Wadhwani Center for Government Digital Transformation

The first Global Dialogue would be most successful if it does three things that no existing AI governance forum currently does well. First, it should function as a structured learning exchange grounded in implementation evidence, not only in normative declarations. The most consequential AI deployments today are unfolding in public sector contexts across the Global South, often at scale, under resource constraints, and serving populations with limited literacy and connectivity. Governance norms developed without this evidence base will not survive contact with deployment reality. Second, it should produce actionable guidance for high-stakes public sector AI, not only for frontier model risks. Most of the world's population will encounter AI not as a chatbot but as the system mediating their access to welfare, healthcare, or grievance redress. The current debate underweights this reality. Third, it should establish a permanent feedback mechanism between national implementation experience and global norm development. This means treating the Dialogue not as a single biennial event but as a continuous process where evidence from deployments flows into evolving frameworks, and revised frameworks flow back to inform deployment practice. Success would also mean visible representation of practitioners and affected communities, not only ministerial delegations and policy advocates. The principle of leaving no one behind, foundational to the 2030 Agenda, must apply to participation in AI governance itself, not only to its outcomes. A Dialogue that produces a polished communiqué of agreed principles without changing how implementation experience reaches global frameworks would be a missed opportunity.

• Safe, secure and trustworthy AI
• AI capacity-building
• Social, economic, ethical, cultural, linguistic and technical implications of AI
• Transparency, accountability, and human oversight

These four priorities reflect where the gap between current global governance frameworks and public sector AI deployment reality is widest, and where engagement now will have the greatest downstream impact. Safe, secure and trustworthy AI is foundational, but the existing literature is heavily oriented toward frontier model risks and high-resource regulatory contexts. The more urgent governance question for most of the world's population is whether the public sector AI system mediating their access to welfare, healthcare, or grievance redress is reliable, contestable, and trustworthy in the operational sense. This dimension of trustworthiness is materially underdeveloped in current frameworks. AI capacity-building is the precondition for any governance framework to function. Norms, principles, and accountability requirements only operate if the institutions deploying AI have the capability to interpret and apply them. In most public sector contexts, particularly in the Global South, institutional capability is the binding constraint. Without sustained capacity investment, governance frameworks remain aspirational. The social, economic, ethical, cultural, linguistic and technical implications of AI is the priority where the inclusion question lives. The dominant governance debate implicitly assumes a text-literate, smartphone-connected user. This assumption excludes a substantial share of the global population. Frameworks that do not address linguistic diversity, low literacy, and voice-first interaction will not meet the principle of leaving no one behind. Transparency, accountability and human oversight is where global frameworks are weakest relative to the pace of public sector deployment. Where AI mediates access to entitlements or services, the standards for auditability, redress, and meaningful human review must be more demanding, not less. Current frameworks have been developed largely with private sector and consumer-facing AI in mind, and do not yet address the specific accountability requirements of state-deployed AI in low-recourse contexts.

The most material gap in the listed themes is the absence of explicit attention to voice-first and low-literacy AI governance. The global governance debate implicitly assumes a text-literate, smartphone-connected user. This assumption excludes a substantial portion of the world's population. In India alone, functional illiteracy affects hundreds of millions of people. Across the Global South, a large share of potential AI system users cannot read a consent notice, navigate a grievance portal, or interpret a text-based AI output. Voice-first AI systems, which interact with users through spoken language in regional dialects, represent the only viable path to inclusive AI for these populations. They raise distinct governance questions that the current thematic framework does not yet address. How is meaningful informed consent obtained from a user who receives an automated phone call? What does an audit trail look like for a voice interaction? What does human oversight mean when the user has limited literacy and no access to a written redress mechanism? Frugal AI is a related blind spot. Systems designed for low-end devices, intermittent connectivity, and feature phones operate under constraints that the current safety, accountability, and transparency literature does not adequately reflect. A third cross-cutting issue is lifecycle governance. Most current discussion focuses on AI development and pre-deployment risk. The harder challenge is what enables a deployed AI system to remain reliable, accountable, and contestable over time as data, users, and institutional context drift. The Dialogue should treat post-deployment governance as a first-order theme.

