Sixteen AI companies sign voluntary safety pledges at Seoul Summit

Sixteen of the largest AI developers and deployers — Microsoft, Amazon, Google, Meta, OpenAI, Anthropic, Cohere, G42, IBM, Inflection AI, Mistral AI, Naver, Samsung Electronics, Technology Innovation Institute, xAI, and Zhipu.ai — [signed the Frontier AI Safety Commitments](https://www.cio.com/article/2119325/big-tech-companies-commit-to-new-safety-practices-for-ai.html) on Tuesday at the AI Seoul Summit in Korea. The signatories agreed to publish safety frameworks that detail how they will measure risks, define a “safety handbrake” for when those risks become intolerable, and commit not to deploy a model if its risks cannot be held to a certain threshold.

This is not the Bletchley Declaration. That November 2023 agreement was between governments — the EU, US, China, and others — and established a shared understanding of frontier AI risks. The Frontier AI Safety Commitments operate at the organizational level. Companies choose their own thresholds. They decide what “intolerable” means. And there is no enforcement mechanism.

The question is whether voluntary, self-defined guardrails produce meaningful safety or merely a veneer of responsibility.

Maria Koskinen, AI Policy Manager at AI governance vendor Saidot, pointed to the contrast with the EU AI Act. The EU AI Act defines “systemic risk” and provides guidance on risk management for general-purpose AI models. The Frontier commitments leave threshold-setting to the signatories. Koskinen noted that the EU approach “gives more certainty not only to organizations implementing these commitments but also to those adopting AI solutions and individuals being impacted by these models.”

The difference is material. A company that defines its “intolerable” risk as a catastrophic event with a 1-in-10,000 chance is operating differently from a company that sets the bar at a 1-in-100 chance of moderate harm. Both can claim compliance. The commitments do not specify a baseline.

Pareekh Jain, CEO of Pareekh Consulting, called the move “a step towards ethical AI.” He noted that the commitments can help guide CIOs in understanding AI-related risk and risk-management actions as they deploy the technology.

That framing is revealing. The commitments are positioned as a guide for enterprise buyers, not as a binding constraint on frontier labs. A CIO evaluating a model from a signatory company can point to the published safety framework and ask questions. That is better than nothing. But the framework is written by the vendor, not verified by a third party, and the thresholds are set by the vendor.

The 16 signatories include the usual suspects: the five hyperscalers and frontier labs that dominate the AI conversation. But the list also includes Samsung Electronics, Naver, G42, and Zhipu.ai — companies with significant AI ambitions in consumer electronics, search, sovereign AI, and the Chinese market respectively. The Technology Innovation Institute is a UAE-based research organization. The breadth suggests the commitments are designed to be inclusive enough to attract a wide range of signatories, which may come at the cost of specificity.

Koskinen acknowledged the voluntary nature: “While the commitments are voluntary, with no enforcement … they still set a precedent for other AI organizations to follow.”

Precedent is real. The Bletchley Declaration established that governments see frontier AI as a governance object. The Frontier AI Safety Commitments establish that companies are willing to publicly state they will measure risk and define thresholds. That is progress from a world where safety was entirely internal and unspoken.

But precedent is not enforcement. The commitments do not include auditing, verification, or consequences for non-compliance. A company could publish a safety framework, define a high threshold, and then modify its deployment plans without public notice. The “safety handbrake” is self-applied.

The commitments also do not address the compute supply chain. Frontier AI models require massive training clusters. The companies that build those clusters — NVIDIA, AMD, TSMC — are not signatories. The data centers that host them are not signatories. The commitments focus on model-level risk, not infrastructure-level risk, which is where concentration and single points of failure live.

For AI builders, the practical signal is mixed. The commitments create a new document genre: the safety framework that accompanies a model release. CIOs and enterprise buyers should expect to see these frameworks alongside model cards and system cards. They should read them critically, asking whether the thresholds are meaningful and whether the “intolerable” definition is specific enough to constrain behavior.

The next test comes at the AI Action Summit in France early next year, where the signatories have pledged to release their threshold definitions. If those definitions are vague or uniform across companies with very different risk profiles, the commitments will have failed their first real test. If they are specific and varied, they may actually constrain behavior.

The commitments are a signal, not a solution. They tell the market that safety is a topic companies are willing to discuss publicly. They do not tell the market that safety is enforced. That gap is where the real work begins.