Indian Council Of Medical Research (ICMR)

The first Global Dialogue on AI Governance would be successful if it delivers following outcomes. First, it should establish the Dialogue as a credible, inclusive, and development-oriented platform for sustained international cooperation on AI governance. The value of this first meeting will lie not only in the exchange of perspectives, but in creating confidence that the United Nations can serve as a space where all countries, particularly developing countries, can participate meaningfully in shaping the global conversation. Second, it should identify a practical agenda for cooperation to bridge AI divides. This should include capacity-building, access to digital and computational infrastructure, support for multilingual and context-sensitive AI systems, and strengthening of national capabilities for evaluation, governance, and responsible deployment. For many countries, inclusive AI governance will depend on whether global discussions translate into real support for domestic readiness and institutional capacity. Third, the Dialogue should help advance a shared understanding that trust in AI requires governance across the full lifecycle of AI systems. Governance should not be limited to design-stage principles alone, but should also include transparency, accountability, human oversight, interoperability, and mechanisms to address evolving risks after deployment. If the first Dialogue can lay the foundation for continued cooperation that is inclusive, practical, and responsive to differing national contexts, it would represent a meaningful and forward-looking success.

• Transparency, accountability, and human oversight
• Safe, secure and trustworthy AI
• AI capacity-building
• Interoperability of governance approaches

Priorities are guided by the need to ensure that AI governance remains both inclusive and operational. Safe, secure and trustworthy AI is a priority because public confidence in AI will depend on whether systems are reliable, resilient, and fit for deployment in high-impact sectors. As AI moves into health, public services, and other socially significant domains, governance must address not only performance, but also safety and institutional trust. AI capacity-building is essential because equitable participation in the AI ecosystem requires more than access to applications. It requires infrastructure, skills, datasets, evaluation capability, and governance capacity. For developing countries, this is central to bridging AI divides and ensuring that AI supports national development priorities. Interoperability of governance approaches is important because AI is developing across jurisdictions, sectors, and institutional settings at great speed. Greater interoperability can promote coherence, reduce fragmentation, support mutual learning, and facilitate practical cooperation while respecting national circumstances and policy space. Transparency, accountability, and human oversight remain indispensable because trust in AI cannot rest on voluntary claims alone. It requires clear responsibility, meaningful review mechanisms, and oversight that remains effective throughout the operational life of AI systems.

Few cross-cutting issues merit more explicit attention. First, diversity, equity, and data sovereignty should be reflected more directly within global AI governance discussions. Many countries, particularly in the Global South, remain underrepresented in datasets, benchmarks, and model evaluation ecosystems. Governance should therefore address not only access to AI systems, but also fair participation in shaping datasets, standards, and evaluation norms. Second, there is a need to recognize full lifecycle and post-deployment risks more explicitly. Emerging challenges such as model drift, adversarial manipulation, post-training compromise, and vulnerabilities across complex AI supply chains may not be adequately addressed through static or one-time compliance approaches. These issues are increasingly relevant in high-impact sectors and deserve greater policy attention. Third, context-specific evaluation is an important emerging concern. AI systems developed and tested in one setting may not perform equitably or safely in linguistically, culturally, or institutionally different environments. This has implications for fairness, trust, and real-world effectiveness, especially in developing country contexts. Finally, greater attention may be needed for public-interest digital infrastructure for AI governance, including shared testing environments, multilingual resources, benchmark repositories, and institutional mechanisms for audit and incident reporting.

In India, and more broadly across developing-country contexts, the selected thematic areas are already shaping both the opportunities and the governance pressures of AI deployment. On the opportunity side, rapid progress in digital public infrastructure and digital health platforms has shown that AI-enabled systems can expand access, improve continuity of care, strengthen real-time monitoring, and support service delivery in underserved settings. Interoperable, consent-based health architectures and AI-assisted public platforms also create a pathway for more scalable and inclusive deployment. At the same time, the most significant challenge is that governance maturity is not advancing at the same pace as deployment. In practice, this creates four linked risks. First, capacity gaps persist in infrastructure, skilled workforce, multilingual usability, and last-mile connectivity, which can widen exclusion if AI systems are not designed for rural and marginalized populations. Second, insufficient interoperability across public and private systems reduces efficiency and limits the full benefit of digital ecosystems. Third, accountability remains blurred when AI influences decision-making, particularly in relation to liability, transparency, and meaningful human oversight. Fourth, current governance approaches remain stronger on data protection and ex ante compliance than on post-deployment risks such as model drift, adversarial manipulation, and silent system compromise. For India and the wider region, this also creates a major strategic opportunity. Governance gaps can be addressed not only through regulation, but through institution-building: locally governed and diverse datasets, auditable evaluation frameworks, lifecycle monitoring, and interoperable standards. Initiatives such as MIDAS show how countries can move from data dependence to data sovereignty while improving trust, representativeness, and accountability. If supported through international cooperation, these advances could enable a more equitable and trustworthy AI ecosystem for the Global South.

