IDinsight

The first Global Dialogue on AI Governance will be a success if it closes the distance between governance principles and the practical realities of AI deployment in development contexts. We believe this requires an outcome of agreement on a minimum evidentiary standard for AI investments. Not whether a tool has been built, but whether it demonstrably improves outcomes for the communities it is meant to serve. The AI Evaluation Playbook is a living framework developed by IDinsight, the Agency Fund, and the Center for Global Development to establish common evaluation standards for AI in the development sector. Its four-level framework spanning model, product, user, and impact evaluation provides a structured methodology that funders, governments, and implementing organisations can adopt and adapt across contexts. Achieving this will depend on two enabling conditions. The Dialogue must surface and legitimise the voices of governments and practitioners deploying AI in low-resource, high-stakes contexts, from community health systems to public education. These actors have hard-won evidence about what responsible AI implementation actually demands. Their experience should shape governance frameworks, not be retrofitted into them after the fact. Equally, the Dialogue must translate that shared standard into concrete commitments: funders formalising evaluation requirements, governments adopting common benchmarks, and implementing organisations treating continuous learning as a core output rather than an afterthought.

Safe, secure and trustworthy AI;Social, economic, ethical, cultural, linguistic and technical implications of AI

Safe, secure and trustworthy AI is foundational to everything we do. When AI is deployed in low-resource, high-stakes contexts, the risk of unintended harm is real. Hence, safety and trust have to be earned through rigorous, continuous evaluation. The AI Evaluation Playbook, co-developed by IDinsight with the Agency Fund and the Center for Global Development, provides a practical framework for doing exactly that: assessing AI systems across four levels, from model performance to real-world impact. Our experience working with governments in countries such as Kenya, Ethiopia, and Senegal has reinforced a simple truth: safety and trustworthiness are not abstract standards. They come down to whether a tool works reliably in low-connectivity environments, whether it respects the agency of the people using it, and whether it actually delivers on its promises. The social, economic, ethical, cultural, linguistic and technical implications of AI are equally central to our approach. IDinsight embeds a dignity lens across all of its AI work, asking not only whether an intervention achieves its technical objectives, but whether it is respectful, inclusive, and culturally appropriate. In contexts where training data is mismatched, design decisions are made far from the communities affected, and people have limited recourse when things go wrong, these questions are not secondary. They are central. Our tools are built to lower language and digital literacy barriers, and our evaluations are designed to surface these dimensions explicitly rather than treat them as a footnote. Governance frameworks that fail to engage seriously with these implications will not serve the communities most affected by AI. We believe these two priorities, taken together, represent the minimum standard for responsible AI deployment in development contexts.

Global AI governance has made progress on questions of safety and security. It has not yet engaged adequately with a more fundamental question: what constitutes credible evidence that an AI system is working as intended, particularly in low-resource, high-stakes development contexts. A community health tool can be secure, transparent, and human rights-compliant on paper and still fail to improve health outcomes in practice. Without a shared methodology for measuring real-world impact, governance frameworks risk becoming compliance exercises rather than genuine accountability mechanisms to the communities they serve. We would encourage the Dialogue to treat evaluation standards as a cross-cutting governance issue in its own right. The AI Evaluation Playbook offers a practical starting point for what sector-wide minimum standards.

On safe, secure and trustworthy AI, Governments and implementing organisations are making significant investments without adequate frameworks to know whether those tools are performing reliably, let alone whether they are improving lives. On the social, economic, ethical, cultural, linguistic and technical implications of AI, the risks are equally tangible. AI systems deployed in low-resource contexts are frequently built on training data that does not reflect local realities, in languages that marginalise significant portions of the population, and with design assumptions shaped by engineers far removed from the communities affected. IDinsight's work in Senegal and Ethiopia on bilingual teacher support and community health tools has shown that cultural and linguistic fit is not a secondary concern. It determines whether a tool is used at all, and whether it strengthens or undermines the judgment and agency of the people it is meant to support. What both gaps share is a sector that has prioritised speed of deployment over depth of accountability. The opportunity before the global governance community is to change that default, by treating evidence generation, cultural fit, and dignity not as enhancements to responsible AI deployment, but as baseline requirements for it.

