Kokilaben Hospital, Mumbai, India Founder – BeResponsibleAI

The first Global Dialogue on AI Governance will be successful only if it moves beyond principles and produces operational clarity. In my view, five outcomes are critical: (1) A move from principles to measurable governance: We already have sufficient global principles (OECD, UNESCO, etc.). What is missing is a way to measure, compare, and enforce responsible AI across systems and jurisdictions. The dialogue should initiate a framework that enables this transition. (2) Interoperability across regulatory ecosystems: The fragmentation between EU, US, India, and other regulatory approaches is growing. The Dialogue should enable a translation layer that allows systems to align across jurisdictions without duplicating compliance effort. (3) Real multi-stakeholder participation grounded in practice. Not just representation, but decision-relevant input from clinicians, engineers, and operators who deploy AI in high-risk environments. In healthcare, governance failures translate directly into patient harm. (4) A capacity-building roadmap that is implementable, not aspirational Capacity-building must include tools, frameworks, and governance infrastructure, not just training or policy guidance. Without this, responsible AI remains theoretical for most of the world. (5) Recognition that high-risk domains require different governance models Healthcare, finance, and critical infrastructure cannot rely on generic AI governance. These domains require deterministic safety boundaries, traceability, and human accountability as enforceable conditions—not recommendations.

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

My work sits at the intersection of clinical practice and AI governance implementation, and these priorities reflect what actually matters in deployment-not just theory. Safe, Secure and Trustworthy AI: In healthcare, safety is not a principle-it is a constraint. AI systems that influence diagnosis or treatment must demonstrate predictable behavior, bounded outputs, and reproducible logic. Purely probabilistic systems are insufficient in isolation for high-risk decision-making. Transparency, Accountability, and Human Oversight: Every AI-assisted clinical decision must be traceable and auditable, with clear ownership. Accountability cannot be delegated to a model. The final decision-and responsibility-remains with the clinician. Social, Ethical, and Cultural Implications: Healthcare AI operates across diverse populations. Bias, language accessibility, and cultural context are not secondary concerns-they directly affect outcomes. These must be measured and monitored, not just acknowledged. AI Capacity-Building: There is a structural gap globally: very few professionals understand both domain context (e.g., medicine) and AI governance. Without building this dual capability, systems will either be blindly adopted or unnecessarily rejected. Both are dangerous.

Two issues are not being adequately addressed in current global discussions: (1) Deterministic vs. Probabilistic AI Governance: Most global governance discussions are centered around generative and probabilistic AI. However, in high-risk domains like healthcare, deterministic and rule-based systems remain critical because they allow for traceability, validation, and reproducibility. These are fundamentally different paradigms and require different governance approaches. Treating them under a single framework creates blind spots. (2) Consent Architecture and Data Sovereignty in AI Systems: Current data protection frameworks (GDPR, DPDP, HIPAA) do not adequately address AI training and secondary data use at scale. We need: 1. Portable consent mechanisms 2. Auditability of data usage across AI pipelines 3. Clear patient-level control over how data contributes to AI systems Without this, we risk eroding trust in healthcare AI before it matures.

From a healthcare perspective in India and the Asia-Pacific region: Challenges: (1) Regulatory ambiguity for AI systems: India has a data protection framework, but AI-specific governance—especially for clinical use—is undefined. This creates uncertainty for developers and clinicians. (2) Lack of sector-specific standards: Unlike the EU, there is no clear classification of healthcare AI as high-risk with corresponding obligations. This results in uneven adoption and unclear liability. (3) Capability gap at institutional level Hospitals and medical institutions are not equipped to evaluate AI systems. This leads to: Blind trust or complete rejection. Neither is acceptable. Opportunities: (1) Ability to build governance natively: India has the opportunity to embed governance into digital health infrastructure from the start, rather than retrofitting it later. (2) Scalable models for low-resource settings: If governance works in India, it can work in most of the world. (3) Market-driven demand for trust: Clinicians and patients are already asking: "Can I trust this system?" This creates natural pressure for responsible AI adoption.

The AI dialogue can be valuable if it becomes more than a discussion platform. (1) Create convergence without forcing uniformity: Different regions will regulate differently. The dialogue should enable alignment without homogenization. (2) Establish minimum operational expectations for high-risk AI: Especially in domains like healthcare, we need baseline enforceable conditions, not just ethical guidelines. (3) Enable knowledge transfer with implementation focus: Sharing frameworks, failures, and real-world deployments is more useful than sharing principles. (4) Support domain-specific governance tracks: Healthcare AI governance cannot be solved as a subset of general AI governance. It needs dedicated technical and clinical input. (5) Introduce continuous evaluation mechanisms: Governance must be iterative and lifecycle-based, not event-based.

