UN: The first Global Dialogue on AI Governance opens in Geneva

UN: The first Global Dialogue on AI Governance opens in Geneva

🔎 Geneva becomes the center of the world again, this time for AI

On July 6 and 7, 2026, Geneva's Palexpo will host an unprecedented event: the first UN-organized Global Dialogue on AI Governance. The initiative, jointly led by the International Telecommunication Union (ITU) and UNESCO, brings together member states, private sector giants, and civil society on the same stage for the first time.

The timing is not coincidental. In eighteen months, the global regulatory landscape has exploded into a mosaic of incompatible national frameworks. Europe is applying its AI Act, the United States is moving forward with voluntary standards for AI models, and China is legislating at breakneck speed on recommendation algorithms and foundational models. The result: a diplomatic gap that the UN is trying to fill before the fractures become irreparable.

The stakes go beyond simple diplomacy. The models dominating current benchmarks — OpenAI's GPT-5.5 (98.2 in agentic, 91 in general), Google's Gemini 3 Pro Deep Think (95.4/90), Anthropic's Claude Opus 4.7 (94.3/90) — are reaching performance levels that make governance urgent, not optional. When Anthropic calls for a global AI pause, and 80% of the code generated in certain sectors already comes from Claude, the signal is clear: technology has gotten ahead of institutions.

This Geneva dialogue is not just another summit. It is the UN's first concrete attempt to produce an operational multilateral framework, supported by a multidisciplinary annual report presented in plenary session. No empty declarations: the goal is to define standards, monitoring mechanisms, and minimum obligations.

The essentials

• The first UN Global Dialogue on AI Governance will be held on July 6-7, 2026 at Palexpo in Geneva, co-organized by the ITU and UNESCO.
• The event brings together member states, the private sector, and civil society around a multidisciplinary annual report on the state of AI governance.
• The context is marked by a global regulatory fragmentation and the arrival of GPT-5.5 and Claude Opus 4.7 level models, making inaction costly.
• The goal is not an immediate binding treaty, but the creation of a sustainable international coordination framework with monitoring mechanisms.

Recommended tools

The unprecedented format of the Dialogue: three pillars, one plenary session

The Global Dialogue does not follow the usual pattern of UN summits. According to the official UN communique, the structure is based on three distinct but interconnected pillars, each with its own working sessions before a summary plenary.

The first pillar brings together the Member States. Not just the usual G7 members — whose Evian summit brought together Altman, Amodei and Hassabis for the first time — but all 193 member countries. The ITU insists on the representativeness of the Global South, which is often absent from discussions on technical standards.

The second pillar is dedicated to the private sector. OpenAI, Google, Anthropic, xAI (Grok 4.1, 90 in general), DeepSeek (V4 Pro Max, 88) and Moonshot AI (Kimi K2.6, 84) are invited to present their internal safety practices. The format imposes cross-discussions with regulators, not marketing keynotes.

The third pillar mobilizes civil society: researchers, NGOs, digital rights organizations. It is this pillar that feeds into the annual multidisciplinary report, presented in plenary session on July 7. This report, coordinated by UNESCO according to its own presentation of the event, covers the ethical, legal, economic, and societal dimensions of AI.

The annual report is the real innovation of the framework. Unlike the final declarations of summets that end up gathering dust on a shelf, this document is designed to be updated every year with measurable indicators. It will serve as a reference baseline for future negotiations.

Why the UN, why now

The legitimate question: why is the UN stepping in now, when the ITU and UNESCO are already working separately on these issues?

The answer lies in the gap between existing frameworks. The ITU published its AI standardization framework in 2024. UNESCO has had its Recommendation on the Ethics of AI since 2021. But neither has the political legitimacy to rally states around common obligations. The Global Dialogue is precisely designed to build this bridge between technical standardization (ITU), ethical principles (UNESCO), and political negotiation (General Assembly).

The ITU communiqué from June 2026 sets out the context unambiguously: "Without international coordination, we risk a regulatory patchwork that hinders innovation in developing countries while leaving exploitable gaps in the most regulated countries."

The figures back up this urgency. In 2025, the global AI market exceeded $500 billion according to consolidated estimates. Investments in foundational models have tripled in two years. Yet, fewer than 30% of countries have a national framework specific to AI. The Geneva dialogue aims to close this governance gap before massive adoption renders any retroactive framework ineffective.

