Barcelona School of Management at Universitat Pompeu Fabra/I4T Global Knowledge Network

Success requires the Dialogue to move from principles to architecture. Three concrete outcomes would mark it as genuinely consequential. First, a shared normative commitment that information is a public good — not merely a commodity — and that AI systems operating as information infrastructure carry corresponding public obligations. This builds on the UNESCO Windhoek Declaration (2021) and needs to be operationalized, not just reaffirmed. Second, agreement on a minimum floor of accountability obligations that applies to large AI systems regardless of jurisdiction — including mandatory transparency reporting on training data composition, systemic risk assessment for information ecosystem impacts, and researcher access to platform and AI data for independent oversight. Third, a concrete commitment to inclusion in governance design itself — not only as a value to be protected but as a procedural requirement. Global South regulators, civil society organizations, and communities whose languages and cultures are underrepresented in AI training data must have a structured role in the governance frameworks the Dialogue recommends. Outcomes drafted exclusively by the countries and companies that dominate AI development will replicate the structural inequalities they purport to address.

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

The I4T Global Knowledge Network works at the intersection of platform governance, freedom of expression, and information integrity across multiple regional contexts. Our selection reflects that intersection. Social, cultural and linguistic implications sit at the core of our work. AI systems trained overwhelmingly on data from a handful of dominant languages and cultures reproduce structural inequalities at unprecedented scale and speed. Communities across the Global South - whose media ecosystems, indigenous knowledge traditions, and languages are marginalised in training datasets - experience AI not as a neutral tool but as infrastructure that systematically misrepresents or erases them. This demands urgent, specific action, not only general commitments to inclusion. Human rights protection provides the normative foundation. The right to receive information from diverse sources, to participate in cultural life, and to non-discrimination are directly implicated by how large language models are trained, deployed, and regulated. These are not new rights - they are existing obligations that need to be applied coherently to AI systems. Transparency and accountability are the mechanism through which all other commitments become enforceable. Transparency that serves only regulators is necessary but insufficient; the Dialogue should develop layered frameworks that also serve researchers and - critically - affected communities themselves. Interoperability of governance approaches is essential because AI regulation is being designed simultaneously across dozens of jurisdictions, and fragmentation risks becoming a tool for regulatory arbitrage. Shared methodological tools and mutual recognition mechanisms would allow smaller and less-resourced regulators to participate meaningfully in governance rather than simply receiving frameworks designed elsewhere.

Two issues cut across all four clusters but appear in none as a named theme. The convergence of AI and platform regulation. Large language models are not merely AI applications - they are information infrastructure, and they increasingly operate as search engines, recommendation systems, and content mediators simultaneously. Yet AI governance and platform governance continue to be designed in parallel tracks, creating significant accountability gaps. The Dialogue should explicitly address the hybrid status of AI systems in the information ecosystem and recommend that existing platform accountability frameworks apply to them - systemic risk assessment, independent auditing, transparency reporting, and researcher data access - without waiting for new AI-specific legislation. Cognitive autonomy and epistemic design. No current governance theme names the risk that AI systems, through design choices rather than any single output, systematically erode users' capacity for independent reasoning. Chatbot attachment mechanics, persuasive interface design, and the substitution of generative outputs for genuine information-seeking are not harms that can be caught by content moderation or factual accuracy audits. They require a distinct regulatory category that evaluates AI design against standards protecting epistemic self-determination. This is closely related to media and information literacy, but it demands that obligations sit with designers, not only with educators.

On social, cultural and linguistic implications: Communities whose languages are marginalized in AI training datasets already experience distorted or absent representation in LLM outputs. Journalists, educators, and civil society organizations working in African, Indigenous, and lower-resource languages report that generative AI tools reproduce dominant-language framings when engaging with their communities' histories, political contexts, and cultures. This compounds existing inequalities in the media ecosystem rather than offsetting them. On human rights: In several regions around the world, AI-assisted content moderation deployed by global platforms has disproportionately suppressed political speech in local languages, partly because lower-resource languages receive less investment in moderation quality assurance. The same asymmetry is now appearing in generative AI outputs. Without binding non-discrimination auditing requirements, these harms are invisible to regulators and unactionable by affected communities. On transparency and accountability: Many regulators lack the technical capacity, legal authority, and data access to conduct meaningful oversight of AI systems operating in their information ecosystems. The DSA provides a model, but it applies only within the EU. Outside that jurisdiction, transparency is largely voluntary, and companies face no obligation to share the data that would allow independent assessment of systemic harms. On interoperability: The proliferation of national AI regulatory initiatives — often designed without coordination — is already producing divergent standards that larger actors can navigate and smaller ones cannot. Regional regulatory cooperation mechanisms are underdeveloped, and Global South regulators risk becoming rule-takers rather than rule-makers in frameworks that shape their information environments. The primary opportunity is that the convergence of platform and AI governance, if addressed now, could produce frameworks genuinely capable of applying across contexts, but only if the dialogue invests in shared methodology and inclusive design from the outset.

