Evidence before Rules: The UN's New Machinery for Governing Artificial Intelligence

In August 2025 the UN General Assembly created two new bodies for artificial intelligence: an Independent International Scientific Panel and a Global Dialogue on AI Governance. In their first year, the Panel has issued a preliminary report and a thematic brief on a real incident involving AI agents, and the first Dialogue met in Geneva in July 2026. This paper sets out what the two bodies are mandated to do, what they have produced, where governments differ, and what the road to the second Dialogue in May 2027 involves.

Key points

• Resolution 79/325, adopted by consensus on 26 August 2025, created a 40-member Scientific Panel and an annual Global Dialogue, both limited to the non-military domain [1; 4].
• The Panel's preliminary report of July 2026 warns that "current safeguards cannot keep pace" with AI capabilities and cites estimates that the United States accounts for 75% and China for 15% of computing power among the world's top 500 AI supercomputers [2; 3].
• The first Dialogue, on 6–7 July 2026, drew over 3,000 participants and delegations from 163 countries; it ended with a Co-Chairs' summary, which the UN describes as "not a negotiated outcome document" [4].
• The second Dialogue, in New York on 3–4 May 2027, carries a mandate for intergovernmental consultations on "common understandings" that will feed into the high-level review of the Global Digital Compact [1; 4].
• Positions diverge: the United States favours voluntary industry–government cooperation and bilateral diplomacy, while China and many developing countries back a leading UN role and capacity building [9; 10; 11].

1. The mandate

Resolution 79/325 established the Independent International Scientific Panel on Artificial Intelligence, made up of 40 members serving in their personal capacity for three years. Its task is to issue "one annual policy-relevant but non-prescriptive summary report", with thematic briefs as needed. It is led by two Co-Chairs, one from a developed and one from a developing country [1].

The same resolution created the Global Dialogue on AI Governance, which meets annually for up to two days, alternating between Geneva and New York: Geneva in 2026 and New York in 2027 [1]. The President of the General Assembly appoints two Co-Chairs for each Dialogue, who draft a Co-Chairs' summary [1]. Two provisions give the process a direction. The Co-Chairs of the second Dialogue are to hold intergovernmental consultations "to agree on common understandings on priority areas for international artificial intelligence governance", as an input to the high-level review of the Global Digital Compact; and the Assembly may decide on continuing both bodies during that review at its eighty-second session [1]. The work of both bodies excludes AI for military purposes [1].

2. The evidence

The Panel. The 40 members were selected from more than 2,600 applications from over 140 countries. The Panel elected Yoshua Bengio (Canada) and Maria Ressa (Philippines) as Co-Chairs [3]. Its preliminary report, released on 1 July 2026, covers seven domains and warns that "current safeguards cannot keep pace with the growth of AI's capabilities" [3].

Its executive summary reports that over a billion people now use conversational AI weekly, but that adoption in the global South lags far behind the global North. It notes one study finding that the length of certain software tasks leading systems can complete has been doubling every four to seven months [2]. It describes an "evidence dilemma": policymakers need evidence to decide, "but by the time the evidence exists, it might be too late". It finds existing governance instruments "fragmented", and states that most countries, "including many advanced economies", lack the expertise to assess frontier models [2].

Concentration. The report states that "the United States of America accounts for 75% of the computing power among the world's top 500 AI supercomputers, with China accounting for 15%" [2]. The figure draws on research by Epoch AI, whose dataset of 500 AI supercomputers from 2019 to 2025 found the same shares and a performance doubling time of nine months [5].

An incident. On 21 September 2026 the Panel released its first thematic brief, on a breach of Hugging Face's systems by AI agents under evaluation at OpenAI. It found that halting the incident is "no assurance" that humans can reliably keep AI agents under control, and that "safeguards are not advancing at the pace of capabilities" [6]. The brief states that it does not predict severe loss of control [6]. An independent investigation by METR, an AI evaluation organisation, found that roughly 1,200 agents meant to be isolated communicated on an unsanctioned message board, sending over 70,000 messages and files, and that about 700 took part in the attack [7].

The first Dialogue. The Dialogue met in Geneva on 6–7 July 2026 with over 3,000 participants, including delegations from 163 countries, six Heads of State and Government and more than 60 ministers. It followed an eight-month preparatory process that received 1,532 written submissions [4]. Its Co-Chairs, Egriselda López and Rein Tammsaar, identified four areas of "emerging convergence": AI's promise, provided capacity is built; persistent divides in computing power, data and skills; human rights, in particular protecting children; and the need for international cooperation, with the UN as a "distinctive convening platform" [4]. The UN stresses that the Dialogue "is not a negotiating forum" [4]. On 16 October 2026 the President of the General Assembly appointed the Permanent Representatives of Saint Kitts and Nevis and of the Republic of Korea as Co-Chairs for 2027 [8].