A significant detail: the choice of Geneva rather than New York. The city already hosts the ITU, the WHO, the CERN, and dozens of international organizations. It is a technical diplomatic ecosystem, not a political one. The message is clear: this dialogue is primarily a matter of expertise, not posturing.

AI models at the heart of the negotiations

AI is not governed in the abstract. The Geneva negotiations are directly informed by the reality of the models deployed in 2026, and the agentic benchmark scores are on everyone's lips.

GPT-5.5 dominates with a score of 98.2 in agentic tasks, followed by Gemini 3 Pro Deep Think (95.4) and Claude Opus 4.7 Adaptive (94.3). These three models can now plan, execute, and correct complex task chains autonomously. The leap from the previous generation (GPT-5, 78.1) is dizzying.

In the generalist category, the hierarchy is tighter: Gemini 3.1 Pro (92), GPT-5.5 (91), GPT-5.4 Pro (91), Claude Opus 4.7 (90), Gemini 3 Pro Deep Think (90), and Grok 4.1 (90) are clustered within less than two points of each other. This convergence means that no single player has a sufficient technological monopoly to dictate the rules alone.

The presence of open-source or self-hosted models in the agentic leaderboard is also a subject of negotiation. Kimi K2.6 (88.1, self-host) and GLM-5 Reasoning (82, self-host) show that agentic autonomy capability is no longer the exclusive domain of American giants. How do you govern models that anyone can deploy on their own server? This is one of the thorniest questions in the dialogue.

Research on multimodal dialogue systems is also fueling the debates. The study Enhancing Consistency in Multimodal Dialogue System Using LLM with Dialogue Scenario shows that the consistency of dialogue systems remains a major technical challenge, with direct implications for the safety of mainstream deployments. Furthermore, the Dialogue Robot Competition 2023 demonstrated the current limitations of conversational systems integrated into humanoid robots, a domain where governance is virtually non-existent.

The three negotiating blocks

Model security and evaluation

The first negotiating block focuses on the evaluation and certification mechanisms for foundational models. The idea is not to create a new technical standard — the ITU handles that — but to define what level of transparency developers must provide to regulators.

The United States, with its voluntary standards, is pushing for an enhanced self-assessment system. The EU wants to export the prior notification mechanism from its AI Act. Global South countries are demanding access to evaluation results, not just reports written by companies.

The tension is real. When Claude Opus 4.7 or GPT-5.5 are deployed via API in 150 countries, who has jurisdiction? The country of incorporation, the country where the servers are hosted, or the country of the end user? The UNESCO report proposes a "subsidiary competence" mechanism that remains to be negotiated.

Copyright and training data

The second block concerns the legal regime of training data. This is the most contentious issue, as it directly affects the business models of developers.

In its report, UNESCO documents the data licensing practices used by major labs. The finding is unequivocal: the majority of current models have been trained on corpora whose legal status remains unclear. The negotiations are attempting to define a minimum baseline of due diligence that every developer should comply with, regardless of their national jurisdiction.

Inclusivity and equitable access

The third block is driven by developing countries. The argument is simple: models like DeepSeek V4 Pro Max (88 in general) and Kimi K2.6 (84) prove that AI innovation is no longer the exclusive preserve of Silicon Valley. Yet, the rules of the game — benchmarks, safety standards, transparency requirements — are still defined by a small number of players.

The dialogue aims to create mechanisms for equitable participation in standardization bodies. Concretely, this could take the form of funding so that researchers from the Global South can participate in the ITU's work, or open data pools for training in underrepresented languages.

Dialogue robots: what research tells us about current limitations

AI governance is not just about LLMs. Dialogue systems integrated into physical robots pose specific regulatory challenges that the Geneva Dialogue is only beginning to touch upon.

The study Dialogue system with humanoid robot documents the architectures used to integrate language models into robotic platforms. The central issue: the feedback loop between the physical world and the language model creates risks that do not exist in a pure chatbot. A robot that misinterprets a gesture can make irreversible physical decisions.

Research on Personality-adapted multimodal dialogue systems adds another layer of complexity. When a dialogue system adapts its personality in real time based on the user, who defines the boundaries of this adaptation? Current regulatory frameworks, designed for static systems, are outdated.

The annotation scheme for Dependency Dialogue Acts, however, offers an interesting path for governance. By formally structuring dialogue acts (question, assertion, request, etc.) and their dependencies, a common language is created to audit the behavior of conversational systems. This type of standardized metric is exactly what the ITU is looking to integrate into its standards.

