Centre for Responsible AI (CeRAI), IIT Madras

A successful first 'Global Dialogue on AI Governance' must move beyond declaratory principles and deliver outcomes that are implementable, inclusive, and grounded in real-world impact, particularly for communities in the Global South who stand to benefit most yet remain most underserved. 1. Inclusion as a fundamental principle: Governance for the 'Common Person' lens would be a commitment to ensuring AI systems serve marginalised and resource-constrained communities—not just advanced markets. Identifying and mitigating risks faced by end-users, such as smallholder farmers or frontline health workers, must be embedded into governance design from the outset. 2. Lifecycle-Based Governance as a Foundational Principle: AI governance without lifecycle thinking is incomplete. Risk identification and assessment, accountability mechanisms, and oversight must span design, development, deployment, and post-deployment phases—not remain isolated interventions at any single stage. 3. Accountability across all AI actors: The dialogue must define the differentiated yet collaborative responsibilities of developers, fine-tuners, deployers, intermediaries, and end-users. Each actor's role requires separate consideration and coordinated accountability across the AI lifecycle. 4. Technology-Agnostic, Context-Sensitive Frameworks: Governance frameworks must apply across all AI types, such as discriminative, prescriptive, generative, and agentic, while also remaining flexible to risk severity, domain sensitivity, and the decision-making stakes of the deployment context. 5. Operationalising Responsible AI as Features: The dialogue should prioritise actionability by advancing transparency, safety, security, and understandability beyond privacy. Each of these ought to be viewed as a responsible AI "feature" that AI systems seek to incorporate. Actionability must include mandating AI audits, both product and process audits, alongside structured incident reporting mechanisms for all AI systems. 6. Thinking Beyond Foundation Models: The Global South often requires lightweight, traditional AI systems suited to low-infrastructure environments. Governance frameworks must be resource-conscious and applicable to simple algorithmic tools that solve critical, localised problems. 7. Human-in-the-Loop Feedback Mechanisms: Success requires institutionalising structured pathways through which real-world user experiences continuously inform AI systems across their lifecycle, ensuring AI adoption genuinely serves humanity rather than merely advancing it technically.

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

1. Safe, Secure, and Trustworthy AI: Safety and security must be embedded as design principles from the earliest phases of AI development, not treated as post-deployment add-ons. This requires human-centric safety evaluations that reflect Global South contexts, alongside mandatory AI incident reporting and structured safety disclosures. Advancing AI audits, both tool-based and process-based, covering internal mechanisms (such as ISO 42001) and independent third-party audits (which also include ISO 42001), is a concrete, actionable outcome the Dialogue can deliver to operationalise trustworthiness globally. 2. AI Capacity Building: The dialogue offers a unique opportunity to establish common AI and AI governance terminology accessible across nations and user communities with varying technical maturity. Capacity building must go beyond training developers; it must empower deployers, regulators, civil society, and end-users to meaningfully participate in AI governance processes. 3. Transparency, Accountability, and Oversight: Transparency requires balancing material disclosure with intellectual property protections, a tension requiring global alignment. Accountability must adopt a "law-plus" framing: combining binding legal obligations with voluntary frameworks to accommodate diverse national regulatory approaches while ensuring risks are identified and mitigated across the AI lifecycle, harms are addressed, and incidents are reported both proactively and responsively when adverse consequences occur. Oversight must clarify where human judgement is essential, distinguishing "human-in-the-loop" from "human-on-the-loop" across different deployment contexts. 4. Social, Economic, Ethical, Cultural, Linguistic, and Technical Implications: Western-centric AI models and governance frameworks inadequately serve the Global South's diverse socio-cultural, economic, and linguistic realities. Deployments in high-stakes sectors like health, agriculture, and justice are undermining fundamental rights and eroding user trust. Evaluations and impact assessments must systematically account for these dimensions to ensure AI genuinely serves all of humanity.

