What happens when the companies building the world's most powerful AI also get to grade their own homework? This week, we may be about to find out.
The plan
According to a September 24 report from The Information, OpenAI, Google and Anthropic are building an industry-led body — tentatively called the Standards Authority for Frontier AI, or SAFA — that would write safety rules for frontier models. The target: a launch in late 2026 or early 2027. The catch: it would operate with no government oversight.
What it would actually do
The remit is the unglamorous machinery of AI safety: supporting third-party groups that test models before release, writing the playbook for how labs report safety incidents, codifying the industry's voluntary safety commitments, and setting qualifications for independent AI auditors. One working group is even considering giving SAFA its own testing team — which would turn it from a standards publisher into a genuine gatekeeper.
The people — and the irony
The labs have reportedly approached Sriram Krishnan, the former White House AI policy adviser who publicly declared "there will not be an FDA for AI," to run it as CEO. Former White House science director Arati Prabhakar is also in the mix, and former Secretary of State Condoleezza Rice is reportedly under consideration for board chair. The model for the whole thing is FINRA, the self-regulatory body that polices Wall Street brokers — an idea floated by Google DeepMind's Demis Hassabis.
Here's the twist: just one day before the SAFA report broke, OpenAI's Sam Altman and Anthropic's Dario Amodei stood before the UN Security Council asking world governments to set global AI rules. A day later, reporting said their companies plan to write the rulebook themselves.
Why it matters
With no government involvement, SAFA couldn't legally block a model release — it would run on influence, not law. Critics worry three giants writing the rules could box out open-source developers and smaller rivals: Cohere's CEO called the arrangement "a cartel by any other name." And there's already a real antitrust lawsuit, filed September 21, accusing Anthropic, OpenAI, Google and SpaceXAI of coordinating an "AI slowdown cartel."
The honest caveat
None of the three labs has publicly confirmed SAFA. All of this is reported, not announced — but the details are specific enough, and the reaction loud enough, that it belongs on your radar.
My take
Standardized pre-release testing is genuinely useful — right now, every lab grades itself by its own rubric. But SAFA's real test won't be its rulebook; it will be whether anyone beyond the three labs that created it ever gets a seat at the table. If it becomes a club for the biggest players, it won't build trust — it'll just move the gatekeeping indoors. Watch who gets invited.
Comments
Post a Comment