Google, OpenAI and Anthropic are moving to launch a joint self-regulatory body for frontier artificial intelligence, provisionally called the Standards Authority for Frontier AI, that would set shared rules for safety testing and pre-release review of the most capable models. The three companies are aiming for a launch by the end of this year or in early 2027.
The effort, first reported by The Information and covered widely since, would have the group operate independently of any government, developing frameworks for pre-deployment safety testing, standards for who qualifies as an independent auditor, and common rules for reporting safety incidents. Reporting on the plan, including from Computerworld, describes it as an attempt to give the companies' existing public safety pledges practical, checkable substance.
Why now
The initiative traces back to a proposal Google DeepMind chief executive Demis Hassabis floated in July for an outside body to evaluate frontier models. According to multiple outlets, the three labs pivoted toward pure self-regulation after a draft White House executive order that would have imposed binding federal testing requirements was shelved. The companies have reportedly approached former White House AI policy adviser Sriram Krishnan and ex-OSTP director Arati Prabhakar for leadership roles, alongside potential technical advisers.
OpenAI has framed the move as central to managing AI's pace of advancement rather than just its content. "Fully autonomous [recursive self-improvement] is not happening today, and we should not pursue it unless and until it can be done safely," the company said, adding that shared standards "may be as important to pacing the frontier as alignment research itself."
AI companies are in no position to provide these kinds of assurances.
Carmi Levy, technology analyst
Skeptics note the obvious tension: the same companies racing each other to ship more capable systems would also be writing and policing the rules for testing them, with no government enforcement backstop if a member falls short. The plan lands the same week researchers published papers questioning whether current audit methodologies can reliably catch dangerous model behavior at all, an irony not lost on critics of self-regulation.
None of the three companies has published formal governance documents for the group yet. What happens next hinges on whether other major labs — and eventually regulators in the EU and elsewhere — treat the body as a credible complement to law, or as an attempt to preempt binding oversight before it arrives.