Khora

A successful first Global Dialogue should not pretend to solve AI governance in two days. It should do something more useful: establish a shared way of seeing where AI is becoming infrastructure before public language, institutional capacity, and democratic oversight have caught up. Three outcomes would matter most. First, the Dialogue should distinguish real convergence from diplomatic consensus language. Many actors can agree on "trustworthy AI" while meaning different things by trust, safety, evidence, rights, sovereignty, and accountability. The Dialogue should make those differences visible rather than hide them. Second, it should produce practical interoperability between governance approaches. The world does not need one universal AI regime, but it does need ways for different legal and institutional systems to compare risks, share evidence, audit systems, and cooperate across borders. Third, it should widen AI governance beyond technical safety. AI now mediates administration, education, healthcare, culture, employment, security, language, and public imagination. Governance must therefore include not only governments, labs, and companies, but also civil society, cultural practitioners, educators, workers, patients, local communities, and people directly exposed to automated decisions. The strongest outcome would be a process that makes AI governance legible to publics, not only negotiable between institutions. That means clear transparency duties, accountable human oversight, rights-based safeguards, and channels through which affected people can contest errors and describe consequences in ordinary language. Success would be a Dialogue that treats AI not only as a technology to be managed, but as a force already reshaping what institutions know, what publics can see, and what societies are being asked to accept as inevitable.

• Social, economic, ethical, cultural, linguistic and technical implications of AI
• Transparency, accountability, and human oversight
• Protection and promotion of human rights
• Interoperability of governance approaches

I selected these priorities because AI governance is not only a question of model safety. It is also a question of institutional power, public evidence, cultural mediation, and human consequence. The social, economic, ethical, cultural, linguistic and technical implications of AI should be central because AI systems increasingly shape how people access knowledge, work, welfare, healthcare, education, culture, and public services. These effects will not be evenly distributed. They will depend on language, disability, income, geography, public capacity, and the strength or weakness of local institutions. Interoperability of governance approaches is urgent because different regions will regulate AI through different legal traditions and political values. The aim should not be false uniformity. It should be practical comparability: shared vocabularies, reporting structures, risk categories, evaluation methods, and cooperation channels. Human rights must remain a baseline, especially where AI affects surveillance, migration, welfare, education, healthcare, policing, employment, political participation, or access to public goods. Rights should not enter the process after systems have already been procured, deployed, and normalized. Transparency, accountability and human oversight are where governance becomes real. People need to know when AI is being used, what it is being used for, who is responsible, how decisions can be challenged, and where human judgment remains answerable. These priorities also reflect my work across immersive technology, public communication, cultural experience, and institutional adoption. AI governance must be able to speak both to technical systems and to lived public experience. Otherwise, it risks governing the machine while missing the world the machine is quietly reorganizing.

One cross-cutting issue is the governance of public imagination. AI systems do not only automate tasks. They change what institutions believe is possible, what publics are told is inevitable, and what societies learn to treat as normal. This matters for procurement, education, healthcare, defence, culture, media, and public administration. Governance should therefore examine not only deployed systems, but also the demonstrations, benchmarks, narratives, and institutional incentives that make certain AI futures appear unavoidable before they have been publicly examined. A second issue is experiential evidence. Many AI harms and benefits are not captured well by technical evaluation alone. People encounter AI through interfaces, delays, denials, recommendations, generated language, surveillance, synthetic media, exclusion, dependency, and altered trust. Governance processes need ways to collect this lived evidence from affected groups, not only expert assessments from vendors, labs, or regulators. A third issue is synthetic and immersive mediation. AI-generated environments, avatars, virtual agents, simulations, spatial interfaces, and emotionally responsive systems will increasingly shape education, therapy, work, memory, culture, and public participation. This raises questions about consent, manipulation, embodiment, identity, emotional influence, accessibility, and the boundary between communication and simulation. A fourth issue is institutional dependency. Public bodies may adopt AI faster than they develop the capacity to audit, contest, understand, maintain, or exit these systems. The danger is not only spectacular failure. It is quiet administrative dependence, where accountability thins out because no one fully owns the system anymore. The Dialogue should treat these as governance questions: how AI alters perception, evidence, trust, institutional capacity, and the public's ability to refuse bad futures before they are built.