The AI Dialogue has a real opportunity to do something most bilateral and regional governance efforts cannot: bring together a genuinely diverse set of actors around a shared understanding of what responsible AI deployment requires in practice. Its most valuable contribution would be to shift the conversation toward the contexts where the consequences of getting AI wrong are most severe. Much of the existing international architecture has been shaped by high-income countries, reflecting their infrastructure, their risk profiles, and their priorities. The Dialogue has both the mandate and the platform to change that. Governments, practitioners, and communities in the global South have hard-won evidence about what AI deployment actually looks like on the ground. That experience should be shaping global governance frameworks, not being accommodated by them after the fact. The infrastructure gaps are concrete. AI-powered teaching tools are already in classrooms across Sub-Saharan Africa, yet the tools underpinning them frequently lack access to locally relevant curriculum content, perform poorly in local languages, and generate insights that never travel across borders. IDinsight's work building open-source Knowledge Graphs for Ghana, Senegal, and Zambia's education sector demonstrates what shared, publicly available infrastructure can do and how much of it the governance community has yet to engage with at scale. In practical terms, the Dialogue can help broker agreement on shared evaluation standards, strengthen coordination between existing regional and multilateral initiatives, and establish accountability mechanisms that track real changes in practice rather than commitments on paper. For this to translate into lasting impact, the Dialogue needs to generate learning across sessions, build on regional convenings, and create feedback loops that ground the governance conversation in field experience. The goal should be a global process that becomes more evidence-based, more geographically representative, and more actionable with each iteration.

The Global Digital Compact and the work of the UN Secretary-General's Advisory Body on AI have laid important normative groundwork. The OECD AI Principles and the African Union's Continental AI Strategy represent significant regional efforts to establish shared standards, and the Dialogue should treat these as foundations. At the practitioner level, initiatives like the AI Evidence Alliance for Social Impact, supported by IDRC and FCDO, are generating rigorous field evidence on the effectiveness, risks, and scalability of AI interventions in development contexts. The AI Evaluation Playbook, developed by IDinsight, the Agency Fund, and the Center for Global Development, is building shared evaluation standards that funders and implementers can adopt across contexts. These efforts represent exactly the kind of evidence infrastructure that global governance frameworks need to draw on. What is also missing from current governance conversations is a frank conversation and realisation of what evidence-based AI funding in development should actually look like. Effective AI funding requires a fundamentally different evidentiary standard than traditional programming. Conventional funding models demand proof of effectiveness before committing resources, but AI tools evolve faster than those cycles allow. Rigid requirements will systematically miss the iteration that makes these tools both effective and safe. What the sector needs is a funding posture grounded in continuous evidence generation, one that treats learning and iteration as core outputs rather than preconditions for investment. The Dialogue is uniquely placed to make that case directly to the decision-makers who need to hear it, connecting the actors generating field evidence with those setting funding standards, and ensuring that what is learned in practice shapes what is decided in policy.

Governments, particularly those in low- and middle-income countries, are deploying AI in public systems right now. Their contribution is not theoretical. They can speak to what responsible implementation actually demands, where governance frameworks fall short in practice, and what support they need to make better decisions. Creating structured space for this kind of testimony, beyond prepared statements, should be a design priority. Practitioners and implementing organisations bring field evidence that rarely makes it into policy conversations. The Dialogue should create mechanisms for that evidence to be submitted, synthesised, and formally considered, rather than treated as anecdote. Pre-session written submissions, structured working groups, and dedicated practitioner panels would all help close this gap. Funders have an outsized influence on what gets built and how it gets evaluated. Their participation should be accompanied by an expectation of transparency about their current standards and openness to revising them based on what the Dialogue surfaces. Civil society and community representatives bring accountability. They can speak to the lived experience of AI deployment in ways that technical and policy actors cannot, and their inclusion should be structural rather than symbolic. Equally important is creating pathways for the voices of ordinary citizens to inform the Dialogue directly. Structured citizen data collection, whether through surveys, community consultations, or participatory research conducted ahead of each session, would give the Dialogue an evidence base that reflects how AI is actually experienced by the people most affected by it. This is not a secondary input. In contexts where AI is being deployed in health, education, and social protection, citizen perspectives are primary data, and treating them as such would fundamentally strengthen the legitimacy and relevance of any governance framework that emerges. In terms of format, the Dialogue would benefit from moving beyond plenary debate toward more participatory working sessions designed to generate concrete outputs. The measure of a successful session should not be the quality of the statements delivered, but the specificity of the commitments made and the evidence produced.

