South Korea and the United Kingdom co-hosted the AI Seoul Summit, the second meeting in the series launched at Bletchley Park in November 2023. The agenda split cleanly. On the corporate track, leading developers made public safety commitments. On the governmental track, states agreed to deepen cooperation on evaluation science. What each track delivered, and what each left open, is still the clearest read on where AI governance stood in 2024.
The Frontier AI Safety Commitments
The headline document was the Frontier AI Safety Commitments, signed by sixteen major AI companies, including developers from the United States, China, the United Kingdom, and the UAE. Companies pledged to publish safety frameworks describing how they assess risks from frontier models; to define thresholds of risk they treat as intolerable; and, critically, not to develop or deploy models if those thresholds could not be kept below that line, including a commitment to pull back where mitigations fail.
The meaningful shift was public principle. Leading developers said, on the record, that some capabilities are too dangerous to deploy, and that they would define red lines and act on them. The industry moved from vague assurances toward stated limits. The limits remained self-defined.
Why voluntary commitments are not enough
- Companies set their own thresholds. Each developer defines intolerable risk and how to measure it. There is no common standard and no external body validating the definitions.
- No independent verification. Nothing confirms that an internal safety framework is followed, or that evaluations are honest. Compliance rests on the developer's word.
- No enforcement. Breaking a pledge carries reputational cost at most. The pledges are not law.
- They can be revised or dropped. Voluntary commitments under competitive pressure can quietly erode when a rival races ahead. That is the fragility of self-regulation.
The folk objection is that asking labs to self-police is better than nothing. True, and insufficient. A promise the promiser can define, measure, and break alone is a placeholder for governance. Placeholders only help if something harder replaces them on purpose.
The network of safety institutes
The governmental track produced something with more durable potential: momentum toward an international network of AI safety institutes. Building on the United Kingdom's and United States' newly created bodies, states agreed to cooperate on the science of evaluating frontier models, sharing methods, aligning testing, and building the technical capacity any future verification regime would need. Public bodies that can actually assess whether a model is dangerous are the connective tissue a treaty eventually requires. See the Foundation's guide to what an AI safety institute is.
The Seoul Declaration, endorsed by the assembled governments, reaffirmed human-centric, safe, and trustworthy AI and continued international cooperation. Like Bletchley's text, it was a statement of direction rather than a set of obligations.
The pattern and the gap
Seoul fits a known pattern in the governance of dangerous technologies: acknowledgment, then voluntary commitment, then (if the process succeeds) binding rules. The corporate pledges and the institute network are the second stage. The unfinished work is the third: common standards, independent verification, and legal obligation.
The safety institutes are the most promising bridge, because the capacity they build is exactly what a binding regime would use. The risk is that the summit process stalls at the voluntary stage, that the world settles for pledges because pledges are easier than treaties, while capabilities keep advancing. Seoul made the gap between commitment and governance visible. Closing it is the remaining task. The Foundation's position is that closing it means moving from company pledges to preventive international law before self-defined red lines dissolve under race pressure. Related reading: why voluntary AI safety commitments fail as a sole strategy.
Common questions.
What were the Frontier AI Safety Commitments?
Voluntary pledges signed by sixteen leading AI companies at the May 2024 Seoul AI Safety Summit. Companies committed to publish safety frameworks, define thresholds of risk they consider intolerable, and not develop or deploy models that cross those thresholds without adequate mitigation, including pulling back where mitigations fail. It was the first time major developers publicly committed to the principle that some capabilities are too dangerous to deploy.
Are the Seoul commitments legally binding?
No. They are voluntary corporate pledges. Each company sets its own risk thresholds and measurement methods, there is no independent verification that frameworks are followed, and there is no legal enforcement; the only consequence for breaking a commitment is reputational. They can also be revised or abandoned under competitive pressure, which is the central weakness of self-regulation.
What was the AI safety institute network?
A commitment by governments at Seoul to build an international network of national AI safety institutes, public bodies, starting with the UK's and US's, that develop the science of evaluating frontier models. They cooperate on testing methods and technical capacity. This matters because independent evaluation expertise is exactly what any future verification regime or binding treaty would require.
Did the Seoul Summit achieve binding AI governance?
No. It produced voluntary corporate commitments and a non-binding governmental declaration, plus progress on the safety institute network. It advanced the process from acknowledgment (Bletchley) toward commitment, but stopped short of common standards, independent verification, and legal obligation. The open question is whether the summit series converts these voluntary steps into binding governance before frontier capabilities outrun it.