In Denmark and Europe, AI governance is developing quickly, but the practical gap is between high-level principles and the daily realities of adoption. My sector sits between technology, culture, public communication, immersive media, education, healthcare, and institutional innovation. From that position, the challenge is clear: AI is entering workflows, services, creative production, public engagement, and decision-support systems faster than many institutions can build the capacity to understand, audit, contest, or explain them. The most significant challenge is not only technical risk. It is institutional dependency. Public bodies, cultural organizations, SMEs, educators, and healthcare-adjacent actors may use AI systems without enough internal knowledge about data provenance, model limits, accountability chains, rights impacts, or exit options. This creates a weak layer between formal governance and lived consequence. A second challenge is cultural and linguistic. Smaller language communities and public cultures risk becoming dependent on systems optimized elsewhere, trained through unequal data conditions, and evaluated through benchmarks that do not capture local meaning, trust, vulnerability, or democratic context. A third challenge concerns synthetic and immersive media. AI-generated avatars, simulations, spatial interfaces, and emotionally responsive agents create opportunities for education, therapy, accessibility, cultural heritage, and public participation. They also raise unresolved questions about consent, manipulation, identity, embodiment, emotional influence, and the difference between informing people and placing them inside persuasive environments. The opportunity is that Europe can lead in governance that is not only compliance-based, but public-facing and human-readable. Strong governance could help SMEs, cultural institutions, public bodies, and civil society adopt AI responsibly, document impacts, build trust, and create better forms of public evidence. For my region and sector, the most important task is to make AI governance usable before AI becomes invisible infrastructure.

The AI Dialogue can advance international cooperation by becoming a practical translation layer between governance systems, regions, sectors, and forms of evidence. AI governance is already developing across many forums: human rights, standards, safety, development, public-sector adoption, digital infrastructure, open-source models, education, labour, and industrial policy. These efforts are necessary, but they often use different vocabularies and respond to different institutional pressures. The Dialogue can help make this landscape more legible and more usable. Its role should not be to impose one global model of AI governance. Its value would be to help Member States and stakeholders compare approaches, identify areas of convergence, surface real disagreement, and develop practical cooperation around shared problems: transparency, accountability, risk assessment, capacity-building, public procurement, rights protection, and cross-border impacts. The Dialogue can also help widen what counts as relevant evidence. Technical expertise is essential, but it is not sufficient. AI governance also needs evidence from public-sector practitioners, workers, educators, patients, cultural institutions, SMEs, civil society, local communities, and people affected by automated decisions. International cooperation will remain incomplete if it connects institutions at the top while missing the consequences below. The UN can add value by linking AI governance to human rights, sustainable development, institutional capacity, cultural diversity, and public legitimacy. It can ask questions that narrower forums may not hold open for long enough: who is able to shape AI systems, who is exposed to them, who can contest them, and which forms of dependency are being created. The Dialogue should therefore become a standing place where the world compares not only AI rules, but also AI consequences.

The AI Dialogue should connect existing initiatives rather than duplicate them. It should build on the Global Digital Compact, the Independent International Scientific Panel on AI, UNESCO's Recommendation on the Ethics of AI, the OECD AI Principles and AI Policy Observatory, the Council of Europe Framework Convention on AI, the EU AI Act, the G7 Hiroshima AI Process, ITU's AI for Good work, ISO/IEC standards processes, Internet Governance Forum discussions, and regional processes in Africa, Latin America, Asia and the Arab region. Each of these mechanisms has a different strength. Some provide scientific assessment, some ethical principles, some legal obligations, some technical standards, some development and capacity-building channels, and some multistakeholder participation. UNESCO's Recommendation is a global ethics instrument adopted by UNESCO Member States, with human rights, dignity, transparency, fairness and human oversight among its anchors. The Council of Europe's Framework Convention is the first international treaty on AI and human rights, democracy and the rule of law. The UN process itself is now tied to the Independent International Scientific Panel and the Global Dialogue under General Assembly resolution A/RES/79/325. The added value of the Dialogue should be synthesis, participation and follow-through. It can map where existing initiatives align, where they leave gaps, and where lower-capacity actors need practical support. It can also bring in voices often peripheral to technical and regulatory forums: affected communities, public-service implementers, cultural institutions, educators, SMEs, creative sectors, and civil society. The Dialogue should become the place where AI governance is made readable across systems: what exists, what conflicts, what is missing, who is accountable, and how publics can understand or contest AI systems that increasingly shape daily life.