The Dialogue's most valuable could be serving as the space where AI governance and platform governance are brought into coherent alignment at the international level. Existing international AI governance efforts like the OECD AI Principles, the Council of Europe AI Convention, the G7 Hiroshima process or UNESCO's own Recommendation on the Ethics of AI have produced important normative foundations. But we still need to work on an accountability architecture that applies consistently to AI systems operating as information infrastructure across jurisdictions. Concretely, the Dialogue can advance international cooperation by establishing a shared interpretive framework that clarifies how existing international human rights law, including rights to information, non-discrimination, and cultural participation applies to large AI systems. This is not a call for new rights but for authoritative guidance on existing obligations, which national regulators urgently need. We also need peer-learning and mutual recognition mechanisms that allow regulators in lower-capacity jurisdictions to draw on risk assessments, audit methodologies, and transparency data produced elsewhere. Governance capacity is as unevenly distributed as AI development itself. It is important for the Dialogue to treat capacity-sharing as a structural feature of its design, not an add-on. Third, by providing a forum where the fragmentation of national AI regulatory frameworks can be actively managed, identifying minimum interoperability requirements that preserve regulatory diversity. We need a baseline of accountability that applies regardless of where a system is deployed or where its developer is incorporated. The Dialogue could build the normative and methodological infrastructure that makes enforcement coherent.

Several existing initiatives provide essential foundations that the Dialogue should connect with rather than duplicate. UNESCO's Recommendation on the Ethics of AI (2021) and the Windhoek Declaration on Information as a Public Good provide the normative anchors the Dialogue should build from. Connecting these two frameworks explicitly, establishing that AI systems operating as information infrastructure carry public good obligations, would give the Dialogue a distinctive normative contribution rather than another principles document. The ITU's AI for Good initiative and the OECD.AI Policy Observatory provide technical and policy data infrastructure that the Dialogue should leverage for comparative regulatory mapping. Regional initiatives like the African Union's AI Continental Strategy, the CELAC digital agenda, and emerging ASEAN AI governance frameworks represent governance development that is often invisible in Geneva-centred processes. The Dialogue should create structured linkages with these bodies. The I4T Global Knowledge Network's Periodic Table of Platform Regulation offers a methodology for mapping governance frameworks comparatively across dimensions and jurisdictions directly applicable to AI governance interoperability questions. We would love to engange through this scheme.

Governments should contribute not only as regulators but as responsible deployers. Many states are already procuring and deploying AI systems in public services affecting millions of people. Their experience as buyers and accountability-holders is as relevant as their role as rule-makers. Civil society and affected communities need more than observer status. The Dialogue should establish substantive working tracks where civil society organizations, particularly those representing linguistic minorities, indigenous communities, journalists, and human rights defenders, contribute to drafting, not only to commenting. This requires dedicated resourcing: translation, travel support, and advance document sharing in accessible formats. Academic and research networks should be formally integrated into the Dialogue's evidence base. Independent research findings, including on AI impacts in underrepresented languages and regions, should have a defined pathway into Dialogue outcomes, reducing dependence on company-produced transparency reports. The private sector should participate under conditions of declared interests and proportionate influence. Companies should not sit on the same structural footing as communities whose information environments they shape. The Dialogue should adopt a tiered structure — a high-level political track for commitments, and substantive technical-policy tracks where detailed framework work happens with broader participation. Outcomes from the technical tracks should feed directly into the political track, rather than operating in parallel with no connection. Regional consultations should be deliberative and not just extractive, findings need to make their way into the agenda.

The most significant absences in current global AI governance discussions are not accidental, they reflect the same structural inequalities that AI systems themselves reproduce. Communities speaking lower-resource languages have almost no representation in the bodies designing the training data standards, transparency requirements, and cultural representation benchmarks that will determine whether AI serves or marginalizes them. Inclusion requires more than translation. Governance pprocesses must treat linguistic diversity as a design requirement, with dedicated tracks for assessing AI impacts on non-dominant language communities. Indigenous peoples and knowledge communities face a specific harm because their cultural knowledge, oral traditions, and community data have in many cases been scraped into training datasets without consent, compensation, or governance participation. They need a formal role in the frameworks governing data provenance, benefit-sharing, and cultural representation as rights-holders with specific claims and not just as "vulnerable communities." Journalists and independent media from the Global South are on the frontlines of AI's information ecosystem impacts, experiencing AI-assisted disinformation, algorithmic suppression of local-language content, and displacement of editorial work. Their operational knowledge of how AI affects journalism practice should be a primary input too. Regulators from low- and middle-income countries are building governance frameworks under resource constraints. This requires advance funding, genuine agenda influence, and follow-up technical support so that engagement translates into capacity rather than presence for its own sake. Inclusion must be structural.

Structured red-teaming sessions in which civil society, researchers, and affected communities are invited to identify the failure modes of proposed governance frameworks before they are adopted, not after. This is standard practice in technical security and maybe we can help it become standard in governance design. Evidence translation tracks that connect independent research directly to policy deliberation. Rather than academic papers circulating in parallel, the Dialogue should commission short, structured evidence briefs with defined word limits and plain-language summaries that feed into specific agenda items. The I4T Network's experience suggests that when research is formatted for decision-makers rather than peer review, uptake is better.

The African Union's AI Continental Strategy and the CELAC regional digital agendas demonstrate that meaningful governance frameworks are being developed outside the Euro-Atlantic axis and that the Dialogue should treat these as sources, not recipients, of governance design. Wikipedia's language-equity model makes explicit the disparity in coverage across language communities and treats it as a governance problem requiring structural intervention. I think this is an interesting template for how AI systems could be required to report on and address representational gaps across languages and cultures. Brazil's approach to platform regulation under the LGPD and ongoing legislative debates on platform accountability illustrates how mid-income countries are building regulatory capacity that combines human rights foundations with operational enforceability - a model directly relevant to AI governance in similar contexts. The I4T Global Knowledge Network's Periodic Table of Platform Regulation offers a comparative methodology for mapping governance dimensions across jurisdictions. Applied to AI governance interoperability, it allows regulators to identify minimum convergence points without requiring full harmonization. This is practically useful for smaller regulators seeking to anchor their frameworks to international standards. All these have in common accountability obligations are binding, measurable, and subject to independent verification, which is key.