TechAngels

A successful first Global Dialogue on AI Governance would, in my view, achieve at least three things. First, it should help frame the right governance questions by clarifying the responsibilities and corresponding extent of accountability of the main actors across the AI value chain: frontier model developers, infrastructure providers, deployers, governments at both central and local level, and civil society. Too often, AI governance debates remain high-level, while real-world responsibility is fragmented across multiple layers. Second, it should move beyond principles alone and identify areas where practical convergence is possible, especially around transparency, safety, auditability, and redress. Even if full consensus is not immediately possible, success would mean a clearer shared understanding of where minimum common expectations can emerge. Third, it should create a foundation for continued multistakeholder cooperation rather than remain a one-off discussion. The most valuable outcome would be a process that connects policymakers, technical builders, companies, and affected communities in a structured way, with a focus on implementation and not only aspiration. In short, success would mean leaving the Dialogue with greater clarity on who is responsible for what, where international alignment is realistic, and how to turn broad governance goals into workable mechanisms.

• Safe, secure and trustworthy AI
• Interoperability of governance approaches
• Open-source software, open data and open AI models

I selected safe, secure and trustworthy AI, interoperability of governance approaches, and open-source software, open data and open AI models because these are three areas where governance must become both more practical and better aligned with incentives. First, safe and trustworthy AI is essential, but regulation should not rely only on sanctions. It should also create incentives for actors across the AI stack to invest in transparency, auditability, safety-by-design, and meaningful risk controls. A good legal framework should not merely punish failures; it should make responsible behaviour easier, more visible (replicable), and more valuable. Second, interoperability of governance approaches is vital to ensure a common baseline of values, guardrails, and responsibilities across jurisdictions. AI systems are developed and deployed across borders, so fragmented approaches create unnecessary legal uncertainty and weaken effective implementation. Third, open-source software, open data, and open AI models remain critical for innovation, scrutiny, and competition. Open ecosystems can strengthen research, accessibility, and independent evaluation. They should therefore be included in governance frameworks through proportionate safeguards, rather than treated as outside the conversation. Over time, a more common global baseline for open-source governance could become especially valuable: a shared framework that is intelligible across jurisdictions and increasingly readable by both humans and AI systems, including agents operating autonomously across digital environments. Overall, my priority is governance that combines accountability with incentives: frameworks that preserve innovation while encouraging all stakeholders to adopt trustworthy practices in a realistic and scalable way.

Yes. A major cross-cutting issue is that AI does not fit the traditional subject-setting paradigm. The listed themes are useful, but they remain too compartmentalized for a technology that can act across domains, scale rapidly, and combine knowledge, agency, and automation in ways that challenge existing governance structures. One emerging issue is horizontal misuse risk: AI-enabled fraud, cybercrime, surveillance, manipulation, reputational attacks, and potentially autonomous harmful action. These are not simply narrow use cases. They are systemic risks that can affect the fabric of society at scale. Another is the problem of evaluability. At its current stage, AI can already exploit weaknesses in human oversight by gaming tests, falsifying results, or producing the answers evaluators expect to hear. This means governance cannot rely only on static benchmarking or formal compliance. A further issue is the insufficient treatment of externalities, especially environmental cost, AI-enabled mass surveillance, and autonomous use in warfare. These deserve explicit attention as cross-cutting threats, not only as sub-parts of safety or economic impact. In short, AI governance needs a stronger category for systemic and horizontal risks: those arising not from one sector alone, but from the interaction of scale, autonomy, deception, and misuse across society.

Being in Romania, a country bordering Ukraine, these governance gaps are not abstract. They are felt more directly through the regional security environment, the growing relevance of dual-use technologies, and the increasing exposure to cross-border cyber and information threats. One major challenge is the accelerating development of autonomous and AI-enabled weapons systems. Once deployed, such systems may operate at a speed, scale, and level of autonomy that exceeds meaningful human control in practice. This raises not only safety concerns, but also questions of accountability, escalation, and the adequacy of existing legal frameworks. Unlike traditional regulatory systems, these technologies are not constrained by borders in the same way their effects are not. A second challenge is the rapid improvement of AI-generated media, voice cloning, and automated deception. These capabilities are already making cybercrime, fraud, impersonation, and disinformation more sophisticated, more scalable, and harder to detect. For countries in our region, this creates heightened risks not only for individuals and businesses, but also for public trust, democratic resilience, and institutional credibility. At the same time, there is also opportunity. These pressures can push countries in the region to invest earlier in governance capacity, cyber resilience, verification mechanisms, and more practical cooperation between government, private sector, and civil society. They also create an opportunity for Europe's eastern flank to contribute important real-world perspectives to global AI governance debates, especially on security, resilience, and implementation under pressure. Overall, the most significant challenge is that capability growth is moving faster than governance adaptation. The opportunity is to build frameworks that are realistic, preventive, and responsive before these risks become even harder to contain.