Different stakeholders should contribute different kinds of evidence, not only different opinions. Governments can bring legal mandates, public-interest obligations, procurement experience and regulatory lessons. Technical experts can clarify system capabilities, risks, limits and evaluation methods. Industry can explain deployment realities, but should not dominate the definition of feasibility. Civil society can identify harms, exclusions and accountability gaps. Cultural institutions, educators, healthcare actors, SMEs, workers and affected communities can show how AI is experienced in daily life, not only how it is described in policy. I would recommend a structure with three layers. First, short plenary sessions should frame the key tensions: rights and innovation, safety and openness, interoperability and sovereignty, automation and human accountability. Second, smaller thematic sessions should be designed around concrete cases rather than general speeches: AI in public services, education, health, labour, culture, synthetic media, procurement and local-language contexts. Third, each session should produce a short "governance record": points of convergence, unresolved disagreements, practical needs, capacity gaps and proposed follow-up actions. Participation should not depend only on who can attend Geneva. The Dialogue should include written inputs, remote participation, regional pre-dialogues, youth and civil-society consultations, and accessible summaries in multiple languages. The format should make one thing difficult to avoid: affected people and implementers should be heard before governance language becomes too clean.

Underrepresented voices include people who experience AI systems without having the language, access or institutional power to influence them. This includes workers managed or evaluated by automated systems; people affected by welfare, migration, policing, education, credit, health or hiring algorithms; smaller language communities; persons with disabilities; children and young people; older people; Indigenous communities; artists and cultural workers; teachers; nurses and frontline public servants; SMEs; local governments; and civil-society groups outside the usual global policy circuits. One missing perspective is the public-sector implementer: the person inside a municipality, hospital, school, cultural institution or agency who is asked to adopt AI before the organization has enough capacity to audit it, explain it, or exit it. Another missing perspective is the person subject to AI-mediated decisions who cannot easily know whether AI was used, who is accountable, or how to appeal. These voices could be included through regional consultations before the Dialogue, funded participation for civil society and lower-resource actors, remote testimony, short case submissions, public evidence hearings, and structured formats where affected communities respond to expert claims rather than being added as symbolic witnesses. The Dialogue should also include cultural and linguistic mediation: translation, plain-language summaries, accessible formats, and community-facing explanations of what AI governance decisions may mean in practice. Inclusion should be measured by whether underrepresented groups can alter the agenda, not merely appear inside it.

The most useful formats would move beyond keynote speeches and allow disagreement, evidence and implementation problems to become visible. I would recommend five formats. First, "case clinics" where stakeholders examine real governance situations: AI in a school, welfare office, hospital, cultural platform, hiring process, border system, or local-language service. Each case should ask: who benefits, who is exposed, who is accountable, and how can the system be challenged? Second, "evidence panels" pairing technical experts with affected communities, public-sector implementers and civil society. This would prevent technical claims from floating above lived consequences. Third, "governance stress tests" where participants examine whether existing rules would work under plausible scenarios: synthetic media in elections, automated denial of public services, AI-generated health advice, emotional AI in education, or public-sector vendor lock-in. Fourth, "interoperability labs" where regulators, standards bodies, companies and civil society compare terms, reporting methods, risk categories and audit practices across regions. Fifth, "public legibility sessions" where complex governance proposals are translated into plain-language implications: what citizens can know, contest, refuse, appeal or audit. The Dialogue could also use immersive or interactive demonstration spaces carefully, not as technology showcases, but as governance encounters. Participants could experience how AI systems classify, recommend, simulate, persuade or exclude, then discuss what oversight would actually require. The goal should be engagement that leaves a record: concrete disagreements, capacity needs, accountability gaps and actions that can be followed up after Geneva.

Some of the most useful AI governance approaches are not only those that regulate models, but those that make AI visible before it becomes institutional furniture. Risk-based regulation, human-rights frameworks, technical standards and policy observatories are all important. The EU AI Act, UNESCO's Recommendation on the Ethics of AI, the Council of Europe Framework Convention on AI, OECD AI policy work, ISO/IEC standards processes and national AI registers all offer useful foundations. But the next layer of governance should focus on legibility, contestability and dependency. One strong practice is a public register of AI systems used by public authorities, written in language ordinary people can understand. Citizens should know when AI is involved in welfare, health, education, policing, migration, employment or public communication. A second practice is mandatory "consequence assessment," not only impact assessment. Institutions should document who is exposed to the system, what decisions it influences, how errors are challenged, what human oversight means in practice, and how the institution can exit the system if it becomes unsafe, opaque or too dependent on one vendor. A third practice is public evidence collection. AI evaluation should include frontline workers, affected communities, educators, patients, artists, SMEs, local officials and civil society. They can detect forms of harm that technical benchmarks miss: delay, denial, humiliation, linguistic exclusion, loss of trust, automation bias, and the slow thinning of human responsibility. A fourth practice is governance sandboxes for culture, education, health and public services, not only for industry. These should test consent, accessibility, emotional influence, synthetic media, human oversight and local-language performance. The best AI governance makes systems inspectable while they are still optional. Once AI becomes invisible infrastructure, accountability becomes archaeology.