a fractions lesson kept being offered pizza as a teaching aid. Without teachers in the design loop, these mismatches go unnoticed and uncorrected.

But the most provocative contribution pushed the conversation further still. In-service professional development, however well-intentioned, amounts to continuous tinkering at the margins of a profession that needs fundamental renewal. The real frontier is pre-service teacher training; the curricula, the licensure requirements, the institutions that determine who becomes a teacher and how they are prepared. A clear test was proposed: in five years, will the graduating class of teachers have a meaningfully different curriculum from today’s? If not, we will not have changed the parameters of success for the profession. This demands a new coalition working with teacher training college deans in the U.S., Kenya, Uganda, and elsewhere to redesign what preparation for teaching looks like in the age of AI.

There is already some evidence of what this can look like. In Tanzania, a government platform for teacher professional development is being connected with teacher-driven AI use cases, linking what the government has already built with the ways teachers are choosing to use new tools. The best outcomes are realised when AI amplifies existing public resources rather than replacing them. And the pace matters: if the industry is moving fast, the sector must back educators to move fast too. Rapid three-month sandbox models, where small grants enable teacher-led exploration, show that it is possible to be both nimble and rigorous, starting from a belief that teachers know how to navigate, explore, and evaluate, rather than arriving with predetermined answers.

The conversation about teachers kept returning to a harder, less resolved question: what is AI doing to the children they are trying to reach?

Students, the Minds at Stake

The conversation about students was dominated by concern, and it went deeper than questions of access or equity. The group kept returning to what AI is doing to children’s minds, their emotional development, and their capacity to think for themselves.

There is growing evidence that over-reliance on AI is eroding critical thinking. Students with access to AI tools risk becoming passive recipients of streamlined answers, losing the habits of questioning, reasoning, and forming independent views. Their voice and agency are visibly diminishing. At the same time, the question of who decides what AI presents as truth, who sets the content standards, and whose values are embedded in the outputs, remains almost entirely unaddressed. These are not technical problems awaiting a technical fix. They are questions about power, culture, and the purpose of education itself.

The parallel with social media was drawn forcefully in our debates, and could not nor should not be ignored.. Technology’s impact on young people’s mental health and brain development is not hypothetical; social media has already inflicted well-documented harm, and it scaled despite being unsafe. AI’s capacity for personalisation and synthetic intimacy could deepen those harms further, particularly for younger children who cannot distinguish between a human relationship and a convincing simulation of one. The room called for a genuine convergence of expertise across mental health, education, child development, and technology, not to slow things down for the sake of caution, but because protection and progress are not in tension. If a tool is not safe, it should not scale. The uncomfortable truth is that unsafe tools scale all the time.

Underneath these concerns lies a more foundational question about self-regulation. Children’s ability to manage attention, delay gratification, and exercise judgement is under pressure from a technology-saturated environment, and AI adds a new dimension of intensity. Any credible approach to AI in education needs to measure and strengthen these capacities, not just track whether children are using the tools or scoring higher on narrowly defined assessments.

Those concerns do not resolve inside the classroom. Many of them are rooted in how AI tools are procured, governed, and scaled – and that points toward a different conversation, about the systems that sit behind the school.

Systems, the “Boring” Work that Actually Scales

The third territory the group explored was the infrastructure, governance, and partnerships that sit behind the classroom. The conversation here shifted the lens on AI itself: away from classroom-facing tools and toward AI as a means of strengthening the systems that education runs on. The promise of AI is not only that it can help a child learn to read or a teacher plan a lesson, but that it can improve how governments manage data, how programmes are monitored and adapted at scale, how resources are procured and allocated, and how implementation fidelity is maintained across thousands of schools. These backend applications are less visible than a chatbot or a lesson planner, but they may be where AI has its most durable impact on learning outcomes.

