Google, OpenAI, Anthropic Plan Frontier AI Standards Body

What it claims

Citing The Information, GovInfoSecurity reports that Google, OpenAI and Anthropic have agreed to set up the Standards Authority for Frontier AI (SAFA), targeted for end of 2026 or early 2027. People reportedly approached for positions include Sriram Krishnan (former White House AI policy adviser), Arati Prabhakar (former Biden administration technology official), Condoleezza Rice and venture capitalist David Friedberg. Beth Barnes (METR) and Paul Christiano were approached as scientific consultants. Both have recent ties to OpenAI. METR investigated the Hugging Face incident, in which OpenAI agents accessed third-party systems, and OpenAI recently appointed Christiano to its foundation's board. Readers judging how independent SAFA's evaluations would be should weigh those links.

How it works

The design comes from Demis Hassabis's July 14 essay. SAFA would be a FINRA-style self-regulatory body that writes common assessment standards and runs benchmark testing. Hassabis said funding "would need to be substantial and likely mostly come from industry" in order to hire top talent and secure compute. As a self-regulatory organization it would need no approval from Congress or the president. To hold any legal authority, though, it may have to register with a government agency, the way FINRA registers with the SEC, which approves all of FINRA's rule changes. The article names no candidate agency for SAFA.

Limits

• The report is secondhand, via The Information. The article does not say whether the labs will be SAFA's main funders.
• It names no standards, benchmarks, test procedures or disclosure requirements. There is nothing to evaluate or adopt yet.
• The existing government route is voluntary pre-deployment testing at the federal AI safety institute, created under a June executive order. The White House was supposed to issue clear guidelines for this process, and the Trump administration has refused to publish them. Labs currently have no public criteria to test against. Trump has said publicly that he sees no need for more AI regulation.
• No federal AI bill has made it out of committee, so state-level rules still dominate.
• Nvidia's Jensen Huang has taken an accelerationist position against the industry's "pacing" consensus. That suggests SAFA may not cover every major player.