India is deploying AI in welfare, agriculture, healthcare, and public administration at a scale few countries can match. The governance gaps we encounter are not theoretical. The most immediate gap is the absence of accountability frameworks suited to high-volume automated public sector AI. When a voice system like ASHA places calls each month to millions of PDS beneficiaries, the question of what happens when the system errs is not addressed by any existing governance instrument. There is no agreed standard for what constitutes a false negative in a welfare context, no redress pathway designed for a beneficiary without a smartphone or internet access, and no audit requirement that reflects the operational reality of these deployments. The second gap is linguistic and cognitive. Most AI safety and transparency frameworks assume text-based interfaces and literate users. India's linguistic diversity, with over 400 languages and hundreds of millions of citizens with limited literacy in any official script, makes frameworks designed for English or even Hindi interfaces structurally exclusionary at the outset. The third gap is the absence of guidance on lifecycle governance. India has many AI proofs of concept that fail to transition into sustained operation. The reasons are governance reasons such as unclear ownership, weak post-deployment monitoring, the absence of mechanisms for handling context drift, and limited horizontal learning across government institutions. The opportunity is significant. India has accumulated operational experience deploying AI for structurally excluded populations. The IndiaAI Mission, the Digital Public Infrastructure stack, and iGOT Karmayogi represent governance infrastructure being built in real time. Global processes have not yet created a mechanism to learn from these deployments systematically. The Dialogue could change that.

The Dialogue's most valuable role would be to function as a structured learning exchange, not only as a forum for normative agreement. Most international AI governance processes are organised around norm-setting, agreeing on principles, declarations, and eventually instruments. This work is necessary but insufficient. The more pressing need is a mechanism through which countries, particularly in the Global South, can share implementation experience in ways that inform how those norms evolve. Concretely, the Dialogue could do three things that no existing mechanism does well today. First, it could maintain a living repository of implementation case studies from diverse deployment contexts. Welfare AI in South Asia, agricultural advisory systems in sub-Saharan Africa, judicial decision support in Latin America, and similar work elsewhere. These are precisely the environments where governance gaps are most consequential. Second, it could facilitate South-South knowledge transfer directly, without routing it through Geneva or New York. Countries deploying AI in resource-constrained, high-diversity contexts have more to learn from each other than from regulatory models developed in high-income jurisdictions. Third, it could establish a feedback loop from national implementation back into global norm development. At present, norms are developed at the global level and expected to flow downward into national practice. The Dialogue could reverse this dynamic, allowing the evidence base from national deployments to revise governance frameworks over time. The Dialogue will fall short if it becomes another forum for member states and stakeholders to reiterate agreed principles. Its distinct value lies in being the place where implementation evidence shapes what those principles actually mean in practice.

The Dialogue should connect with existing initiatives rather than duplicate them. Within the UN system, the Global Digital Compact and the work of the Independent International Scientific Panel on AI provide directly relevant foundations. The UNESCO Recommendation on the Ethics of AI already represents a multilateral consensus on principles. The Dialogue should treat these as starting points and focus on the implementation and evidence layers that they do not currently address. The Global Partnership on AI has produced substantive output on responsible AI, where India has been an active participant. Multi-stakeholder bodies such as the Partnership on AI have generated technical work on transparency and accountability. Connecting practitioners with these technical bodies through the Dialogue's process would add value that neither currently provides on its own. National and regional initiatives also offer transferable lessons. India's Digital Public Infrastructure approach treats foundational systems for identity, payments, consent, and data exchange as a public good with embedded governance. Countries in other regions, including parts of Africa and Latin America, are building comparable infrastructure. The Dialogue should create channels for these initiatives to inform each other directly. Sectoral programmes within the UN system, particularly those linked to the Sustainable Development Goals on health, education, and poverty, generate field evidence on AI in service delivery contexts. This evidence currently sits outside the AI governance conversation. The Dialogue should pull it in. The added value the Dialogue can uniquely bring is convening authority. It is the only UN-mandated platform that can bring political, technical, civil society, and practitioner constituencies into the same room. That power should be used to integrate existing knowledge, not to generate new declarations that sit alongside rather than connect with what already exists.

The Dialogue risks replicating the participation patterns of existing forums unless it designs deliberately against them. Civil society and practitioners from the Global South should be involved in agenda-setting, not only invited as observers or panellists. The thematic framework should be tested with practitioners before each session convenes, and their input should visibly shape the agenda rather than be noted in summary documents. Private sector participation should distinguish between industry associations and individual companies, and between frontier AI developers and firms deploying AI in resource-constrained public sector contexts. Their interests are not identical, and conflating them distorts the conversation. The format should include working sessions organised around specific implementation challenges rather than thematic panels. A session on how to design redress mechanisms for low-literacy users of welfare AI would generate more actionable output than a general panel on transparency and accountability. A structured summary of inputs from each consultation should be published, including areas of disagreement, not only points of consensus. Trust in the process depends on participants seeing that their inputs were considered, including where they were contested. The 2026 Geneva session should not be the primary participation opportunity. Regional pre-consultations with structured input mechanisms, similar to the consultation hosted at the India AI Impact Summit, should be a formal part of the process, with outputs that feed traceably into the Geneva agenda. The Dialogue should also adopt an open consultation cycle between sessions, where the working draft of the framework is published and revised in response to structured input, rather than treating the biennial sessions as discrete events.

