PTC Hodling

The first Global Dialogue on AI Governance will be successful if it produces actionable outcomes rather than reaffirming shared values already expressed in existing declarations. Three concrete outcomes would mark genuine progress. First, agreement on a common governance baseline applicable to AI systems operating at scale — covering safety, security, accountability, traceability, human oversight, and data governance. This baseline should be specific enough to inform national regulation and institutional policy, not only aspirational enough to attract consensus. Second, recognition that the integrity of AI-generated content requires dedicated governance mechanisms. As AI systems increasingly rely on AI-produced data, the absence of provenance and verification standards creates compounding risks for institutional decision-making and public trust. Third, practical guidance on implementation how governments and organizations can structure oversight of AI systems through defined roles, metrics, and controls. Without architecture, governance commitments remain formal rather than functional. Success, in short, means the Dialogue produces outputs that institutions can act on.

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

Safe, secure and trustworthy AI is the prerequisite. Any technology, incl. AI should be safe. Planning, operational support, reliability directly affects the quality of institutional decisions. Governance frameworks that do not address safety at the system level will be insufficient. AI capacity-building ensures that governance commitments can be implemented. Many institutions currently lack the technical and regulatory capacity to deploy, monitor, and audit AI systems. Capacity is the enabling condition for all other priorities. Interoperability of governance approaches addresses a structural gap. AI systems increasingly operate across jurisdictions and sectors. Where governance frameworks are incompatible, oversight becomes fragmented and enforcement gaps emerge. Transparency, accountability, and human oversight ensure that responsibility remains clearly assigned. Humans must retain authority over consequential decisions and be able to identify where and how AI systems have influenced outcomes. This is a governance requirement, not a technical preference.

Governance of agent hierarchies. Frameworks for individual model outputs are not suited to layered agent systems. Especially, when task-level agents, organizational agents, and national-level AI will operate in real time across one logical decision chain. Agents require standards on scope of authority, intervention points, and responsibility assignment when outcomes emerge from multi-agent coordination. Integrity of AI-generated content across generations. Mandatory labeling of AI-generated content is infrastructure module. Without provenance standards, synthetic outputs enter training pipelines for subsequent systems, compounding errors across model generations invisibly. The reliability of AI systems will erode structurally if this is not addressed at the ecosystem level. Human capital and the skill shift. As AI agents absorb analytical and operational functions, required human skills are shifting from technical execution toward goal-setting and AI oversight. Governance frameworks have not addressed this transition. Redefinition of education and professional standards - is required to clode the capacity gap.

Across the region, AI adoption pressure is producing a counterproductive pattern: organizations are bypassing process documentation, automation, and data quality controls in favor of deploying AI directly. The assumption is that AI can substitute for structured organizational knowledge. It cannot. AI systems depend on the quality of the data they are grounded on. Where processes are undocumented or inconsistently executed, AI outputs reflect those gaps, and institutions that skip process discipline will find their AI systems underperform, erode trust, and require the foundational work to be done retroactively. The viable path is parallel, not sequential: implement lightweight process automation and structured data capture alongside AI deployment, using these inputs to progressively enrich AI systems with actual organizational context. This closes the loop between process execution and model performance iteratively, without requiring full process maturity upfront. The governance implication is direct: capacity-building frameworks must treat process maturity as part of AI readiness standards. Without it, investment in AI produces capability that is fragile, unauditable, and misaligned with organizational objectives.

The internet in its development had several defining moments: web design with animation, its own subcultures, trolls and communities, and separately the darknet with its own rules around adult and extremist content. Each of these required a response, many of them late. AI will have its own version of these trends. Some will be negative, and they cannot be fully excluded, only minimized. The question is whether the response will be coordinated before these trends scale, or after. The most important thing the Dialogue can establish is a shared answer to a simple question: what can be trusted, and what cannot. Not in abstract ethical terms, but operationally: how trust in AI systems and AI-generated content is established, verified, and withdrawn. Who is responsible at each level. Without this, every platform and every government will make that determination independently. The internet showed us where that leads. The Dialogue has the opportunity to agree on the baseline before AI ecosystems reach the point where coordination becomes structurally difficult.

