Theirworld

A successful first Global Dialogue on AI Governance should be judged not by the breadth of participation, but by whether it reduces ambiguity in how AI is governed globally. First, it should shift from abstract principles to decision-making clarity. Concepts like fairness and accountability are widely endorsed, but rarely operationalised. Success would mean outlining how governments should classify risk, assign responsibility across the AI lifecycle, and intervene when harms occur.Second, the dialogue should produce early-stage interoperability between governance models. Complete alignment is unrealistic, but agreement on minimum standards for high-risk systems would meaningfully reduce regulatory fragmentation and prevent governance arbitrage.Third, it must correct the current imbalance in participation by enabling agenda-setting power for the Global South, not just representation. This includes recognising differences in regulatory capacity, data ecosystems, and development priorities.Fourth, the dialogue should generate institutional follow-through, such as pilot collaborations, shared testing frameworks, or coordinated regulatory experiments. Without this, global dialogue risks becoming performative. Finally, success requires grounding discussions in existing, measurable harms—including bias, exclusion, and technology-facilitated violence rather than focusing solely on speculative future risks.Ultimately, the dialogue should move AI governance from principles to procedures, and from coordination to implementation.

• Safe, secure and trustworthy AI
• Interoperability of governance approaches
• Protection and promotion of human rights
• Transparency, accountability, and human oversight

These priorities reflect a view that the core challenge in AI governance is not identifying principles, but making them enforceable across uneven global contexts. Safe, secure and trustworthy AI is often framed as a technical objective, but in practice it is a governance question: who defines acceptable risk, and who bears responsibility when systems fail? Interoperability of governance approaches is essential because fragmentation is no longer a future risk-it is the current reality. Without some coordination, gaps between regulatory systems will be exploited, weakening both oversight and public trust. Protection and promotion of human rights provides a necessary anchor in a rapidly evolving space. As AI systems increasingly shape access to services and opportunities, governance must be grounded in established normative frameworks rather than reactive regulation. Transparency, accountability, and human oversight are what make governance credible. Without enforceable mechanisms such as auditability, traceability, and clear liability-these principles remain aspirational. Together, these priorities focus on shifting AI governance from values-based alignment to institutional responsibility and practical enforcement.

A central issue that remains under-addressed is the political economy of AI governance specifically, how power is concentrated in a small number of firms and countries that shape both the development of AI systems and the terms of their regulation. Governance discussions that do not account for this risk becoming structurally limited. Second, there is a persistent implementation gap. Many frameworks assume regulatory capacity that does not exist in practice, particularly in lower-resource settings. Without investment in institutional capability, governance risks becoming uneven and ineffective. Third, there is insufficient focus on measuring real-world impact. Current approaches prioritise compliance whether systems meet predefined standards-rather than whether they produce equitable and beneficial outcomes in practice. Finally, everyday harms including gendered abuse, exclusion from services, and biased automated decision-making remain peripheral in many high-level discussions. This reflects a broader tendency to prioritise long-term or existential risks over ongoing, measurable harms. Addressing these issues would require shifting AI governance from a primarily technical and normative exercise to one that engages with power, capacity, and lived experience.

In my context, the governance gaps in AI are not just theoretical they shape how I understand both the risks and limitations of current policy responses. Through my work on a policy brief examining automated moderation and gendered online harm, I found that many AI systems already deployed at scale fail in ways that are not easily captured by existing governance frameworks. For instance, content moderation systems often struggle to identify context-specific or gendered abuse, leaving users exposed despite the presence of "safety" mechanisms. This reflects a broader issue: governance frameworks tend to prioritise system design over lived impact and accountability when harms occur. In the Indian context, this challenge is amplified by the scale and diversity of users, where linguistic, cultural, and socio-economic factors complicate both system performance and regulatory oversight. At the same time, mechanisms for transparency and redress remain limited, making it difficult for affected individuals to challenge automated decisions. I have also observed that much of the governance discourse is shaped by developments in a few regions, while countries like India are often positioned as adopters rather than agenda-setters. This creates a gap between global standards and local realities. However, this also presents an opportunity. India's experience with large-scale digital systems creates space to develop more grounded, implementation-focused governance models—particularly those that centre accessibility, inclusion, and public accountability. These experiences have shaped my interest in AI governance as a field that must move beyond principles and engage directly with how systems affect people in practice, especially in underrepresented contexts.