The most systematically underrepresented voices in global AI governance are the people most directly affected by AI in public sector contexts. Welfare beneficiaries, rural citizens, persons with limited literacy, and users who interact with government AI through voice rather than text interfaces. This is not a gap that can be addressed only by inviting more civil society organisations from the Global South to Geneva. Most civil society organisations active in global governance forums are themselves urban, English-speaking, and primarily focused on policy advocacy rather than service delivery. They represent affected communities. They are not those communities. Practical inclusion requires investment in intermediary mechanisms. Practitioners who work directly with affected populations, frontline government officers, community health workers, and local NGOs with implementation experience, need to be resourced to participate. This means funding travel, providing interpretation, and designing sessions that do not require fluency in Geneva negotiating culture to contribute meaningfully. It also requires rethinking what counts as evidence. A documented case from a frontline AI interaction in a welfare context, where an automated verification has gone wrong and the household has no recourse, is governance-relevant evidence. The Dialogue should create mechanisms to collect, verify, and present this kind of structured practitioner testimony alongside the research papers and policy documents that currently dominate governance deliberations. Governments of smaller and lower-income countries are also underrepresented relative to their share of AI deployments and affected populations. The Dialogue should consider dedicated capacity support for these governments to prepare substantive inputs, including drafting support and access to subject matter expertise, not only logistical support to attend sessions.

Four formats would materially improve the quality of engagement over standard panel discussions. The first is structured problem clinics. Rather than thematic panels at which participants deliver prepared statements, the Dialogue could organise sessions around specific governance challenges submitted in advance by practitioners. Participants would work through a real problem, for example how to design a redress mechanism for a voice-based welfare AI system serving low-literacy users. The session output would be a documented set of approaches rather than a summary of positions. The second is a practitioner testimony track. A dedicated segment in which frontline implementers, the people who actually operate AI systems in public service delivery, describe what they encounter, what works, and what fails. This is distinct from case-study presentations and would require structured facilitation to draw out governance-relevant insights rather than operational descriptions alone. The third is an iterative input mechanism between sessions. Rather than treating the 2026 Geneva session as the primary moment of engagement, the Dialogue could publish a working draft of its emerging framework after each consultation and invite structured responses. This would allow participants to see how their inputs have or have not shaped the document and respond accordingly. The fourth is a structured failure review. A session in which governments and organisations present AI deployments that did not work, what failed, why, and what governance implications follow. The current discourse rewards success stories and discourages honest reporting of failure, which is precisely where the most governance-relevant learning sits. Together, these formats shift the dynamic from performance, where participants deliver positions, to deliberation, where positions evolve in response to evidence and argument.

Three examples from India's public sector AI ecosystem are directly relevant to the Dialogue's mandate. The first is ASHA, an automated voice-based grievance verification system developed by WGDT. It places calls in regional languages to PDS beneficiaries, reaching millions of households each month to verify whether grievances have been resolved. By reaching populations that cannot navigate a written redress portal, ASHA shows that voice-first AI can extend democratic accountability to citizens otherwise structurally excluded by smartphone-dependent or text-dependent interfaces. The governance lesson is that inclusive design is itself an accountability mechanism, not only an accessibility feature. The second is iGOT Karmayogi, the Government of India's online learning platform for civil servants. WGDT has supported over four million course completions on the platform across 30+ courses on AI and digital governance. It demonstrates that sustained AI capability-building can be embedded within national government infrastructure rather than delivered through one-off project funding. The governance implication is that capability investment requires a permanent institutional home within the state. The third is the CAG AI Champions cohort, in which audit officers identified, designed, and presented AI projects with WGDT support. Nine projects were approved and five were showcased at Audit Diwas in November 2025. The lesson was that structured peer cohorts, anchored by an empowered single point of contact within the institution and an external public accountability deadline, change adoption behaviour more effectively than training alone. Across the three examples, the recurring pattern is that effective AI governance in the public sector requires four elements working together. A technical system, sustained human capacity investment, an empowered institutional champion, and an external accountability mechanism. No element is sufficient on its own.