University of Florida

UNESCO's own have not yet articulated coherently the problem of epistemic monoculture: a governance frame that treats epistemic and cultural diversity as a structural safety condition, not an inclusion concern. Three concrete conversations would be interesting: First, the Dialogue should recognize that AI systems around the world are converging on similar architectures, trained on overlapping data, optimized against the same benchmarks. Of the world's roughly 7,000 languages, fewer than 100 appear meaningfully in major AI models. Existing frameworks treat this narrowing as a technical artifact or a fairness concern. Naming it as a structural risk, e.g. a system has fewer responses available to it than the world it must operate in, gives Member States a category they currently lack. Second, the Dialogue should commit to measures that track not only visible harms but the conditions under which AI remains trustworthy over time. Some harms can be undone; others cannot. Indicators of how similar AI systems are becoming, which languages and knowledge traditions they meaningfully represent, and where local capacity to develop AI is being lost should sit alongside benchmarks of capability and risk. Third, the Dialogue should not become a parallel principles document. Its distinctive value lies in coordinating with the existing forums -- including AI for Good and the WSIS Forum, with which it shares Geneva in 2026 -- to establish that diversity is a precondition for AI that is safe, resilient, and genuinely beneficial, not a value to be balanced against deployment speed.

• Safe, secure and trustworthy AI
• Social, economic, ethical, cultural, linguistic and technical implications of AI

Technical and social need to be understood together - the urgent governance question is what happens at their intersection. AI systems become unsafe and untrustworthy in part because the cultural, social, and ethical assumptions built into them go unexamined. A model trained predominantly on English-language data, optimized against benchmarks developed within a narrow set of institutions, and aligned to preferences expressed by a non-representative slice of users will perform reliably in some contexts and fail unpredictably in others. Those failures are not bugs to be patched after deployment - they are the predictable consequence of building systems on a narrow cultural base and treating that base as neutral. Safety is therefore inseparable from the question of whose worlds a system has been built to navigate. The Renwick Program for Safe, Ethical and Beneficial AI at the University of Florida works directly at this intersection. Our research examines how cultural assumptions embedded in AI systems produce blind spots that surface downstream as safety failures, fairness failures, and erosion of local capacity to develop and govern AI. Current work includes a recent paper in AI and Ethics and the development of an AnthroBench framework for evaluating cultural competence in large language models - both aimed at giving governance bodies the empirical tools to ask these questions concretely. We would bring to the Dialogue both technical fluency and an anthropological lens, with the goal of helping establish that trustworthy AI and culturally aware AI are not parallel agendas but the same agenda seen from two angles.

Yes, questions around epistemic diversity, representation and inherent brittleness because of homogeneity in models as outlined above.

We lack leadership to make important decisions around AI governance, while also harboring the companies that develop potentially dangerous technology without that necessary oversight.

It could bring together different perspective and disciplines necessary to come up with really creative and innovative solution. It would also mean that currently underrepresented voices are heard. Dialogue can surface asymmetries the existing landscape obscures. Most current frameworks treat AI-producing and AI-consuming nations as a single audience for the same rules. Naming this asymmetry and treating the capacity to participate in AI development, not only to regulate it, as a governance question would shift the international conversation in a way no other forum is positioned to do. The Dialogue can coordinate without consolidating. Its value lies less in writing new principles than in ensuring that the existing fora, e.g. AI for Good, the WSIS Forum, the Council of Europe Convention, etc. - do not work at cross-purposes. A clear coordination mandate is more useful than another principles document. International cooperation on AI requires that Member States be able to engage technically, not only diplomatically. Capacity-building for evaluation, governance, and the ability to develop AI locally should be among the Dialogue's defining contributions.

The AI Safety Institutes Network, OECD's AI Policy Observatory, and the standards work of ISO/IEC JTC 1/SC 42 and NIST are building the evaluation infrastructure international cooperation will require, complemented by multistakeholder bodies such as IASEAI (the International Association for Safe and Ethical AI), whose 2026 annual meeting was hosted at UNESCO House. The Dialogue can ensure this combined work reflects linguistic and cultural diversity in evaluation rather than defaulting to a narrow set of benchmarks. The WSIS Forum and the AI for Good Global Summit share Geneva with the inaugural Dialogue in 2026. Coordinated programming across the three would make 2026 a coherent moment rather than three parallel ones. The Dialogue's distinctive added value should be to weave normative, legal, regional, technical into a coordinated whole, with particular attention to the Global South initiatives the current landscape underweights.

Member states give legitimacy, and more useful when grounded in domestic expertise rather than diplomacy, which means including cultural experts not ministers. Civil society can provide testimony, including communities most directly affected by AI development and deployment. Academics can provide empirical and conceptual tools, e.g. for model evaluation, development of new measures (e.g. cultural diversity), etc. Industry, because they built and deploy. Indigenous communities to question knowledge framework on which AI systems are built on.

see above

Multilingual and hybrid participation as default, not exception. International cooperation requires that non-anglophone member states and civil society participants engage substantively, not only diplomatically. Translation and interpretation should extend beyond the UN6 where regional consultations identify the need, and hybrid participation should be funded so that representation does not depend on travel budgets. Red-teaming sessions that challenge consensus positions before deployment. Cross-regional problem framing.

The CARE Principles for Indigenous Data Governance and Te Mana Raraunga (Māori Data Sovereignty Network) have been adopted by research institutions, governments, and data repositories.