This research, although prior to the Dialogue, is cited in UNESCO's annual report as evidence that the scientific community has already identified problems that regulators are slow to address. The gap between academic research and legislation is one of the driving forces behind the Global Dialogue.

What Geneva will not resolve

To be honest: the first Global Dialogue on AI Governance will not produce a binding treaty. That is not its goal, and claiming otherwise would be dishonest.

What Geneva can produce is a common framework. A shared language among regulators, developers, and researchers. Measurable indicators for the annual report. Political commitments that, even if non-binding, create sufficient reputational pressure to influence behavior.

The parallel with the Paris climate agreements is instructive. In 2015, no one thought that NDCs (nationally determined contributions) would be enough to solve global warming. But they created an architecture of transparency and peer pressure that transformed the political landscape. The Geneva dialogue aims for the same effect for AI.

The limitations, however, are real. The UN has no audit power over private labs. The annual report will depend on the voluntary cooperation of companies. And self-hosted models like GLM-5 Reasoning or Kimi K2.6 can be deployed without any state knowing about it.

The position of key players

OpenAI and the calculated transparency posture

OpenAI arrives in Geneva from a position of strength. GPT-5.5 (98.2 agentic, 91 general) and GPT-5.4 Pro (91.8/91) dominate the benchmarks. OpenAI's strategy is clear: accept moderate transparency obligations in exchange for institutional legitimacy that hinders competitors.

OpenAI's business model relies on the API. The more transparency regulators require, the higher the barrier to entry for new players becomes. This is a rational line of reasoning, and UN negotiators are aware of it.

Anthropic and the advocate of caution

Anthropic holds a unique position. On the one hand, its CEO called for a global pause at a time when Claude was generating 80% of the code in certain contexts. On the other hand, Claude Opus 4.7 Adaptive (94.3/90) and Claude Sonnet 4.6 (81.4/83) are among the most powerful models on the market.

Anthropic's posture in Geneva is that of the responsible laboratory asking for rules for everyone. This is consistent with their safety approach (Constitutional AI), but it is also a competitive strategy: strict rules penalize less disciplined players to a greater extent.

Google and the standards-based approach

Google, with Gemini 3 Pro Deep Think (95.4/90) and Gemini 3.1 Pro (92), is betting on technical standardization. The company is pushing for the dialogue to result in ITU standards rather than UN legal obligations. The distinction is important: a standard is adopted by technical consensus, while a legal obligation is imposed by a treaty.

Emerging players: DeepSeek, Moonshot AI, xAI

DeepSeek (V4 Pro Max, 88) and Moonshot AI (Kimi K2.6, 88.1 agentic) represent the new guard. Their presence in Geneva is symbolic: they show that AI governance can no longer be a bilateral US-China dialogue. xAI with Grok 4.1 (90 general) adds further complexity with a model that ranks among the best but comes with a less conventional regulatory approach.

Autonomous agents: the taboo topic of the dialogue

There is an elephant in the room at Palexpo, and nobody really wants to talk about it: autonomous AI agents. Yet, this is precisely the category where scores have exploded.

GPT-5.5 at 98.2 in agentic means that the model can execute complex task chains with near-human reliability. Claude Opus 4.7 Adaptive at 94.3, GPT-5.4 Pro at 91.8: these figures indicate that we have crossed a qualitative threshold. For those who want to understand the practical implications, our guide to creating your first autonomous AI agent details the typical architecture of these systems.

The governance issue is radical. A classic LLM answers a question. An autonomous agent decides to act, chooses the tools, executes, evaluates the result, and iterates. The chain of responsibility becomes blurred: who is responsible if an agent deployed by a French company, powered by an American model, executes an illegal action on a server located in India?

The UNESCO report mentions autonomous agents, but without proposing a specific framework. Negotiators favor a use-case approach rather than one based on the nature of the system. This is pragmatic, but ultimately insufficient.

Multimodality and avatars: blurring boundaries

Another area where governance lags behind technology: multimodal systems and AI avatars. Current models are no longer simple text generators. They see, hear, speak, and sometimes take physical form.

Research on multimodal dialogue systems, particularly the study on enhancing consistency in multimodal systems, shows that the fusion of modalities creates new risk vectors. An avatar that combines voice generation, vision, and adaptive personality (as documented in the study on personality-adapted multimodal dialogue systems) can manipulate a user much more effectively than a text-based chatbot.

