Seoul Frontier AI Safety Commitments: The Signatory List and What They Pledged

The Frontier AI Safety Commitments are a set of voluntary pledges that AI companies sign, first announced at the AI Seoul Summit in May 2024. They are not a law — no company can be fined for missing a commitment — but they are the direct precursor to binding requirements like California’s SB 53 and New York’s RAISE Act, which turned the commitments’ central ask (publish a safety framework) into a statutory Frontier AI Framework requirement. As of the UK government’s official list, 20 organisations have signed: the original 16 from the Seoul summit in May 2024, plus 4 more added in February 2025. This guide lists every signatory, explains what they actually committed to, and shows how the voluntary commitment relates to the binding requirement that followed it.

The short answer

Signatories to the Frontier AI Safety Commitments pledge to do three things: identify and assess severe risks across a frontier model’s lifecycle, govern that process internally with clear accountability, and be publicly transparent about it. In practice, the commitment that matters most and is the easiest to check is Outcome 3: each signatory agreed to publish a safety framework — a public document describing how it assesses and manages catastrophic risk — in time for the next major AI summit (the AI Action Summit in Paris, February 2025). That published-framework requirement is exactly what SB 53 and the RAISE Act later made mandatory, with legal teeth, for large developers. See What Is a Frontier AI Framework? for how the statutory version differs from this voluntary one.

The full list of signatories

According to the UK government’s official publication, 20 organisations have signed the Frontier AI Safety Commitments to date.

The original 16 (AI Seoul Summit, May 2024)

• Amazon
• Anthropic
• Cohere
• G42
• Google (Google DeepMind)
• IBM
• Inflection AI
• Meta
• Microsoft
• Mistral AI
• Naver
• OpenAI
• Samsung Electronics
• Technology Innovation Institute
• xAI
• Zhipu.ai

4 added later (February 2025)

• 01.AI
• Magic
• MiniMax
• NVIDIA

This list is drawn from the UK Department for Science, Innovation and Technology’s published record, which was last updated on 7 February 2025 to add the four later signatories. If a company you’re researching isn’t on this list, it hasn’t signed the commitments — signing a lab’s own voluntary safety policy, like a Responsible Scaling Policy, is a separate thing from signing this specific multilateral commitment.

What signatories actually pledged

The commitments document sets out three outcomes, broken into eight underlying commitments:

• Outcome 1 — Risk identification, assessment, and management. Signatories agreed to assess severe risks from their frontier models across the AI lifecycle, before and during deployment; to set thresholds beyond which risk is considered intolerable; to define mitigations that keep risk below those thresholds; to describe what they will do if a threshold is breached and adequate mitigation isn’t possible (up to not deploying, or withdrawing, the model or system); and to keep investing in the science of measuring and managing frontier risk.
• Outcome 2 — Accountability. Signatories agreed to hold themselves accountable for meeting the other commitments through internal governance with clear roles, responsibility, and adequate resourcing — including board-level or equivalent oversight.
• Outcome 3 — Transparency. Signatories agreed to be publicly transparent about their implementation of these commitments, except where doing so would disproportionately increase risk or reveal sensitive commercial information out of proportion to the societal benefit — and, specifically, to publish a safety framework explaining how they will approach each of the commitments above.

Of the three, the safety-framework publication commitment under Outcome 3 is the one most directly checkable from outside the company, and the one regulators borrowed most directly: it is the voluntary precursor to the statutory Frontier AI Framework that SB 53 and the RAISE Act now require of large developers. See Responsible Scaling Policy: What It Is and How the Major Labs Compare for how individual labs’ published frameworks compare to one another.

Voluntary commitment vs. statutory requirement

It’s easy to conflate the Seoul commitments with the SB 53 / RAISE Act framework requirement because both center on the same artifact — a published safety framework — but they are not the same thing, and the difference matters for anyone tracking compliance:

• Basis. The Seoul commitments are a multilateral political pledge with no enforcement mechanism. SB 53 and the RAISE Act are state statutes with legal force.
• Who is covered. The commitments apply only to the 20 organisations that chose to sign. The statutes apply to any “large frontier developer” that meets the law’s revenue or compute threshold, whether or not it ever signed anything in Seoul.
• What’s required. The commitments ask signatories to publish a framework describing their approach, in general terms set by the signatories themselves. SB 53 and the RAISE Act specify what the framework must address — capability thresholds, mitigations, third-party assessment, incident response, cybersecurity for model weights — and require it to be kept up to date.
• Consequence of non-compliance. There is none, formally, for the voluntary commitments beyond reputational pressure. Non-compliance with SB 53 or the RAISE Act carries statutory penalties.

Several of the Seoul signatories (Amazon, Anthropic, Google, Meta, Microsoft, OpenAI, xAI, and others among the 20) are also the companies most likely to meet the revenue thresholds that make SB 53 and the RAISE Act’s Frontier AI Framework requirement mandatory rather than voluntary for them. For what that statutory document has to contain, see What Is a Frontier AI Framework? The SB 53 and RAISE Act Requirement, Explained.

Where NIKOLAI fits in

Because the Seoul commitments, individual labs’ voluntary safety policies, and the SB 53 / RAISE Act statutory requirement all use overlapping but not identical language for the same underlying ideas — risk thresholds, mitigations, incident types, governance — comparing what any one signatory has actually published against what the commitments (or the statutes) ask for is harder than it should be. CASRAI’s NIKOLAI dictionary exists for exactly this kind of terminology gap: its N9 (Commitments & Governance) track maps voluntary commitments like these against labs’ published frameworks and against statutory requirements, so the same underlying obligation can be recognized across documents that describe it differently. NIKOLAI’s crosswalks are independent, unendorsed “shadow mappings,” not an official record of compliance.

FAQ

How many companies have signed the Frontier AI Safety Commitments?

20, according to the UK government’s official published list: the original 16 from the AI Seoul Summit in May 2024, plus 4 added on 7 February 2025 (01.AI, Magic, MiniMax, and NVIDIA).

Is signing the commitments the same as complying with SB 53 or the RAISE Act?

No. The commitments are voluntary and have no enforcement mechanism. SB 53 and the RAISE Act are state statutes that separately require any covered “large frontier developer” — signatory or not — to publish a Frontier AI Framework meeting specific statutory content requirements, with legal consequences for non-compliance.

What did signatories actually promise to do?

Three things: assess and manage severe risk across their model’s lifecycle, govern that process internally with clear accountability, and publish a safety framework explaining their approach so outsiders can check it.

Do all 20 signatories have a published safety framework?

Not necessarily as of any given date — the commitment set a target of publishing ahead of the February 2025 AI Action Summit, but whether a specific company’s framework is current, comprehensive, or even publicly posted varies by company and changes over time. Check the individual developer’s own site for its current framework rather than assuming signatory status means a framework is live.

Is Google DeepMind a separate signatory from Google?

No. The official list names the signatory as Google; Google DeepMind is the unit within Google that develops its frontier models and that the commitment and framework publication apply to in practice.