The OECD AI Principles and the accompanying Policy Observatory offer the most developed framework for cross-country monitoring of AI governance approaches, including comparable data on national regulatory developments. The Dialogue should build on this infrastructure rather than create parallel tracking mechanisms. The UNESCO Recommendation on the Ethics of AI provides a baseline ethical framework with broad adoption. Its implementation assessment methodology offers a model for how the Dialogue might monitor follow-through on commitments. The Global Partnership on AI (GPAI) has produced substantive technical work on responsible AI, data governance, and AI in the workplace. Its working group structure is a model for involving both governmental and non-governmental expertise in concrete deliverables. The ITU's AI for Good platform and the Internet Governance Forum provide existing infrastructure for inclusive participation, particularly from developing countries. The added value the Dialogue can bring is specific. All of the initiatives above have mapped the landscape and established norms. None have produced implementation standards or interoperability frameworks that work across jurisdictions. The Dialogue is positioned to fill that gap: moving from principles that everyone agrees on to governance that can actually be executed. What accountable AI deployment looks like in practice. How agent-based systems should be governed at organizational and national levels. What minimum standards apply regardless of jurisdiction. The distance between agreed principles and executable governance is where the Dialogue has the most to contribute.

The Dialogue needs people who have actually deployed AI, governed it, or lived with its consequences. Governments bring regulatory authority and can share what is working and what is not. Companies bring implementation experience and know where governance frameworks are unworkable in practice. Civil society brings independent assessment of harms that neither side has incentive to surface. One group requires specific attention: people who are structurally limited in their ability to verify AI outputs. Low digital literacy, cognitive impairments, elderly users, children, people with disabilities. These groups will interact with AI agents regardless of readiness. AI is described as reducing workload. In practice it removes mechanical tasks while adding intellectual load: outputs must be checked, agents must be instructed, errors must be caught. This burden is not measured. A mandatory feedback loop on every AI interaction is a governance requirement, not a feature. Without correction signals from real users, AI systems have no adequate mechanism to calibrate against actual human experience.

Institutions in developing economies deploying AI without regulatory infrastructure to govern it. They are most exposed and least represented. Operational practitioners inside governments and organizations who actually configure and monitor AI systems. The people with direct knowledge of what AI does in practice rarely have a seat at the table. Business operators in logistics, retail, and services where existing automation is already suboptimal. These organizations are adding AI agents on top of unstable processes, creating compounding complexity. The people managing this on the ground know exactly where governance is failing.

Problem-based working sessions organized around specific governance failures, not thematic categories. Concrete problems generate concrete outputs. Implementation review panels where governments and organizations present what they built, what broke, and what they changed. Real deployments produce more usable knowledge than position statements. AI-assisted documentation with interpretation feedback. Every session summarized by AI in real time, with the summary presented back to participants before closing. Not for efficiency: to show each participant how their input was interpreted after processing. Where meaning was compressed, reframed, or lost. People who see their position distorted in a summary will understand directly why feedback loops and human oversight matter. It also generates evidence of where AI summarization fails under multilingual and contested conditions: exactly what governance frameworks need to address.

The EU AI Act introduced risk-based classification of AI systems. The logic is practical: governance requirements scale with the consequences of failure. High-risk systems in employment, credit, and public services face stricter obligations than low-risk applications. This tiered approach is more implementable than uniform regulation and more honest about where the actual risks are. Singapore's Model AI Governance Framework went further in one important direction: it gave organizations concrete, testable guidance rather than principles to interpret. Sector-specific annexes, worked examples, and self-assessment tools made it usable by practitioners, not only by legal teams. The Coalition for Content Provenance and Authenticity (C2PA) is developing technical standards for labeling AI-generated content at the point of creation. This is the right level to address the problem: not platform policy, but infrastructure. If provenance is embedded at generation, it travels with the content regardless of where it appears.