For practitioners, creating an AI avatar has become trivial — our tutorial for creating your first AI avatar in 10 minutes is proof of this. But from a regulatory perspective, every deployed avatar is an unregulated persuasion system. The Geneva dialogue is starting to identify the problem, but solutions remain in draft form.

The issue of imitating a human being is particularly sensitive. When an avatar reproduces the face, voice, and speech tics of a real person, current legal frameworks (image rights, personality rights) are insufficient because they are territorial, whereas avatars are deployed globally.

Tabular Data and Decisional AI: An Underestimated Angle

All the media spotlight is on LLMs and agents, but an entire area of AI largely escapes public debate: models on tabular data. Yet, this is the domain that has the most concrete impact on the lives of citizens.

Tabular models are the ones that decide on bank loans, medical diagnoses, university admissions, and judicial sentences. They are less sexy than GPT-5.5, but potentially more dangerous because their decisions are applied directly without a human filter.

The model TabPFN, the first foundation model for tabular data, represents a turning point in this field. By bringing the foundation model approach to structured data, it makes possible the massive deployment of decisional AI with very little specific training data. This is a major technical advance, but also a regulatory headache.

The UNESCO report dedicates a section to decisional AI systems, but the Geneva negotiations remain dominated by LLM issues. African and Asian countries, where tabular AI has the most concrete applications (agriculture, healthcare, microfinance), are trying to put this topic back on the table. Their argument is irrefutable: governing AI without governing decisional systems means regulating the storefront and ignoring the server room.

❌ Common mistakes

Mistake 1: Confusing the Dialogue with a binding treaty

What's wrong: many commentators present the Global Dialogue as the equivalent of a Paris Agreement for AI. This is false. The dialogue produces a reference framework and an annual report, not a treaty with sanctions.
The solution: read it as a first step in a long-term process. The first Earth Summit (1972) did not stop global warming, but it created the architecture that led to the Paris Agreement 43 years later.

Mistake 2: Thinking that the US and the EU will dictate the rules

What's wrong: extrapolating the current dominance of American and European models (GPT-5.5, Claude Opus 4.7, Gemini 3) toward regulatory dominance. The dialogue is designed precisely to avoid this scenario.
The solution: follow the positions of Global South countries and emerging actors like DeepSeek and Moonshot AI. Their votes count just as much as those of the United States in the General Assembly.

Mistake 3: Ignoring the technical dimension of the negotiations

What's wrong: treating AI governance as a purely legal or ethical subject. Dependency Dialogue Acts annotation schemas, multimodal coherence metrics, agentic benchmarks: all of this is the technical substrate on which the rules rely.
The solution: inform yourself about ITU standards and the research work cited in the annual report. Governance without technical understanding is empty bureaucracy.

❓ Frequently Asked Questions

Who exactly participates in the Global Dialogue?

The 193 UN Member States, private sector representatives (AI labs, companies deploying AI systems), civil society organizations (NGOs, universities, trade unions), and specialized institutions like the ITU and UNESCO. The official list of participants is available on the UN website.

Will the annual report be public?

Yes. UNESCO has confirmed that the multidisciplinary report will be published in open access after its presentation in plenary on July 7. It will include numerical data, case studies, and recommendations updated every year.

Does this dialogue replace the European AI Act?

No. The Global Dialogue is complementary to national and regional frameworks. It aims to create a layer of international coordination, not to replace existing legislation. The AI Act continues to apply in the EU.

Are open-source models like Kimi K2.6 and GLM-5 affected?

Yes, it is even one of the most debated topics. Self-hosted models pose a specific governance challenge because they can be deployed without any point of contact with a regulator. The dialogue explores governance mechanisms through infrastructures (hosting providers, cloud providers) rather than through the models themselves.

What is the connection with the G7 summit in Évian?

The Évian summit created political momentum among the seven most industrialized countries. The Geneva dialogue broadens this momentum to the entire international community. The conclusions of Évian feed into the negotiations, but do not bind them.

✅ Conclusion

The UN's first Global Dialogue on AI Governance will not change the world in two days. But it marks the moment when AI governance moves from the stage of statements of intent to that of concrete institutional architecture. Between the dizzying scores of agentic models, the global regulatory fragmentation, and the growing influence of emerging players, inaction is no longer an option. The annual report to be presented in Geneva on July 7, 2026, will be the first reliable barometer of our collective capacity to govern what we have created. Follow the developments on the official dialogue website and the ITU page.