The Frontier AI Safety Commitments are voluntary corporate safety commitments issued at the AI Seoul Summit, co-hosted by the United Kingdom and the Republic of Korea on 21 May 2024, by 16 frontier AI organisations. The primary text is published by the UK Department for Science, Innovation and Technology on GOV.UK under the Open Government Licence, first published 21 May 2024 (Source: gov.uk). The commitments bind signatory companies to specific safety practices and serve as the corporate analogue of the state-level The Bletchley Declaration (AI Safety Summit, 1–2 November 2023). They are framed "in furtherance of the Bletchley Declaration", making them a corporate continuation of that earlier framework.
The primary text defines "frontier AI" as "highly capable general-purpose AI models or systems that can perform a wide variety of tasks and match or exceed the capabilities present in the most advanced models", and confines the commitments to frontier models or systems so defined (Source: gov.uk).
Summary of the commitments
Signatories undertake to develop and deploy their frontier AI models and systems responsibly in accordance with the commitments, and to demonstrate how they have done so "by publishing a safety framework focused on severe risks by the upcoming AI Summit in France" — a delivery deadline that tied the Seoul commitments to the subsequent Paris summit. The text also provides that, given the evolving state of the science, signatories' approaches to meeting the three Outcomes may evolve, with public transparency (including reasons) required when they do (Source: gov.uk).
The commitments are organised around three Outcomes and eight numbered commitments (I–VIII).
Outcome 1, that organisations "effectively identify, assess and manage risks when developing and deploying their frontier AI models and systems", covers assessing risks across the model lifecycle, including before deployment and, as appropriate, before and during training (I); setting intolerable-risk thresholds in coordination with home governments (II); identifying and articulating mitigations to keep risks under thresholds, including security controls for unreleased model weights (III); defining explicit processes for threshold breaches, including not deploying at all if mitigations cannot keep risks below thresholds (IV); and continually investing in improving these capabilities, contributing to and taking into account emerging best practice, international standards, and science on AI risk identification, assessment, and mitigation (V).
Outcome 2, Accountability, covers adhering to commitments I–V by developing and continuously reviewing internal accountability and governance frameworks, assigning roles and responsibilities, and resourcing safety work (VI).
Outcome 3, Transparency, covers public transparency on safety implementation — except insofar as disclosure "would increase risk or divulge sensitive commercial information to a degree disproportionate to the societal benefit", with more detailed non-public information shared with trusted actors including home governments (VII) — and explaining how, if at all, external actors such as governments, civil society, academics, and the public are involved in risk assessment and in assessing adherence to the signatory's safety framework (VIII).
Alongside the numbered commitments, signatories affirm a set of current best practices: internal and external red-teaming for severe and novel threats; working toward information sharing; investing in cybersecurity and insider-threat safeguards to protect proprietary and unreleased model weights; incentivizing third-party discovery and reporting of issues and vulnerabilities; developing mechanisms enabling users to understand whether audio or visual content is AI-generated; publicly reporting capabilities, limitations, and domains of appropriate and inappropriate use; prioritizing research on societal risks; and deploying frontier AI to help address global challenges (Source: gov.uk).
The commitments formalize the concept of "intolerable risk", a capability level at which a model cannot be safely deployed even with mitigations. The concept has since been adopted in frontier AI developer frameworks including Anthropic's RSP and OpenAI's Preparedness Framework.
Key claims
Sixteen organisations were original signatories (high confidence), listed in the primary text as: Amazon, Anthropic, Cohere, Google (encompassing Google DeepMind), G42, IBM, Inflection AI, Meta, Microsoft, Mistral AI, Naver, OpenAI, Samsung Electronics, Technology Innovation Institute (TII, UAE), xAI, and Zhipu.ai. The Chinese lab Zhipu.ai signed alongside US and European labs. The primary text confirms that Magic, Minimax, 01.ai, and NVIDIA were subsequently "added to the existing list" of signatories (Source: gov.uk).
Commitment IV, described as the "don't deploy" clause, states: "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." The source characterizes this as the strongest voluntary commitment in the public record (high confidence).
On intolerable-risk thresholds, companies must define thresholds with input from trusted actors including home governments as appropriate, align them with relevant international agreements to which their home governments are party, and publish how thresholds were chosen along with specific examples of situations where models would pose intolerable risk (high confidence). The primary text defines "home governments" as the government of the country in which the organisation is headquartered, and notes that thresholds "can be defined using model capabilities, estimates of risk, implemented safeguards, deployment contexts and/or other relevant risk factors", provided breach is assessable (Source: gov.uk). This element foreshadows the NIST-AISI and UK-AISI evaluation partnerships.
Relation to other frameworks
The commitments feed directly into frontier-lab frameworks including Anthropic's Responsible Scaling Policy (Version 3.1), OpenAI Preparedness Framework V.2, and Frontier Compliance Framework (February 2026). Within the broader AI Safety Cases and Frameworks landscape, they occupy the voluntary-commitments layer, above individual responsible scaling policies but below binding regulation.
The commitments contrast with the Paris AI Action Summit Declaration (2025), which de-emphasizes "safety" framing. Zhipu.ai's signature is one data point for AI Race Dynamics: Chinese frontier labs engaged with Western-led voluntary safety norms in 2024, before the 2025–2026 divergence.
Relationships
- supports: AI Safety Cases and Frameworks
- depends-on: The Bletchley Declaration (AI Safety Summit, 1–2 November 2023)
- related: Seoul Frontier AI Safety Commitments (2024), 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), Managing Advanced Cyber Risks in Frontier AI Frameworks, Paris AI Action Summit Declaration (2025)