The AI Dialogue can play a valuable convening and bridging role in advancing international cooperation on AI governance. Its principal added value lies in providing an inclusive United Nations platform where Member States and relevant stakeholders can exchange perspectives, identify practical areas of convergence, and build greater coherence across a rapidly expanding and fragmented governance landscape. The draft structure itself usefully envisages the Dialogue as a space to connect related initiatives, draw out complementarities, and support continued cooperation beyond a single meeting. The Dialogue can also help ensure that international cooperation remains development-oriented. For many countries, especially in the Global South, AI governance is not only about frontier-model safety; it is also about capacity-building, access to infrastructure, interoperability, context-specific evaluation, and equitable participation in shaping standards and datasets. In this respect, the Dialogue can help bring questions of diversity, equity, and data sovereignty into the centre of global discussions. A further role for the Dialogue is to elevate lifecycle governance as a shared international priority. Emerging risks such as model drift, adversarial manipulation, and post-deployment compromise require cooperation that goes beyond static compliance. The Dialogue can support practical exchange on monitoring, accountability, incident reporting, and human oversight, while encouraging mutual learning across sectors and jurisdictions.

The AI Dialogue should build upon and connect with existing multilateral, technical, and sectoral efforts rather than duplicate them. These include the broader UN-linked processes surrounding the AI for Good ecosystem and the Independent International Scientific Panel, as well as WHO-led efforts on digital health interoperability and certification. In health and related digital sectors, useful reference points also include the European Health Data Space, HL7 FHIR-based interoperability efforts, and virtual capacity-building mechanisms such as Project ECHO. The Dialogue may also benefit from connecting with governance and evaluation frameworks that are already shaping responsible AI practice. These include CONSORT-AI, SPIRIT-AI, TRIPOD-AI, and DECIDE-AI for evaluation and reporting; as well as data-governance efforts such as the OECD Health Data Governance Principles and the African Health Data Space. These mechanisms provide important building blocks on transparency, evidence standards, legal interoperability, and ethical safeguards. From India's perspective, initiatives such as MIDAS and AIKosh also illustrate how countries can operationalize data quality, local stewardship, and auditable evaluation in practice, including through South-South collaboration models. The added value of the AI Dialogue would be to connect these otherwise separate efforts within a legitimate, inclusive, intergovernmental setting. It can help identify complementarities, reduce fragmentation, elevate developing-country priorities, and translate dispersed standards and initiatives into a more coherent agenda for international cooperation on AI governance.

Different stakeholders can contribute most effectively if the Dialogue is structured in a way that combines intergovernmental legitimacy with genuine multistakeholder exchange. Governments should articulate national priorities, regulatory experience, and areas where international cooperation is needed. International organizations can help synthesize global evidence, identify convergence areas, and support implementation pathways. Industry and the technical community can contribute practical lessons on safety, interoperability, standards, and deployment challenges. Academia and research institutions can provide independent evidence, evaluation frameworks, and foresight on emerging risks. Civil society, especially from developing-country contexts, can bring perspectives on rights, equity, inclusion, and real-world impacts in communities and public services. With regard to format, the Dialogue would benefit from a balanced structure combining a high-level governmental segment, a multistakeholder plenary, and smaller thematic discussions organized around clearly framed questions. Breakout sessions should be designed for substantive exchange rather than sequential statements, and could be co-chaired by a Member State and a stakeholder representative. Short scene-setting interventions, moderated discussion, written submissions, and concise rapporteur summaries could improve focus and continuity. Hybrid participation, multilingual access, and advance circulation of guiding questions would also support broader inclusion. A concise Chair's summary identifying practical areas for continued cooperation would add value and help carry momentum beyond the first meeting.

