The summit concluded with a voluntary pledge among 32 governments and 18 leading AI companies to publish system cards detailing model capabilities, limitations, and red-team findings. The agreement also outlines a rapid response protocol for incidents that cross safety thresholds.
Officials said the framework will be iterated every six months to keep pace with frontier model releases. Civil society observers pushed for stronger enforcement, but noted the move as a “meaningful step toward sunlight.”
What changes now
- Baseline disclosure for significant model releases
- Pre-deployment testing for high-risk use cases
- Escalation pathway for newly discovered vulnerabilities
Several companies previewed “nutrition labels” for upcoming systems, summarizing bias testing, safety mitigations, and energy consumption. Regulators indicated those templates could inform future policy.
The signal here is collaboration over confrontation, at least for the next phase,” said one participant. “It buys time to align on the hardest questions.