The AI Dialogue can help create the international governance infrastructure that no single nation can credibly build alone. For AI governance to work, it must be respected not as the initiative of one country or one regulatory bloc, but as a framework with sufficient neutrality, legitimacy, and global buy-in to serve as common ground. The real value of the Dialogue is therefore not only in producing statements, but in building shared concepts, trusted processes, and practical coordination mechanisms that states and other stakeholders can continue to rely on even when their interests diverge.

The AI Dialogue should build on existing initiatives that already provide parts of the governance architecture, while adding a more neutral and globally connective layer. These include, in particular, the EU's AI regulatory framework, including the AI Act; UNESCO's Recommendation on the Ethics of Artificial Intelligence, which applies across all 194 UNESCO Member States; the OECD AI Principles, updated in 2024 to reflect general-purpose and generative AI; and practical risk-management tools such as the NIST AI Risk Management Framework. It should also connect with multistakeholder platforms such as GPAI/OECD.AI and the broader UN/ITU "AI for Good" ecosystem. The added value of the AI Dialogue would not be to duplicate those efforts, but to help connect them. Today, many useful initiatives exist, but they are fragmented across jurisdictions, institutions, and stakeholder groups. The Dialogue could provide a more legitimate international space to identify areas of convergence, translate between different governance languages, and reduce the risk that AI governance evolves into parallel and competing systems. It could also help bridge an important gap between principles and implementation. Some frameworks are strong on values, others on regulatory obligations, others on technical risk management, and others on industry practice. The Dialogue could help align these layers more coherently and include perspectives from governments, infrastructure providers, model developers, application companies, civil society, and regions that are often rule-takers rather than rule-makers. In that sense, its value would be to function less as another standalone initiative and more as connective governance infrastructure: a forum that increases interoperability, legitimacy, and practical coordination across an already crowded field.

Different stakeholders should contribute according to their role in the AI value chain. Governments should bring legitimacy, public-interest priorities, and policy experience. Private-sector actors — including infrastructure providers, model developers, application-layer companies, and AI-enabled SaaS providers — should contribute technical reality, deployment experience, and visibility into how systems are actually built and scaled. Civil society, academia, and international organizations should help ensure that human rights, accountability, inclusion, and long-term societal impacts remain central. In terms of format, the Dialogue should combine openness with a more structured decision-shaping architecture. A useful model would include: a multistakeholder plenary for broad legitimacy; focused thematic working groups for practical outputs; and a smaller rotating steering group to maintain continuity across sessions and regions. That steering group should be geographically balanced and should not give any actor a veto. The Dialogue would be most effective if it produced not only discussion, but also concrete outputs: shared terminology, areas of minimum convergence, issue papers, and recommendations for voluntary or regulatory uptake. In that sense, the structure should be closer to a durable coordination mechanism than a one-off conference.

The most underrepresented voices in global AI governance are often those who are likely to be most affected by AI systems while having the least power to shape them. This includes smaller and lower-capacity states, communities already facing structural disadvantage, workers exposed to automation and algorithmic management, children and young people, educators, creators, and populations vulnerable to surveillance, exclusion, or digital manipulation. In many cases, AI does not create inequality from zero; it amplifies existing asymmetries at scale and with greater speed.

Meaningful engagement during the AI Dialogue will likely require a format that is hybrid, recurring, geographically distributed, and participatory by design.

Effective AI governance is more likely to emerge from a combination of public frameworks, technical standards, and concrete product practices rather than from abstract principles alone. One useful example is the publication of model cards and system cards by companies such as Anthropic and OpenAI. These practices can improve transparency around model capabilities, limitations, evaluations, and safety assumptions, and they give downstream users a better basis for understanding how systems behave in practice. One concrete governance practice is application-layer traceability: users should be able to see which model produced an output, whether routing or fallback occurred, and what constraints materially shaped the result.