The voices most underrepresented in global AI governance discussions are, with few exceptions, the same communities most directly affected by AI deployment in practice. Governments in low- and middle-income countries are often present in formal sessions but underrepresented in the substantive conversations that shape outcomes. They tend to receive governance frameworks rather than co-design them. The Dialogue should actively commission and platform evidence from these governments, not just invite their participation. Civil society organisations working at the community level, particularly those operating in local languages and outside major urban centres, have direct visibility into the social and cultural implications of AI deployment that international organisations and technology companies do not. Dedicated funding for their participation, including translation support and accessible submission formats, would meaningfully broaden the evidence base the Dialogue draws on. Researchers and evidence organisations based in the global South are generating rigorous work on AI deployment in development contexts that is systematically underrepresented in global policy conversations. Creating formal pathways for this research to inform Dialogue proceedings would considerably strengthen the quality of governance decisions. Inclusion is ultimately a design choice. The Dialogue's format, submission processes, and convening structures will determine whose knowledge counts. Getting that design right is as important as any substantive agenda item.

The most effective formats will be those that treat participants as contributors of evidence and experience, not an audience for prepared remarks. Working sessions structured around real cases would shift the Dialogue from abstract principle to practical accountability. Rather than debating what responsible AI deployment should look like, participants would engage with concrete examples of what it actually looks like, where it has worked, where it has failed, and what the governance implications are. This kind of case-based deliberation generates sharper insights and more actionable conclusions than traditional plenary formats. Pre-session evidence submissions, systematically synthesised and presented at the opening of each working group, would ensure that the conversation is grounded in what is actually happening on the ground rather than what participants assume to be true. This would also create a more level playing field between well-resourced delegations and those with less capacity to prepare formal interventions. Structured commitment tracking, in which participants are invited to articulate specific changes they will make based on what they have heard, and in which those commitments are publicly recorded and followed up on at subsequent sessions, would shift the Dialogue's culture from deliberation toward accountability. The gap between what is said in formal sessions and what changes in practice is the Dialogue's greatest risk. Building in mechanisms to close that gap from the outset would set an entirely different tone. Finally, regional preparatory convenings that provide structured input to global sessions would ensure that the Dialogue reflects a genuinely diverse evidence base. IDinsight and its partners are already building this architecture through convenings at the France-Africa Summit, GITEX Africa in Nairobi, and the second session in New York in 2027. Connecting these regional moments to the global Dialogue would make the whole process considerably more than the sum of its parts.

One of the most concrete contributions IDinsight and its partners have made to addressing AI governance challenges in the development sector is the AI Evaluation Playbook, developed together with the Agency Fund and the Center for Global Development. The Playbook establishes a four-level evaluation framework spanning model performance, product engagement, user experience, and real-world impact. It gives funders and implementing organisations a common methodology for assessing whether AI systems are working as intended, for whom, and at what cost. Critically, it is not a theoretical framework. It has been applied across development contexts in health, education, and social protection, generating concrete lessons about what rigorous AI evaluation requires in practice and where current approaches fall short. Those field applications have considerably strengthened the Playbook. Evaluations conducted with government partners and implementing organisations across Africa and South Asia have tested the framework against real deployment challenges, including low-connectivity environments, linguistically diverse user populations, and high-stakes public systems where the consequences of failure are severe. The result is a methodology that reflects the realities of AI deployment in low-resource contexts, not just the assumptions of those designing tools from a distance. The Playbook was formally launched at the Skoll World Forum in April 2026, ahead of the Global Dialogue on AI Governance in Geneva. It is designed as a living framework, updated based on ongoing field evidence and structured input from the practitioner community. We see it as a practical starting point for the kind of sector-wide minimum standard that the Dialogue could help legitimise and scale. The governance gap in AI evaluation is real and consequential. A tested, field-grounded framework that funders and governments can adopt is part of what it means to close that gap.