As part of its Global Dialogue on AI Governance, the United Nations General Assembly invited organizations around the world to say which AI governance measures they would like to see put in place. 1,534 organizations and individuals answered: governments, companies, universities, civil society, technical bodies and international organizations, from every region of the world. We analyzed the responses to this consultation.
The consultation reveals widely shared priorities. Taken together, a large share of them sketch the outline of an international AI governance regime built on binding risk thresholds for the most advanced systems.
We built an explorer that lets you browse every submission and trace the exact quotes our figures rest on: aidialoguereport.org
1. Safety leads the agenda — by participants' own choice
Before drafting their response, each contributor was asked to indicate which AI governance theme(s) — drawn from Resolution 79/325 — their submission addressed. Here is how often each theme was selected by participants:
- 73% — Safe, secure and trustworthy AI
- 71% — Transparency, accountability and human oversight
- 61% — Social, economic, ethical, cultural, linguistic and technical implications
- 53% — AI capacity-building
- 49% — Protection and promotion of human rights
- 44% — Interoperability of governance approaches
- 23% — Open-source software, open data and open AI models
Safety comes first and oversight right behind it, each cited by more than 70% of submissions and well ahead of the other themes the UN put forward. In other words, even before getting into the substance of their proposals, most participants placed the management of risk at the heart of their concerns.
2. Governing AI is no longer up for debate
Nine submissions out of ten propose at least one governance measure (1,358 of 1,534): audits, technical standards, regulation, monitoring and incident reporting, the creation of institutions, international coordination mechanisms… Adding capacity-building and infrastructure brings the figure to 94% (1,442 submissions); only 92 — 6% — propose nothing of the sort.
Granted, the form explicitly invited governance proposals, so a high figure was to be expected. It is nonetheless striking how few responses argued for the status quo or for governing AI more lightly.
3. A majority wants us to be better equipped to assess, monitor and rein in AI
The measures proposed most often are the ones that make rules enforceable:
- 62% propose accountability and control mechanisms: testing, auditing and evaluation (47%); monitoring and incident reporting (27%); technical and organizational controls (33%)
- 60% propose an international architecture: coordination and networking mechanisms (54%), new institutions (10%) or a binding international agreement (7%)
These two families of measures are not mutually exclusive — they are, to a large extent, complementary. An evaluation that leads to no common rule does nothing to advance safety; and because AI risks are cross-border by nature, no rule would hold for frontier systems without international coordination.
This complementarity shows through in the submissions themselves. While each author tends to describe the part of the system closest to their own immediate concerns, 53% of them combine at least two of these levers – accountability, international coordination, common rules – and two-thirds (66%) call for common rules in one form or another: standards, regulation or an international agreement.
4. Binding instruments outweigh voluntary ones
When submissions point to legal instruments, they lean toward binding ones more often than voluntary ones: half of them (50%) invoke a binding rule — national and regional regulation (47%) or a binding international agreement (7%) — against 39% for non-binding norms and declarations.
This holds across every category of respondent: between 46% and 52% of each group's submissions back binding law, industry included. Where they differ is on the kind of regulation they want. Looking more closely at the submissions that favour regulation, 71% of companies want it to be proportionate to risk, with obligations that grow stricter the more dangerous a system is. Civil society, for its part, places more emphasis on rules being enforced in practice (44%) and on the protection of rights.
5. A panorama of all the measures proposed
Here is the full distribution across all fifteen measure categories, as coded:
- 58% — Capacity-building and technical assistance
- 54% — Coordination and networking mechanisms
- 47% — Testing, auditing and evaluation
- 47% — National and regional regulation
- 43% — Technical standards and benchmarks
- 39% — Non-binding norms and declarations
- 33% — Infrastructure and resource provision
- 27% — Monitoring, incident reporting and registries
- 21% — Technical safety and security controls
- 17% — Organizational governance and internal controls
- 17% — Regulatory sandboxes and pilots
- 10% — Institutional creation
- 7% — Binding international agreement
- 6% — Public procurement and market-shaping
- 4% — Fiscal instruments
Two families dominate, and they complement each other: building capacity (capacity-building, infrastructure) and building oversight (evaluation, regulation, standards). The corpus asks for both at once.
Methodology
The analysis covers all 1,534 written submissions published by the UN Global Dialogue on AI Governance.
The categories were fixed before the corpus was read: 31 risk codes across 9 domains, most drawn from the MIT AI Risk Repository [1], and 15 forms a governance measure can take, adapted from OECD and academic taxonomies of policy instruments [2].
Each submission was then annotated independently by three language models from three different developers, which reduces the chance that they make the same mistakes. A model cannot simply assert an annotation: each one must rest on a word-for-word quote from the submission, and software checks that the quote is really there (this was the case for 99.8% of the 36,843 quotes cited across the corpus). An annotation is accepted outright when the two primary models both back it with verified quotes. Contested codes (52% of the 18,667 decisions) went to a separate arbiter model, Claude Opus 4.8, which upheld 46% of them — each of its decisions likewise having to rest on a quote verified in the text. In all, 13,425 annotations.
To gauge how reliable this process is, a human annotator labelled a sample of 60 submissions. Against that benchmark, the system reaches 94% precision and 93% recall on risks, and 92% precision and 96% recall on governance measures.
Every share reported here is corrected for that measured error rate using a Rogan-Gladen prevalence correction [3]. Differences between stakeholder groups or between regions are reported only when they survive a multiple-comparison correction — the false-discovery-rate correction [4] — applied across every comparison tested; even so, they should be read as exploratory.
Two limits apply to everything above. First, participation was self-selected: these figures describe what the people and organizations who wrote in chose to say, not world opinion; and each submission counts once, whether it came from an individual or a government ministry. Second, absence is only weak evidence: a submission that never mentions a risk has not thereby said the risk does not matter. Every accepted code links back to its quote, and any submission can be audited in the explorer.
Acknowledgements: Félix Dorn and Charbel-Raphaël Segerie led the analysis, writing and implementation. Markov Grey, Arthur Grimonpont and Épiphanie Gédéon contributed input and feedback.
References
- Slattery et al. (2025), The AI Risk Repository: A Comprehensive Meta-Review, Database, and Taxonomy of Risks From Artificial Intelligence, arXiv:2408.12622.
- OECD/EC (2021), STIP Compass Taxonomies Describing STI Policy Data, policy-instruments taxonomy; Maas & Villalobos (2023), International AI Institutions: A Literature Review of Models, Examples, and Proposals, Institute for Law & AI.
- Rogan & Gladen (1978), Estimating Prevalence From the Results of a Screening Test, American Journal of Epidemiology 107(1).
- Benjamini & Hochberg (1995), Controlling the False Discovery Rate: A Practical and Powerful Approach to Multiple Testing, Journal of the Royal Statistical Society, Series B 57(1).