Universidad Politécnica de Madrid

The first Global Dialogue will be a success if it creates a room where technical experts and policymakers are genuinely speaking to each other, not past each other. Today, policy people systematically overestimate what AI can do, and technologists systematically underestimate what governance requires. The result is regulation built on technical misunderstandings, and technical systems deployed without accountability frameworks. A successful Dialogue begins to close that gap by institutionalising what might be called meta-translation: the capacity to move fluently between algorithmic logic and legal reasoning, between what a system does and what the law says it must do. Success also means meaningful inclusion of the Global South, understood not merely as geographic representation but as the inclusion of anyone south of power. The communities most harmed by AI systems are rarely the communities that design governance frameworks. A Black man whose behaviour is systematically flagged as suspicious by a predictive policing model, a trans woman denied services by a biometric classifier, a gig worker in the Global South whose labour trains the models that replace them, these are the constituencies this Dialogue must answer to. Success means their experiences are not an afterthought. Finally, success means outputs that are technically honest. Governance frameworks built on inflated assumptions about AI capabilities will fail. The Dialogue should produce recommendations grounded in what AI systems actually do, not what vendors claim they do.

• AI capacity-building
• Social, economic, ethical, cultural, linguistic and technical implications of AI
• Protection and promotion of human rights
• Open-source software, open data and open AI models

AI capacity-building is urgent because the knowledge gap between those who build AI systems and those who govern them is widening faster than any single regulatory framework can address. Without investment in technical literacy among policymakers and legal literacy among technologists, governance will remain performative. The social, economic, ethical, cultural, linguistic and technical implications of AI cannot be separated. Language models trained predominantly on English-language data encode cultural assumptions that travel poorly across contexts. Economic disruption from automation falls disproportionately on workers in the Global South. These are not separate issues requiring separate working groups, they are the same problem viewed from different angles. Protection and promotion of human rights is the floor, not the ceiling. Existing frameworks, like GDPR, the EU AI Act, or the UN Guiding Principles on Business and Human Rights, are necessary but insufficient. They were designed for known harms. AI systems generate novel harms faster than legal frameworks can name them. The Dialogue must develop adaptive rights frameworks, not static ones. Open data and open AI models matter because concentrated AI capability is concentrated power. If only five companies control the frontier models that governments, courts, and public services increasingly depend on, democratic accountability becomes structurally impossible. Open infrastructure is a governance prerequisite, not merely a technical preference.

The most significant gap in the listed themes is what I call the meta-translation problem. Technical and legal communities operate in mutually incomprehensible registers, and no existing institution is systematically addressing this. Policymakers regulate AI systems based on capability claims that are frequently inaccurate. Technologists build systems without understanding the legal frameworks their products activate. The result is regulation that cannot be enforced and technology that cannot be governed. This is not a capacity-building problem in the conventional sense. It is a structural problem of institutional design. God help us, we do not lack experts in AI. We also do not lack experts in law. We lack people and institutions capable of moving fluently between the two, i.e. the meta-translators. The Dialogue should consider how to cultivate and resource this function at the international level. The second gap is the divergence of trajectories. If the meta-translation problem is not solved, the most likely outcomes are not neutral. On the technical side, the absence of meaningful governance creates conditions in which libertarian techno-authoritarianism, governance by algorithm, outside democratic accountability, becomes the default. On the policy side, the absence of technical literacy creates conditions for surveillance infrastructure built on the false premise that AI systems can reliably identify criminals, terrorists, or threats. Neither trajectory is acceptable, but both are rapidly materializing. The space between them is where this Dialogue must operate.

As a Marie Skłodowska-Curie Doctoral Fellow researching European data spaces and AI governance at Universidad Politécnica de Madrid, I sit at the intersection of two governance failures. The first is structural: the EU's regulatory enforcement capacity is profoundly uneven across member states. Italy's Garante is among the most active data protection authorities in Europe. Spain's AESIA remains critically underfunded relative to its mandate under the AI Act. This is not a minor administrative gap. The EU was constructed on two foundations: fundamental rights and the single market. Both are undermined when enforcement of the same regulation produces radically different outcomes in neighbouring countries, before we even consider the gap between southern and northern Europe. The second failure is personal and political. I am a brown, transgender researcher in Europe. Automated classification systems, whether used for border control, access management, or risk assessment, are not neutral. They are built on training data that encodes existing hierarchies of power. A system that cannot reliably classify a trans body, or that flags a brown face as suspicious, is not a technical problem awaiting a technical fix. It is a governance failure. I would be harmed twice by such systems, despite holding a European Commission fellowship. The people with less institutional protection than I have would be harmed without recourse. The opportunities slipping through our fingers are real. European Union's conception of data spaces, if governed well, could redistribute data access and create infrastructure that serves everyone. Technologies like Tim Berner Lee's Solid protocol, which give individuals genuine control over their own data, remain marginalized despite years of development. The gap between the EU's stated ambitions for data sovereignty and the actual adoption of tools that deliver it is where global governance needs to act.

The AI Dialogue needs to be more than a dialogue. The current trajectory of AI governance is one of divergence: major technology companies are building parallel infrastructures that operate outside democratic accountability, while powerful states are developing incompatible regulatory frameworks. This is not a technical problem. It is a political one, and it is a recipe for sustained international instability. The Dialogue's most important function is to resist that divergence. Not through harmonisation for its own sake, but because the alternative, a world in which the five largest AI companies each operate under different rules in different jurisdictions while actively lobbying against coherent governance, produces outcomes that harm everyone outside the most powerful institutions. Companies will seek the least regulated jurisdiction. States will race to attract them. The people who pay the price are always the same people. The Dialogue should evolve toward something with genuine authority, a standing council with technical expertise and enforcement mechanisms, not a recurring conference. Multilateralism is imperfect. But the alternative is not bilateral agreements between the powerful. The alternative is no agreement at all.