1. Lifecycle-Embedded Inclusivity and Participatory Governance: A critical cross-cutting issue is the integration of social, economic, ethical, cultural, and linguistic implications within lifecycle-based accountability frameworks. Framing accountability as a proactive, actor-dependent, lifecycle-focused, risk mitigation mechanism in a "law-plus" manner, incorporating voluntary standards/frameworks, naturally embeds participatory approaches at every stage. India's experience is instructive, as we follow a 'law-plus' approach for managing AI risks. 2. Social, cultural, linguistic, and ethical contexts are diverse and vary significantly across countries and communities. So far, there is no universal checklist that can be looked up for such evaluations. The Dialogue must ponder conversations in those lines and enable community-centric assessment frameworks to evolve organically across Global South regions, ensuring locally significant risks are not overlooked. 3. Operationalising Human Rights Beyond the EU Frame: While human rights protection is a globally acknowledged principle, its operationalisation remains largely EU-centric, that is, tied to specific legal instruments not practically replicable elsewhere. The Dialogue must deliberately avoid imposing an EU regulatory framework on all nations. 4. Interoperability Requires Conceptual Clarity: A foundational distinction must be established before interoperability can be meaningfully discussed: AI governance is not synonymous with AI regulation. Conflating the two risks stalling dialogue by implying that all nations must have formal regulatory frameworks in place before cooperation is possible. Recognising governance as "law-plus", combining legal obligations with voluntary standards, codes of practice, frameworks and institutional norms, allows diverse national approaches to align without demanding regulatory uniformity. What ultimately enables interoperability is not harmonised regulation but a shared commitment to using AI for the public good through responsible and ethical innovation. Rigid regulatory expectations risk stifling the innovation and adoption that underserved communities need most; a governance-first approach, grounded in shared principles, is both more inclusive and more practically achievable on a global scale.

1. Safe, Secure, and Trustworthy AI: The Diversity Challenge and Opportunity for India: India's linguistic, cultural, and literacy diversity makes AI safety and trustworthiness uniquely challenging to implement. Existing safety benchmarks and evaluation datasets are predominantly designed for high-resource languages, resulting in weaker protections and more frequent model failures in local language contexts, disproportionately affecting women, children, and marginalised communities who lack effective redress mechanisms. The opportunity lies in leveraging India's technical, legal, and governance talent to build safety frameworks that are genuinely contextual, multilingual, and scalable. 2. Transparency, Accountability, and Oversight: Conceptual and Institutional Gaps: Core governance terms remain widely misunderstood. Transparency is frequently reduced to complete model disclosure rather than structured, audit-based disclosure meaningful to users. Accountability is narrowly associated with post-harm legal liability rather than proactive, lifecycle-based risk mitigation. India's AI Governance Guidelines offer a constructive step through a graded liability regime and value chain transparency. Pre-deployment impact assessments are rare, post-deployment monitoring is largely absent, and regulatory oversight mechanisms are underdeveloped, particularly for public services affecting underrepresented groups. 3. AI Capacity Building: Scaling Relevance Across a Diverse Nation: Defining what AI capacity meaningfully entails for different individuals, like farmers, frontline workers, regulators, and citizens, remains unresolved. Generic training programmes fail to address the contextual needs of a country as diverse as India. The opportunity lies in designing differentiated, community-relevant capacity building that spans technical literacy, governance comprehension (law plus and lifecycle focused), and civic awareness at scale. 4. Localised AI Safety Infrastructure as an Emerging Priority: There is growing recognition of the need for localised AI safety infrastructure: multilingual evaluation datasets, independent auditing capacity, and structured post-deployment monitoring. Civil society organisations and public-interest researchers are increasingly filling institutional gaps by grounding evaluations in real-world use and lived experiences, a model the Global Dialogue should recognise and support. Risk of technological lock‑in and dependency on a small set of western models: An important concern for the Global South is the risk of technological lock‑in and dependency on a small set of Western AI developers, especially if governments and institutions rush to deploy frontier AI models without building local capacity, evaluation infrastructure, and governance safeguards. This can entrench asymmetric power over critical digital infrastructure, weaken policy autonomy, and crowd out investment in context‑appropriate, locally governed AI systems.

1. Flexible, Inclusive AI Governance Frameworks: AI Governance is not limited to regulation; it is "law-plus," combining legal obligations with voluntary standards, frameworks, institutional norms, ethical values, and codes of practice. Governance frameworks must be flexible enough to reflect each country's strengths and strategic priorities without imposing a single regulatory model globally. Countries and organisations participate in the AI ecosystem in different capacities as model developers, fine-tuners, deployers, intermediaries, or end-users, and governance must treat each actor's role as equally valid and important. 2. Governing Across the AI Stack: The AI stack comprises infrastructure, data, models, and applications, and governance obligations apply meaningfully at each level. Respective entities should have the flexibility to govern the layer of the stack that is most strategically relevant to them, rather than being expected to regulate the full stack uniformly. 3. Consensus on Responsible AI "Features": The Global Dialogue must identify and reach consensus on core "features" of Responsible AI, such as safety, transparency, fairness, accountability, and understandability, that should be globally valued and implemented. Critically, the Dialogue must go beyond identifying features to illustrating how each can be practically operationalised, with concrete examples across different deployment contexts. A "features-first" approach delivers actionable, implementation-ready guidance rather than abstract principles.

