The Frontier AI Safety Commitments are a set of voluntary, non-binding corporate safety commitments adopted by 16 frontier AI organisations at the AI Seoul Summit on 21 May 2024 and updated on 7 February 2025. Co-hosted by the UK Government and the Republic of Korea, the summit produced the commitments as the company-facing counterpart to the state-level Bletchley Declaration, with signatories publicly committing to specific practices on risk assessment, thresholds, mitigation, accountability, and transparency. The document covers frontier AI developers and is structured around three outcomes.
Full title: Frontier AI Safety Commitments, AI Seoul Summit 2024 Issuing bodies: UK Government and Republic of Korea (summit co-hosts); voluntarily adopted by 16 frontier AI organisations Date: 21 May 2024 (updated 7 February 2025) Legal status: Voluntary corporate commitments, non-binding Scope: Frontier AI developers; risk assessment, thresholds, mitigation, accountability, transparency
Status and timeline
The commitments were issued at the AI Seoul Summit, co-hosted by the UK and the Republic of Korea, to continue the process initiated by the Bletchley Declaration. Where the Bletchley Declaration is a state-level political document, the Seoul Commitments are directed at companies, with frontier developers publicly committing to specific safety practices. Sixteen organisations adopted the commitments on 21 May 2024, and the text was updated on 7 February 2025. The official text is published by the UK Department for Science, Innovation and Technology on gov.uk (first published 21 May 2024) (Source: gov.uk).
Signatories undertook a delivery deadline alongside the eight commitments: to demonstrate implementation "by publishing a safety framework focused on severe risks by the upcoming AI Summit in France" — the February 2025 Paris AI Action Summit (Source: gov.uk). The 7 February 2025 update to the gov.uk publication coincided with that deadline, days before the Paris summit opened, as signatories published frontier safety frameworks. The text also provides that organisations' approaches to meeting the three outcomes may evolve given "the evolving state of the science," with signatories committing to provide transparency on such changes, including their reasons, through public updates (Source: gov.uk).
Beyond the eight numbered commitments, signatories affirmed a set of current best practices: internal and external red-teaming for severe and novel threats; information sharing; cybersecurity and insider-threat safeguards to protect unreleased model weights; incentivizing third-party vulnerability discovery and reporting; mechanisms enabling users to identify AI-generated audio and visual content; public reporting of capabilities, limitations, and appropriate and inappropriate use domains; prioritizing research on societal risks; and developing frontier AI to help address global challenges (Source: gov.uk).
Scope and definitions
The commitments apply to frontier AI developers. Two terms are defined:
- "Frontier AI" refers to highly capable general-purpose AI models or systems that match or exceed the capabilities of the most advanced models.
- "Severe risks" are those that could cause serious harm to public safety, human rights, or democratic values at societal scale, and that are difficult to reverse.
Key provisions
The commitments are organised under three outcomes covering eight numbered commitments.
Outcome 1 — Risk identification and management
Commitment I requires assessing risks across the model lifecycle — before deployment and, as appropriate, before and during training — taking into account capabilities, deployment context, mitigations, and results from internal and external evaluations, including government and third-party evaluations. Commitment II requires setting explicit thresholds at which severe risks, absent mitigation, would be deemed intolerable; thresholds are to be defined with input from home governments, kept consistent with relevant international agreements, and published with rationale and example scenarios. Commitment III requires articulating mitigations — both behavioural and security mitigations, including protection of unreleased model weights — to keep risks below those thresholds. Commitment IV requires setting out processes for threshold-breach scenarios and states that "in the extreme, organisations commit not to develop or deploy a model or system at all, if mitigations cannot be applied to keep risks below the thresholds." Commitment V requires continually investing in advancing risk-assessment and mitigation capabilities and contributing to emerging best practice, international standards, and science.
Outcome 2 — Accountability
Commitment VI requires adopting and continuously reviewing internal accountability and governance frameworks, including assigning roles, responsibilities, and sufficient resources.
Outcome 3 — Transparency
Commitment VII requires public transparency on the implementation of Commitments I through VI, except where doing so would increase risk or reveal disproportionate commercial information, and requires sharing more detailed information with trusted actors such as a home government or appointed body. Commitment VIII requires explaining the involvement of external actors — governments, civil society, academics, and the public — in risk assessment and in reviewing adherence.
Signatories
The 16 original signatories on 21 May 2024 were Amazon, Anthropic, Cohere, Google/Google DeepMind, G42, IBM, Inflection AI, Meta, Microsoft, Mistral AI, Naver, OpenAI, Samsung Electronics, Technology Innovation Institute (UAE), xAI, and Zhipu.ai.
Subsequent endorsers, per gov.uk updates, were Magic, Minimax, 01.ai, and NVIDIA (Source: gov.uk).
Zhipu.ai, 01.ai, and Minimax are Chinese labs, making the Seoul Commitments one of the few documents with formal US, European, and Chinese frontier-lab co-signatories.
The definitional footnotes tie the commitments' scope to organisational headquarters: "home governments" is defined as the government of the country in which the organisation is headquartered, and thresholds may be defined "using model capabilities, estimates of risk, implemented safeguards, deployment contexts and/or other relevant risk factors," so long as breach is assessable (Source: gov.uk).
Comparison with other approaches
Compared with the Bletchley Declaration, which is state-facing, the Seoul Commitments are company-facing and share the same lineage. Relative to the G7 Hiroshima Code, which has 11 detailed Actions, the Seoul Commitments comprise 8 commitments with stronger "don't deploy" language and an explicit home-government role.
The commitments formalise practices that frontier-lab Responsible Scaling Policies operationalise: Commitment II (thresholds) and Commitment IV (don't-deploy) are mirrored in Anthropic's policy and OpenAI's framework. Compared with the EU GPAI Code of Practice, the Seoul Commitments are voluntary worldwide, whereas the GPAI Code is voluntary but tied to the EU market; signatory overlap between the two is substantial.
Areas of contention
The threshold for "intolerable risk" is not defined quantitatively. Thresholds are chosen by each company, producing heterogeneity across signatories: Anthropic uses ASL capability levels, while OpenAI uses Tracked Category risk levels.
Enforcement is reputational. No mechanism exists to verify compliance or penalise a breach beyond public scrutiny.
Commitment II's reference to "home government input" creates divergent obligations across jurisdictions: US labs work with the US AISI, UK labs with the UK AISI, and Chinese labs with Chinese authorities. Interoperability across these arrangements is presumed but not specified.
Relationships
- depends-on: The Bletchley Declaration (AI Safety Summit, 1–2 November 2023)
- supports: AI Safety Cases and Frameworks
- related: Anthropic's Responsible Scaling Policy (Version 3.1), Anthropic's Responsible Scaling Policy (Version 2.2), OpenAI Preparedness Framework V.2, Frontier Compliance Framework (February 2026), Paris AI Action Summit Declaration (2025), G7 Hiroshima Code of Conduct for Advanced AI (2023), EU General-Purpose AI Code of Practice (Final Version, 2025), UK AI Safety Institute (AI Security Institute)