Sixteen companies sign the Frontier AI Safety Commitments in Seoul

At the AI Seoul Summit, co-hosted by the UK and South Korea as the follow-up to November 2023’s Bletchley summit, sixteen AI companies signed the Frontier AI Safety Commitments — voluntary pledges going further than Bletchley’s general statement of concern by attaching specific process obligations. Signatories spanned North America, Asia, Europe and the Middle East: Amazon, Anthropic, Cohere, Google/Google DeepMind, G42, IBM, Inflection AI, Meta, Microsoft, Mistral AI, Naver, OpenAI, Samsung Electronics, Technology Innovation Institute, xAI and Zhipu AI.

The commitments centred on three obligations. Companies agreed to publish safety frameworks explaining how they would assess risk across a model’s lifecycle and what mitigations they would apply; to define, with input from governments and other trusted actors, thresholds of “severe risk” that would be considered intolerable; and, in the commitment’s strongest language, to “not develop or deploy a model or system at all” where mitigations could not bring an identified risk below that threshold. Companies also committed to internal and external red-teaming, transparency about their approach subject to protecting sensitive commercial and security information, and investment in interpretability and other technical safety research. The frameworks themselves were due ahead of the next major summit, in France in early 2025.

As with the Bletchley Declaration, the commitments were voluntary and carried no enforcement mechanism — a company that published a weak framework, or none, faced reputational rather than legal consequences. Analysts at the time noted the document said little about how the promised thresholds would be defined or who would judge whether a company had honoured its own “unacceptable risk” line, leaving the substance of the commitment largely to be filled in later, by each signatory, in the frameworks it went on to publish.

The Seoul commitments extended the norm, established at Bletchley, of major labs publicly committing to specific safety practices, and set up the following year’s France summit as the venue where those still-unwritten frameworks and thresholds would be judged against what had been promised.