Delegates from 118 countries agreed on draft language establishing a common evaluation regime for the most capable artificial intelligence systems, ending a week of negotiations that several participants described as the most technical the United Nations has ever hosted. The draft requires developers of frontier models to submit standardised safety evaluations to an international registry before wide deployment.
The breakthrough came after negotiators separated the question of evaluation — where consensus proved reachable — from the far harder questions of liability and enforcement, which are deferred to a follow-on protocol. Smaller nations won a provision guaranteeing access to evaluation tooling, arguing that governance without capacity would entrench a two-tier system.
Ratification is not assured. The draft heads to capitals with an eighteen-month clock, and observers note that at least three major AI-producing states secured reservations they may yet widen. Still, veteran diplomats drew comparisons to the early arms-control era: an imperfect first instrument, they argued, beats a perfect one that never exists.