DeshkaAI (Independent AI Stability Research & System Design)

A successful Global Dialogue on AI Governance should move beyond broad principles and produce actionable, system-level outcomes. First, it should establish shared operational standards for how AI systems behave in real-world conditions, especially in high-stakes environments. This includes expectations around decision reliability, handling uncertainty, and ensuring that systems do not act prematurely without sufficient validation. Second, it should define practical mechanisms for human oversight — not only at a policy level, but embedded within system workflows where critical decisions are made. Third, the Dialogue should create collaborative testing and evaluation frameworks that allow different countries and organizations to assess AI systems under common conditions, improving global trust and interoperability. Finally, success would mean bridging the gap between policy and implementation — translating governance principles into clear design patterns that developers and organizations can apply. If the Dialogue achieves this shift from "what AI should be" to "how AI should behave," it will create a strong and lasting foundation for safe and scalable AI systems worldwide.

• Safe, secure and trustworthy AI
• Interoperability of governance approaches
• Transparency, accountability, and human oversight
• Social, economic, ethical, cultural, linguistic and technical implications of AI

The selected priorities are interconnected and critical for building reliable AI systems at scale. Safe and trustworthy AI is the foundation, but safety must be ensured not only through policies, but through system behavior during real-world use. This requires mechanisms that ensure decisions are stable, validated, and context-aware. Interoperability of governance approaches is essential because AI systems operate across borders. Without alignment, inconsistent standards can create gaps in safety and trust. Transparency, accountability, and human oversight are necessary to maintain control over AI systems, especially in high-impact scenarios. However, these must be embedded into system design rather than treated as external checks. Finally, AI capacity-building ensures that all regions - especially developing ones - can adopt and implement these practices effectively, without compromising safety for speed. Together, these priorities support a balanced approach where innovation continues, but within a stable and accountable framework.

A key cross-cutting issue not fully captured is the lack of focus on decision-time stability in AI systems. Most current discussions emphasize ethics, bias, and transparency, but there is limited attention on how AI systems behave at the exact moment of decision - especially under uncertainty, incomplete data, or high-pressure conditions. This gap can lead to premature, inconsistent, or overconfident outputs, even in otherwise well-designed systems. In real-world deployments, such instability can have significant consequences across sectors like healthcare, finance, and governance. Addressing this requires introducing system-level mechanisms that validate context, apply controlled response timing, and allow for structured pauses before critical actions. This "decision stability layer" is not about slowing innovation, but about ensuring that AI systems remain reliable under real conditions. Integrating this perspective into global governance discussions would strengthen the effectiveness of existing frameworks and reduce risks associated with rapid deployment.

In my region (India and the broader Asia-Pacific), AI adoption is scaling rapidly across sectors such as public services, finance, healthcare, and digital infrastructure. While this creates significant opportunities for efficiency and access, governance gaps remain in how AI systems behave under real-world conditions. One of the main challenges is the gap between policy-level guidelines and system-level execution. Many frameworks define what AI should achieve (fairness, transparency, accountability), but do not ensure how systems make decisions in situations involving uncertainty, incomplete data, or operational pressure. This can lead to premature, inconsistent, or overconfident outputs, especially in high-impact environments. Another challenge is uneven capacity across regions, where rapid adoption often outpaces the ability to implement strong safety and validation mechanisms. At the same time, there is a major opportunity to strengthen governance by integrating system-level safeguards into AI design. This includes mechanisms for validating context, introducing controlled response timing, and ensuring stable decision-making before critical actions are taken. For fast-growing digital ecosystems, this approach can bridge the gap between innovation and reliability — ensuring that AI systems are not only scalable, but also safe, consistent, and trusted by users and institutions.

The AI Dialogue can play a key role by transforming global cooperation from principle-based alignment to system-level coordination. While many countries agree on values such as safety, transparency, and accountability, there is fragmentation in how these are implemented in real-world AI systems. The Dialogue can act as a bridge between policy and engineering by defining shared expectations for how AI systems should behave under real conditions — especially in situations involving uncertainty, scale, and high-stakes decision-making. It can also enable collaborative evaluation frameworks, where different regions test AI systems under common scenarios, improving interoperability and trust across borders. Additionally, the Dialogue can facilitate knowledge-sharing between policymakers, researchers, and developers, ensuring that governance principles translate into practical design patterns. By focusing on alignment at the level of system behavior — not just intent — the AI Dialogue can help build globally reliable, stable, and trustworthy AI systems.