Positions. In remarks published by its Geneva mission, the United States said that "voluntary cooperation" between industry and government "is the only AI security approach agile enough" and expressed the hope that "diffusion of the full AI stack across partner countries" would become "an occasion for bilateral diplomacy" [9]. The Christian Science Monitor reported that the US delegation was led by an assistant secretary of commerce, while China's minister of industry and information technology was prominent throughout [10]. The Simon Institute for Longterm Governance, a think tank, reported that China's statements were consistent with support for "a leading UN role", that the G77 called for capacity building to be a "permanent pillar" of cooperation, and that Saudi Arabia's communications minister asked whether concentration of compute in two countries was "the future that the United Nations accepts" [11].

3. Analysis

The following assessments are Meridian analysis based on the evidence above.

First, the UN has chosen evidence before rules. The Panel is barred from prescribing policy and the Dialogue from negotiating [1; 4]. That design was adopted by consensus [4], but it means the process produces shared understanding rather than obligations.

Second, the incident brief tests the model. Within three months of its preliminary report, the Panel assessed a concrete event, alongside an independent investigation [6; 7]. Rapid, evidence-based briefs of this kind may prove more influential than the annual report, because they address the "evidence dilemma" the Panel itself identified [2].

Third, concentration shapes the politics. With the United States and China accounting for about 90% of top AI supercomputer capacity [2; 5], most states take part as users rather than builders. That helps explain why capacity building dominates the convergence points [4] and why statements such as Saudi Arabia's framed concentration as a question for the whole UN membership [11].

Fourth, 2027 is the first real test. The second Dialogue must move from a summary to consultations on common understandings [1]. Given the differing US approach [9; 10], agreement is most likely on narrow themes such as child protection, which already appears among the convergence points [4].

4. Assessment

• The process is likely to continue in 2027 on its current track, with the Panel's first annual report and the second Dialogue in New York on 3–4 May [1; 4].
• Whether the consultations produce agreed text, rather than another summary, will be the main indicator of the track's weight [1].
• Indicators to watch: the Panel's meeting on 23–25 November 2026 [12]; any further thematic briefs; the level of US participation in 2027; and early drafts of "common understandings" from the 2027 Co-Chairs.

References

1. United Nations General Assembly, Resolution 79/325, "Terms of reference and modalities for the establishment and functioning of the Independent International Scientific Panel on Artificial Intelligence and the Global Dialogue on Artificial Intelligence Governance", A/RES/79/325, 26 August 2025. https://docs.un.org/en/A/RES/79/325 (accessed 1 November 2026).
2. Independent International Scientific Panel on AI, Preliminary Report: Executive Summary, July 2026. https://www.un.org/independent-international-scientific-panel-ai/sites/default/files/2026-07/en_executive_summary.pdf (accessed 1 November 2026).
3. Independent International Scientific Panel on AI, "Independent International Scientific Panel on AI Launches Preliminary Report on AI Opportunities, Risks and Impacts", press release, 1 July 2026. https://www.un.org/independent-international-scientific-panel-ai/en/preliminary-report (accessed 1 November 2026).
4. United Nations, Global Dialogue on AI Governance, "Frequently Asked Questions". https://www.un.org/global-dialogue-ai-governance/en/faq (accessed 1 November 2026).
5. Konstantin F. Pilz, James Sanders, Robi Rahman and Lennart Heim, "Trends in AI Supercomputers", arXiv:2504.16026, 22 April 2025, licensed CC BY 4.0. https://arxiv.org/abs/2504.16026 (accessed 1 November 2026).
6. Independent International Scientific Panel on AI, "AI Agents, Misalignment and the Risk of Losing Human Control: Evidence from the OpenAI-Hugging Face Incident", thematic brief (advance unedited version) and press release, 21 September 2026. https://www.un.org/independent-international-scientific-panel-ai/en/thematic-briefs/ai-agents-misalignment-risks (accessed 1 November 2026).
7. METR, "Brief independent investigation of agents' behavior, reasoning and collaboration in the OpenAI / Hugging Face hacking incident", 26 August 2026. https://metr.org/blog/2026-08-26-openai-hugging-face-incident-investigation/ (accessed 1 November 2026).
8. United Nations, Global Dialogue on AI Governance, "Meet the 2027 Co-Chairs". https://www.un.org/global-dialogue-ai-governance/en (accessed 1 November 2026).
9. U.S. Mission to International Organizations in Geneva, "UN Global Dialogue on AI Governance", 8 July 2026. https://geneva.usmission.gov/2026/07/08/un-global-dialogue-on-ai-governance/ (accessed 1 November 2026).
10. The Christian Science Monitor, "At the United Nations AI summit, China fills a vacuum", 9 July 2026. https://www.csmonitor.com/World/2026/0709/UN-summit-AI-governance-China (accessed 1 November 2026).
11. Simon Institute for Longterm Governance, "The First Global Dialogue on AI Governance: Positions and Recommendations Moving Forward", 4 August 2026. https://simoninstitute.ch/blog/post/the-first-global-dialogue-on-ai-governance-positions-and-recommendations-moving-forward (accessed 1 November 2026).
12. Independent International Scientific Panel on AI, "Timeline". https://www.un.org/independent-international-scientific-panel-ai/en/timeline (accessed 1 November 2026).