From my perspective, the main role of the AI Dialogue should be to connect conversations that are currently happening in isolation. In my own work on AI governance, I've noticed that technical discussions, policy frameworks, and real-world harms often sit in separate spaces. This makes governance feel coherent in theory, but fragmented in practice. The Dialogue can help bridge this gap by translating principles into clearer expectations for implementation. It can also play a role in creating baseline alignment on high-risk AI systems, even if countries take different regulatory approaches. Full consensus may not be realistic, but shared minimum standards would still be meaningful. Importantly, the Dialogue should enable more active participation from countries like India, which are often implementing AI at scale but have less influence in shaping global rules. For me, international cooperation becomes meaningful only when it reflects diverse realities, not just shared principles. Finally, its value would lie in enabling practical collaboration such as shared evaluation methods or pilot governance approaches—so that cooperation leads to learning, not just discussion.

From what I've observed, there is already a strong foundation of AI governance initiatives, such as the OECD AI Principles and UNESCO's Recommendation on the Ethics of AI, along with more recent processes like the G7 Hiroshima AI discussions.However, in my experience engaging with AI governance through research, these efforts often feel disconnected from each other and from implementation realities. They provide important direction, but not always clarity on how different frameworks interact in practice. The AI Dialogue could add value by acting as a linking platform helping map how these initiatives relate, where gaps exist, and how they can be applied across different contexts. It could also contribute by bridging global frameworks with local realities, particularly in countries where governance capacity and deployment contexts differ significantly from where many of these standards are developed.For me, the added value is not in creating new principles, but in making existing ones more usable, connected, and implementable across contexts.

Different stakeholders can contribute meaningfully to the AI Dialogue, but only if the structure moves beyond traditional panel-based formats. From my perspective, effective participation requires role clarity. Governments can shape regulatory direction, the private sector can provide insight into system design and deployment, and civil society and researchers can highlight real-world impacts and accountability gaps. However, these perspectives are often presented in parallel rather than in dialogue.The structure should therefore prioritise interactive, problem-focused sessions for example, case-based discussions where stakeholders respond to the same real-world scenario (such as a high-risk AI deployment) from their respective roles. This would make differences in priorities and constraints more visible and actionable.It would also be valuable to include regional breakouts, where participants can discuss context-specific challenges and feed those insights back into global discussions. Finally, there should be clear outputs from each session, such as short recommendations or identified trade-offs, so that participation contributes to cumulative outcomes rather than isolated conversations.

AI governance becomes more effective when it reflects not just those who design systems, but those who live with their consequences.

One effective approach could be scenario-based simulations, where participants are given a shared case—such as an AI system causing harm and asked to respond from their institutional perspective. This would surface differences in priorities, constraints, and governance approaches in a practical way.

some of the most effective approaches to AI governance are emerging not as formal regulations, but as practices that attempt to make systems contestable and accountable in real-world settings.One example is the idea of "contestability by design"building systems where users can question, challenge, or override automated decisions. In my own work examining automated moderation and gendered harm, I found that the absence of meaningful appeal mechanisms often matters more than the model's technical accuracy. Governance, in this sense, is not just about preventing harm, but about ensuring recourse when harm occurs. Another promising approach is the development of independent auditing and red-teaming practices for AI systems. While still evolving, these introduce adversarial testing and external scrutiny into the lifecycle of AI, shifting evaluation from internal compliance to more realistic assessments of risk.I also find participatory governance models particularly important, where affected communities are involved in shaping how systems are evaluated or deployed. Although still limited in scale, these approaches challenge the assumption that governance should be driven only by technical experts or regulators. Additionally, there is growing attention to post-deployment monitoring, recognising that many harms only emerge after systems interact with complex social contexts. Continuous evaluation, rather than one-time approval, is a critical but underdeveloped aspect of governance. Overall, what stands out to me is that effective AI governance depends less on static rules and more on ongoing oversight, the ability to challenge systems, and responsiveness to lived impacts.