The existing initiatives have largely failed to constrain the actors most in need of constraining. The OECD AI Principles are respected by governments that were already inclined to govern responsibly. The Hiroshima AI Process produced a voluntary code of conduct that frontier AI companies signed without consequence. The Global Partnership on AI generated substantial research and negligible policy change. These are not failures of intent. They are failures of power. None of these bodies can compel the actors who matter most to participate meaningfully. The UN is the only genuinely global body. It is deeply imperfect. Its failures on Palestine, on Security Council paralysis, on enforcement, are real and serious. But it is also the only institution where a Francesca Albanese can exist, an independent expert with a mandate to speak truth to the most powerful states in the world. That function does not exist in the OECD. It does not exist in the G7. It does not exist in any bilateral agreement. The AI Dialogue should build on the UN Scientific Panel, on the work of the ITU, and on the Global Digital Compact. But its added value is not coordination. Coordination exists already and has not worked. The added value is legitimacy: the capacity to say, on behalf of the entire international community, that certain uses of AI are unacceptable, and to mean it. That requires the UN. Nothing else is close.

I will say what I feel honestly. The AI Dialogue requires a permanent secretariat, not a recurring conference. Recurring conferences produce communiqués. A permanent body with rotating membership can produce accountability. Rotation should draw from two communities that currently do not speak to each other: frontier technical researchers and legal or policy experts. Not academics who study AI from a distance, but people who have built the systems under discussion. The researcher who designed the diffusion model architecture that underlies most contemporary image generation should be in the same room as the legal scholar drafting the regulation that governs it. Currently they are not, and the gap shows in both the technical illiteracy of regulation and the governance illiteracy of technical development. The most effective engagement format is structured public exchange between frontier AI researchers and legal experts, conducted in public, with the explicit purpose of producing shared vocabulary. Not a joint paper. A shared vocabulary, a common set of terms that both communities accept as accurate descriptions of what these systems do and what the law requires. Without that, every governance document will be written in a language that half the relevant experts do not recognise.

The most commonly cited gap is demographic: Global South communities, women, minority groups. These gaps are real. But the more pressing and less acknowledged divide for me is disciplinary. Even when computer scientists participate in AI governance discussions, they are rarely frontier researchers. The people who understand how large language models actually work, how diffusion models are trained, what the real failure modes of a classifier are, are almost entirely absent from governance spaces. Instead, governance discussions feature researchers whose expertise is in human-computer interaction, participatory design, or AI ethics as a social science. These researchers do important work. But they cannot tell a regulator why a specific technical mandate is unenforceable. They have not written a training loop. The result is that the people building the most consequential AI systems have no interlocutor in governance spaces they respect. They go to Palantir and Meta instead. The Dialogue will not change this by adding more civil society representatives. It will change this by recruiting people who understand both backpropagation and the Charter of Fundamental Rights, and by treating that combination as the scarcest and most valuable resource in the room.

The most effective format is the least glamorous one: direct public engagement. Schools, universities, community centres. A researcher explaining what these systems actually do to a neighbourhood in León México or Lagos Nigeria, and listening to what people actually fear, is worth more than a hundred side events at a Geneva summit. Hackathons produce prototypes that go nowhere. Side events are attended by people who were already in the room. Ministerial roundtables produce language that was agreed before anyone sat down. Public engagement with young people is not a communications strategy. It is a governance strategy. The next generation of AI researchers and policymakers is in secondary school right now. If they grow up understanding both the technical and legal dimensions of these systems, the translation problem solves itself over twenty years. That I think is the most important investment the Dialogue could make, and the least likely to appear in any formal recommendation.

The most effective AI governance approaches share a common feature: they treat open infrastructure as a public good rather than a regulatory target. Aaron Swartz understood this. His work on Creative Commons, RSS feed, and open access was not activism in the conventional sense. It was infrastructure building. He believed that the architecture of information systems determines who has power over them, and that making infrastructure open and interoperable is itself a governance act. He was right, and the world is poorer for having lost him. Tim Berners-Lee's Solid protocol represents the same logic applied to personal data. Solid gives individuals genuine control over their own data through decentralised storage. It is technically sound and legally coherent with GDPR's principles. It has not reached its potential because governments have not adopted it. If governments mandated Solid the way they could mandate open standards, or the way mandating Linux on public sector computers would have transformed technical education worldwide (I wish this was done sooner), data sovereignty would be infrastructure rather than aspiration. The most encouraging recent example is this tech company called Anthropic handing the Model Context Protocol, designed for communication between systems and language models, to the Linux Foundation. Open protocol governance through neutral foundations, with transparent development and community participation, is a model that works. The Linux Foundation has demonstrated this repeatedly. Internet itself is a free and open adaptation. MCP now belongs to that tradition. The EU AI Act's codes of practice process attempted something similar but illustrates the limits of act-by-act governance. The proposed Digital Omnibus is not the problem. It is a symptom. The problem is that reactive regulation tied to specific legislative instruments will always be vulnerable to the next simplification proposal, the next political cycle, the next industry lobbying effort. Governance through open infrastructure standards, publicly mandated and neutrally administered, is more durable than governance through prohibition.