The AI Dialogue should build upon existing initiatives such as the OECD AI Principles, UNESCO's Recommendation on AI Ethics, ITU's AI for Good platform, and various national regulatory frameworks. These efforts have established strong foundations around ethics, rights, and governance principles. However, there remains a gap between these high-level frameworks and how AI systems operate in practice. The added value of the AI Dialogue would be to translate these principles into system-level implementation guidelines. This includes defining how AI systems should handle uncertainty, validate context before decisions, and avoid premature or unstable outputs in critical situations. The Dialogue can also act as a coordination platform to share real-world testing methods, deployment learnings, and design patterns across regions. By addressing the missing layer — how AI behaves at the point of decision — the Dialogue can strengthen existing initiatives and ensure that AI systems are not only ethical in theory, but also reliable and stable in practice.

Different stakeholders can contribute by aligning their roles with both policy and system-level implementation. Governments can define regulatory frameworks and ensure accountability, while researchers and technical communities can translate these into practical system designs and evaluation methods. Industry can contribute real-world deployment insights, and civil society can ensure that human impact and inclusivity remain central. For the format, the AI Dialogue should move beyond static discussions and include structured, outcome-driven sessions. This could involve scenario-based workshops where stakeholders evaluate how AI systems behave in real-world situations, especially under uncertainty or high-stakes conditions. Additionally, smaller working groups focused on specific challenges (e.g., decision reliability, human oversight mechanisms) can produce actionable recommendations rather than broad statements. A hybrid structure combining open dialogue, technical breakout sessions, and collaborative problem-solving would ensure that contributions are not only shared, but also translated into implementable outcomes.

Several important voices remain underrepresented in AI governance discussions. These include practitioners from developing regions, grassroots implementers of AI systems, and individuals working directly with real-world deployment challenges rather than only policy or theory. In addition, there is often limited representation from system designers and engineers who focus on how AI behaves during actual decision-making processes. Their perspective is critical for bridging the gap between high-level principles and real-world system behavior. To include these voices, the Dialogue should actively enable participation through accessible formats, regional representation, and targeted outreach to technical communities and independent researchers. It should also create space for practical insights — such as case studies, system failures, and field experiences — rather than only theoretical or policy-level discussions. Including these perspectives will make governance more grounded, ensuring that policies are informed by how AI systems actually function in diverse and real-world environments.

To foster meaningful engagement, the AI Dialogue should adopt interactive and problem-oriented formats rather than traditional panel discussions. One effective approach would be scenario-based simulations, where participants analyze how AI systems behave in specific real-world situations, such as high-risk decision-making or uncertain data environments. This encourages deeper understanding beyond abstract principles. Another format could include collaborative design sessions, where mixed groups of policymakers, engineers, and researchers co-develop practical solutions or governance mechanisms. Short, focused intervention slots combined with moderated discussions can ensure diverse voices are heard without losing depth. Additionally, creating ongoing digital collaboration spaces (before and after the Dialogue) can help continue discussions and track implementation of ideas. These formats would shift engagement from passive listening to active problem-solving, making the Dialogue more impactful and outcome-driven

Several existing initiatives provide strong foundations for AI governance, including the OECD AI Principles, UNESCO's Recommendation on the Ethics of AI, and ITU's AI for Good platform. These frameworks emphasize fairness, transparency, accountability, and human rights, and have been effective in shaping global understanding and policy direction. In practice, some organizations have also adopted risk-based approaches, human-in-the-loop systems, and post-deployment monitoring to improve reliability. These are important steps toward responsible AI deployment. However, a key gap remains in ensuring how AI systems behave at the moment of decision, especially under uncertainty or real-world pressure. Current practices often focus on outcomes and compliance, but less on the internal decision process itself. A promising approach is to integrate system-level validation mechanisms - such as pre-decision checks, context verification, and controlled response timing - to ensure that AI systems do not act prematurely or inconsistently. Combining existing ethical frameworks with such operational safeguards can significantly improve the stability and reliability of AI systems, making governance not only principle-driven, but also behavior-driven in real-world applications.