Seoul AI Safety Commitments: Signatories, Frameworks & Deadline

Sixteen labs sign, more follow

On May 21, 2024, the UK and South Korean governments published the Frontier AI Safety Commitments, agreed at the AI Seoul Summit by sixteen companies: Amazon, Anthropic, Cohere, G42, Google, IBM, Inflection AI, Meta, Microsoft, Mistral AI, Naver, OpenAI, Samsung Electronics, Technology Innovation Institute, xAI and Zhipu AI. The document was updated on February 7, 2025 to record further organizations that had since agreed to the same text, including Magic, MiniMax, 01.AI and NVIDIA. [2]

Atlas interpretation: The growing signatory list is the point of the exercise as much as the original sixteen. A voluntary pledge with no enforcement mechanism draws whatever force it has from how many relevant companies keep opting in after the cameras leave, and the fact that the document is still being amended eight months on is the closest thing to evidence that it did not just fade with the summit. [2]

Three outcomes, one deadline

The commitments specify three outcomes. Signatories agree to assess frontier AI risks across the development lifecycle, set severity thresholds for when a model's risk is judged intolerable, put mitigations in place, and commit to not developing or deploying a model at all if those risks cannot be adequately mitigated. They agree to keep internal governance and accountability structures, with clear roles and resources assigned to safety. And they agree to be publicly transparent about how they are implementing all of this, short of disclosing commercially sensitive material, while sharing more detailed information with governments as appropriate. [2]

The one concrete deadline in the text: each signatory commits to publish a safety framework describing how it is meeting these outcomes by the next AI summit, then scheduled to be hosted by France. [2]

Atlas interpretation: What the pledge actually binds a company to is narrower than the outcome language suggests. Each signatory sets its own severity thresholds and its own definition of adequate mitigation, and no external body checks the framework against the thresholds or the thresholds against the risk. That is a disclosure commitment dressed in the language of a safety standard, and it is the same shape a domestic law later took in a jurisdiction that could actually enforce one: publish a framework, report incidents, no mandated technical bar. [2]

Between Bletchley and Paris

The Seoul summit was the second in the series the UK started at Bletchley Park in November 2023, this time co-hosted with South Korea. Alongside the corporate commitments, 27 countries and the EU adopted a separate Seoul Ministerial Statement, ten governments signed a Seoul Declaration, the first international network of national AI Safety Institutes was announced, and the UK's own AI Safety Institute opened its first overseas office, in San Francisco. [3]

Google DeepMind published its own Frontier Safety Framework, defining Critical Capability Levels across autonomy, biosecurity, cybersecurity and machine learning research and development, on May 17, 2024, four days before the Seoul commitments were signed. [4]

Atlas interpretation: That timing is a useful data point against reading the Seoul deadline as the thing that produced the frameworks. At least one signatory's framework was already finished before the pledge existed, which suggests the commitment mostly formalized safety work already underway at the frontier labs rather than starting it, and that the real test of the pledge was whether the labs without a framework in hand would produce one under the same voluntary pressure. [4][2]