Several strong foundations already exist: 1. OECD AI Principles – strong at the level of principles 2. UNESCO AI Ethics Recommendation – global normative alignment 3. GPAI – multi-stakeholder collaboration 4. WHO AI for Health Guidance – domain-specific clarity 5. EU AI Act – regulatory structure for high-risk systems However, all of these share a common limitation: They are stronger at defining what should be done than how to implement it in real systems. The added value of the AI Dialogue would be to bridge this gap—from principle to implementation.

The Dialogue should be designed for function, not representation 1. Governments → regulatory experience 2. Academia → evidence and rigor 3. Private sector → implementation reality 4. Civil society → accountability & Clinicians and domain experts → decision impact Healthcare professionals, in particular, must be included in dedicated high-risk AI sessions, not as peripheral voices. A useful structure would include: 1. Pre-dialogue regional consultations 2. Core dialogue with sector-specific working groups 3. Post-dialogue tracking of commitments Without continuity, the dialogue will not translate into impact.

The most underrepresented groups are those closest to impact: • Healthcare professionals in low-resource settings: Doctors, nurses, and hospital administrators in Global South are on the frontlines of AI deployment but rarely have a seat at governance tables. Their practical insights about resource constraints, infrastructure limitations, and patient diversity are invaluable. Inclusion mechanism: Dedicated sessions with virtual participation options, travel fellowships for selected practitioners, and partnerships with medical associations in developing nations. • Patients and patient advocacy groups: The ultimate stakeholders, patients whose data trains AI systems and whose lives are affected by AI decisions, are almost entirely absent from governance discussions. Inclusion mechanism: Structured patient testimony sessions, partnerships with patient advocacy organizations, and mechanisms for patient groups to submit formal position statements • SMEs from emerging economies • Non-Western ethics and social science perspectives: While technically trained AI researchers are well-represented, scholars from ethics, sociology, anthropology, and philosophy, especially from non-Western intellectual traditions, need stronger representation. Inclusion mechanism: Dedicated ethics and humanities panels, cross-disciplinary working groups, and recognition of diverse philosophical frameworks beyond Western ethics. • Indigenous communities These groups should not just be included symbolically, they should have structured input pathways that influence outcomes.

Several innovative engagement formats could make the AI dialogue more dynamic and impactful. The goal should be not just discussion but shared problem-solving: (1) Live AI Governance Simulation Exercises: Participants engage in real-time scenario-based exercises where they must collaboratively govern a hypothetical AI system facing emerging challenges. This hands-on approach reveals practical tensions between principles and implementation, fostering deeper understanding across stakeholder groups. (2) AI-Enabled Multilingual Participation: Deploy responsible AI tools for real-time translation and transcription across all sessions, ensuring language is never a barrier. This should extend beyond major UN languages to include widely spoken regional languages, with AI-powered summarization to make long proceedings accessible to all. (3) Global Town Halls with Distributed Hubs: Instead of concentrating the Dialogue in Geneva alone, establish simultaneous satellite events in regional hubs (Africa, Asia, Latin America) that feed into the main session via live video. Local discussions are synthesized and presented to the plenary, ensuring geographic diversity is built into the format itself. (4) Public Deliberation Assemblies: Randomly selected citizens from diverse backgrounds participate in structured deliberation sessions on specific AI governance questions, with expert testimony and facilitated discussion. Their conclusions inform the official dialogue outcomes, bringing grassroots perspectives directly into high-level decision-making. (5) Innovation Sandboxes: Provide physical and virtual spaces where stakeholders can prototype governance tools, demonstration frameworks, and compliance mechanisms in real time. This transforms the Dialogue from a discussion event into a co-creation laboratory, with tangible outputs that participants can take back to their organizations. (6) Consensus-Building Algorithms: Use collaborative digital platforms that visualize areas of agreement and disagreement in real time across all participants, helping negotiators identify pathways to consensus and ensuring no voice is lost in the noise.

Several models are useful: • EU AI Act → risk-based classification • Singapore AI Verify → practical testing toolkit • Canada AIA → structured impact assessment • WHO AI for Health → domain-specific governance • Model Cards / Datasheets → transparency practices From my own work in healthcare AI: We have found that governance becomes meaningful only when it is embedded into the system itself-through: • Taceability • Validation pipelines • Safety constraints • Audit mechanisms Otherwise, it remains documentation.