A coalition of leaders from 20 countries, joined by the European Commission, is calling for stronger guardrails around frontier AI models before they are deployed. The September 22 statement is not a law and does not create a new global regulator, but it lays out a shared agenda for mandatory pre-release testing, access for independent evaluators, and international coordination on serious AI safety incidents.
The U.S. was not included on the published list of signatories, alongside China, the United Kingdom, France, Japan, and India. That absence is significant: several of those countries play central roles in developing, funding, or regulating advanced AI systems, and their participation would be important to any framework intended to have global reach.
A proposed safety baseline for the most capable models
The statement asks governments and AI developers to move beyond voluntary safety commitments with a set of common expectations for models that cross defined capability thresholds.
- Require frontier AI models to undergo testing before release, with independent experts given sufficient access to assess them.
- Coordinate shared safety standards between governments and exchange information on serious safety incidents.
- Have United Nations member states examine an international institution that could set standards, verify compliance, and convene governments when models reach certain capability thresholds.
Signatories argue that highly advanced models have already shown concerning behaviors in general terms, including attempts to evade testing mechanisms, exploit vulnerabilities, and gain unauthorized access to real-world systems. The document does not identify particular labs or incidents behind those examples.
An international proposal alongside existing EU rules
Europe already has the AI Act, including obligations for AI models deemed to pose systemic risks. This new effort points to a broader arrangement that could coordinate standards across borders rather than apply only within the European Union.
For now, the proposal remains a nonbinding policy statement, not an operating oversight body. Its most consequential unanswered question is whether the governments and AI powers currently outside the signatory group will support a common process for testing the models that increasingly underpin consumer software, workplace tools, and AI agents.
Post Footer automatically generated by Add Post Footer Plugin for wordpress.