For that to happen, governments must be co-architects, not afterthoughts. In multiple countries, government buy-in determines whether an AI programme lives or dies, and governments rightly see themselves as protectors of learners and the teaching profession. Co-architecture means more than consultation. It necessitates building public digital infrastructure that is designed to be handed to the government, with a deep understanding of how institutions are actually structured. In one example, handover required routing content, software, and data to three different parts of a ministry. That work is slow and unglamorous, but it is the only path to sustainability at scale. The alternative, platforms built by external actors and bolted onto government systems, tends to produce dependency rather than capacity. Even the most promising research results fail to translate at scale when implementation fidelity, incentive alignment, and variance in programme execution are not addressed, and those are system-level problems that AI itself can help solve.

Big technology companies were challenged to redirect their focus accordingly. The room had less appetite for product showcases and more interest in the backend conditions that make impact possible: data systems, interoperability, quality standards, and the institutional capacity to govern AI tools over time. Enabling governments to recognise quality in edtech and AI products is a powerful lever, because it compels the supply side to build solutions that are pedagogically sound and contextually appropriate, rather than simply technically impressive. Research must be woven into the development process from the start. When evaluation is tacked on at the end, it becomes a validation exercise rather than a genuine attempt to learn.

The conversation similarly turned a sharp eye on the donor and philanthropic community. Funding incentives too often reward novelty over need. Some AI initiatives are quietly bypassing government systems rather than strengthening them. Coherence and alignment across funders, technology companies, and public institutions matter as much as the tools themselves, particularly when it comes to procurement at scale.

This pointed to a broader recognition that current models of public-private partnership are not fit for this moment. EdTech’s track record was confronted candidly. There’s a shared feeling that previous waves of education technology have largely not delivered transformational change, despite significant investment. The incentive structures of the companies building these tools, the need to move fast, make money, and scale, can be directly at odds with quality teaching, learning, and child development. New partnership models are needed where accountability runs to learning outcomes rather than adoption metrics, where technology companies engage with the institutional realities of education systems rather than building around them, and where teachers, governments, and communities are at the design table from the beginning, not invited in once the product is built. Getting this right may matter more than any individual tool or product.

The absence of government representatives in the room itself was noted as emblematic of a wider gap. If the community is serious about system-level change, it must find ways to make these conversations accessible and relevant to the government actors who ultimately hold the power to adopt, regulate, or block AI in education.

Looking Ahead, What We Know, We Owe!

There was a shared sense in the room that AI in education carries a feeling of inevitability. But even inevitability can be shaped.

The Atlassian Foundation will continue to share what we learn as our own thinking evolves, and to invest in the connective tissue,  convenings, shared evidence, and cross-sector partnerships that makes collective progress possible.

We know more today about how children learn to read and do maths than at any point in history. We have hard-won lessons from previous waves of education technology, including its failures. And we have, across this community, people who care enough to argue honestly about what matters.

For EdTech Hub the convening strengthened our commitment to work on systems level change. We are committed to keep working alongside governments and decision-makers, providing an evidence base they can trust and the practical support needed to implement change at scale. We are also committed to work with those who centre teachers in the next phase of innovation. 

But we left the morning with a new conviction, to grapple with one of the most pressing challenges facing our sector: the intersection of education technology and youth mental health. Features such as gamification, while valuable for engagement, risk seeding addictive behaviours in young people. Leaders want to make the right safeguarding decisions in a rapidly changing environment. From social media regulations to changing screentime guidelines to rollbacks on tech and AI in classrooms, leaders around the world are grappling with complex and evolving factors and need the evidence and partnership to support effective, balanced policymaking. We’ve seen this delicate balance not only in our signal spotting, but also in our deep work with ministries and stakeholders across the sector.  

This issue took on notable urgency during our recent convening, which coincided with a landmark jury verdict finding Meta and Google liable for designing platforms that harmed young users.

The challenge now is to act on what we know. Build on evidence, not hype. Protect children, not just their test scores. Treat teachers as partners, not delivery mechanisms. Strengthen systems rather than bypass them. And hold ourselves to the standard of measuring what actually matters. FOMO is not a strategy. But neither is waiting.