On 20–21 November 2024, AI safety institutes and government-authorised bodies from Australia, Canada, the European Commission, France, Japan, Kenya, Korea, Singapore, the United Kingdom and the United States met in San Francisco for the first meeting of the International AI Safety Institutes network. The meeting ran along three tracks: synthetic content risks, foundation model testing and risk assessment of advanced AI systems. The participants are technical evaluation bodies, not legislatures or enforcement agencies, and that composition shapes what the network can produce.
What the network is, and is not
It is not a treaty body and does not create cross-border legal obligations. Members cooperate on research methods and testing capability; their work is to align evaluation methods, testing practice and the framing of conclusions. It is more accurately understood as a technical standards alliance than as a regulator. Outputs may influence regulation, but they are not regulation. Companies need not treat them as “international rules”, but should watch whether their methods are adopted by domestic regulators, procurement and platform governance.
What the three tracks address
- Synthetic content risks. Deepfakes, voice cloning, content provenance and labelling, with direct links to platform governance, election security and fraud prevention.
- Foundation model testing. Test methods, benchmark design and model access arrangements. The difficulty is completing independent evaluation without disclosing trade secrets.
- Risk assessment of advanced AI systems. Aimed at systems with significant capability and broad impact, focusing on systemic risk and loss-of-control scenarios before deployment.
The tracks correspond to different product stages: synthetic content sits at the application layer, foundation model testing at the development layer, and system risk assessment at the deployment layer. Application-layer companies relate most directly to the first; model or API providers see the latter two transmitted through customers.
How evaluation results enter commercial decisions
Outputs will not automatically become market access conditions, but travel through three channels: public procurement requirements, platform security questionnaires, and investor and insurer due diligence. For companies selling abroad, customers increasingly ask whether an independent evaluation has been done and what it was based on. A set of test records that can be explained externally is more useful than responding to questionnaires ad hoc. The records need not follow any national format, but should include scope, method, conclusions and limitations.
The direct implications of the synthetic content track
Synthetic content risk is the track closest to the application layer. Technical approaches to labelling, provenance and detection are converging, while enforcement differs by jurisdiction. Companies offering generation capabilities across multiple markets need a labelling strategy that can be adapted: how metadata, watermarks, visible labels and user notices are combined depends on the distribution channel and local requirements. Making labelling a fixed part of the generation process is more controllable than fixing it at the publishing stage; leaving responsibility entirely to users is usually not accepted in platform partnership reviews.
Where Chinese companies stand
China is not among the first members, which does not mean it is unaffected. Influence arrives indirectly through standards, customer requirements and open-source communities. For AI companies in Shenzhen, the practical question is whether one set of internal documents can answer requests from European, US and Southeast Asian customers. That requires assembling model descriptions, test records and risk management processes into reusable material in advance, rather than rewriting for each market. In external communication, explaining one’s own safety governance arrangements builds more credibility than asserting conformity with a particular country’s standard.
For multi-jurisdiction evaluation requirements, see multi-jurisdiction compliance; for synthetic content labelling, see AIGC labelling; for model obligations in the EU market, see EU AI Act services.
This article is based on public information about the first meeting in San Francisco on 20–21 November 2024 and the launch of the International AI Safety Institutes network. It is for reference only and does not constitute legal advice.