1. Global South Network for Trustworthy AI: The Dialogue should actively engage with the Global South Network for Trustworthy AI, a civil society-led platform uniting organisations across Asia, Africa, and Latin America to strengthen AI evaluation and governance in diverse, real-world deployment contexts. Its grounded, evidence-based work on multilingual model performance, deployment risks, and documented harms fills critical gaps that mainstream safety and evaluation frameworks overlook. The Dialogue can amplify and institutionalise this work by connecting it to global standard-setting processes. 2. Participatory and Multistakeholder Approaches: The Centre for Responsible AI (CeRAI), IIT Madras, alongside the Vidhi Centre for Legal Policy, has pioneered a participatory approach to AI governance that embeds community engagement at the conception stage of the AI lifecycle, not merely as a consultative afterthought. This framework recognises that meaningful governance requires structured pathways for affected communities, civil society, and domain experts to shape AI systems before deployment. 3. Building a Bottom-Up Governance Agenda: The added value the Dialogue can uniquely bring is the creation of sustained infrastructure for equitable knowledge exchange, enabling a genuinely bottom-up articulation of AI priorities and risks. By connecting community-level insights to global policy processes, the Dialogue can ensure that governance frameworks are grounded in lived realities rather than driven exclusively by high-resource, technology-producing nations.

1. Civil society and grassroots organisations, who are often first to document harms from AI in public services, remain structurally under-represented, despite UNESCO and others stressing the need to equip CSOs from marginalised and Indigenous communities to participate in AI governance. 2. Local and Indigenous AI researchers working on multilingual and low-resource contexts, such as AI4Bharat, Masakhane and Deep Learning Indaba, also lack consistent representation at global governance tables, even though they generate critical evidence on bias, safety and usability for the global majority. 3. Other missing voices include Indigenous peoples, linguistic minorities, youth, and persons with disabilities, who are rarely empowered as critical players rather than consultees. They can be included by holding regional sessions in Global South countries; providing funded participation, visa and accessibility support; live‑streaming deliberations with interpretation and safe channels for remote interventions; and creating a formal global advisory group composed of CSOs, Indigenous, youth, disability and Global South technical networks with co‑decision rights over agendas and outcomes.

1. India AI Governance Guidelines: India's AI Governance Guidelines offer a concrete example of governance that goes beyond regulation, integrating existing laws with voluntary frameworks and international standards. The guidelines recommend utilising current legal frameworks to their fullest extent, spanning: - Digital Personal Data Protection Act, 2023, - Information Technology Law and their recent amendments related to AI-generated content, - Intellectual Property Laws , - Consumer Protection Law Additionally, we have sector specific guidelines and strategies that are developed in alignment with the National AI Governance Guidelines: - RBI's Framework for Responsible and Ethical Enablement of Artificial Intelligence (FREE-AI) for the Finance sector - NHA's Strategy for AI in Healthcare in India 2. AI Accountability and Value Chain: India's AI Governance Guidelines explicitly recognise the AI lifecycle and the differentiated roles of actors within it, recommending the following: - A graded liability regime that assigns accountability proportionally across developers, deployers, and users - Transparency in the AI value chain is a governance requirement, not merely a principle. 3. Risk Mitigation as a Distinct Governance Pillar: Risk mitigation is treated separately from legal accountability, acknowledging that responsible AI implementation requires measures broader than law alone. The other dimensions applicable for robust AI governance include: - ISO 42001 - international standard for AI management systems - Bias Assessment Standards published by the Department of Telecommunications, Government of India - Industry Playbook (NASSCOM Developer's Playbook for Responsible AI) for Implementing Responsible AI, a domestic voluntary framework for industry adoption - Voluntary commitments and AI incident reporting frameworks by industry actors and research communities as an adaptive governance mechanism - Third-party independent audits as a trust-building instrument, advocated as a practical accountability measure Why This Model Matters Globally India's approach demonstrates that effective AI governance does not require a dedicated AI regulation. By leveraging existing legal infrastructure, establishing clear actor-based accountability, and embedding voluntary mechanisms as structured complements to law, the Guidelines offer a replicable, flexible model particularly relevant for Global South nations building governance capacity without extensive regulatory infrastructure.