Global discussions on AI governance still do not adequately reflect the perspectives of developing countries, particularly those facing constraints in infrastructure, data ecosystems, compute access, and regulatory capacity. There is also insufficient representation of public-sector implementers, sectoral regulators, public health practitioners, frontline workers, multilingual communities, women, youth, persons with disabilities, and populations whose data are often used but whose interests are not meaningfully represented in governance design. An important gap also concerns institutions and experts from the Global South who are working on context-specific deployment, dataset stewardship, evaluation, and public-interest digital infrastructure. Without these voices, global debates risk becoming overly shaped by frontier-model concerns in high-resource settings, while questions of equity, diversity, language, access, and data sovereignty remain peripheral. These groups could be included through practical measures: dedicated speaking opportunities for underrepresented regions and communities; support for remote and hybrid participation; multilingual documentation and interpretation; open calls for written submissions; regional and sectoral consultations in advance of the Dialogue; and stronger engagement with public-interest researchers, civil society organizations, and field-level practitioners. It would also be useful to ensure that thematic discussions include participants from real-world implementation settings, not only from policy or technology centres. Inclusion should be treated not as symbolic balance, but as necessary for building a governance framework that is globally credible, equitable, and responsive to diverse contexts.

Meaningful engagement is likely to be strengthened by formats that move beyond prepared statements and allow participants to examine concrete governance questions together. One useful format would be moderated thematic roundtables built around short case studies or policy scenarios, for example on post-deployment accountability, multilingual inclusion, or cross-border interoperability. This would allow participants to respond to practical dilemmas rather than speak only in general terms. A second useful approach would be breakout "policy labs" co-led by a Member State and a stakeholder representative, with limited initial interventions followed by guided discussion and a short list of takeaways. This could make thematic sessions more interactive and solution-oriented. Short "lightning interventions" from diverse participants could also be used to surface emerging issues quickly before deeper discussion. The Dialogue could further benefit from structured written inputs, digital consultation tools, and real-time synthesis by rapporteurs. Interactive polling or moderated digital questions may help identify areas of convergence across participants. A dedicated segment for showcasing practical cooperation models, including capacity-building partnerships, interoperable governance tools, and public-interest digital infrastructure initiatives, could also make the Dialogue more action-oriented. Overall, innovative formats should aim to combine inclusiveness with discipline: fewer long statements, more guided interaction, clearer thematic framing, and stronger follow-through in the final summary. This would help ensure that the Dialogue remains substantive, dynamic, and useful for future cooperation.

A first example is the EU AI Act, which offers a useful risk-based and lifecycle-oriented framework, particularly through requirements on risk management, technical documentation, and post-market monitoring for high-risk systems. Although further strengthening is still needed, especially for routine outcome-linked audits, it remains an important model for structured governance. A second relevant example is the FDA's Predetermined Change Control Plan (PCCP), which provides a practical mechanism for governing pre-specified modifications in AI-enabled medical products, showing how oversight can adapt to systems that evolve over time. A third example is India's Digital Personal Data Protection framework, which is significant in demonstrating a citizen-centric approach to lawful processing, consent, security safeguards, and institutional accountability. While data protection alone is not sufficient for full AI governance, it is an essential foundation for trustworthy deployment. At the level of implementation, MIDAS (Medical Imaging Datasets for India) offers a concrete solution to challenges of representativeness, data quality, and local stewardship. Its dataset-grading approach, built around documentation, technical fidelity, governance, and representativeness, provides a practical model for moving from data dependence toward data sovereignty while improving auditability and trust. Related platforms such as AIKosh illustrate how nationally governed repositories can support responsible data sharing and benchmarking. In addition, large-scale public digital health platforms such as eSanjeevani, especially when linked with interoperable digital health architecture, show how governance can be embedded into public digital infrastructure rather than treated only as ex post regulation. Finally, some of the most concrete governance practices are operational rather than legislative: continuous post-deployment validation, adversarial testing, periodic audits, AI safety boards, sentinel case libraries, and incident reporting mechanisms. These are especially important for high-impact sectors where governance must remain effective after deployment, not only before it.