Google, OpenAI and Anthropic have a fresh solution to the problem of nobody being able to control AI: they will regulate themselves. On 25 September the three announced the Standards Authority for Frontier AI, or SAFA, a self-regulatory body to set safety tests, auditing standards and incident-reporting rules for the industry. Who will it police? The industry. Who is building it? The three most powerful companies in that industry. Truly, the henhouse has commissioned a security review, and the foxes have kindly volunteered to write it.
The origin story is the tell. The idea came from Google DeepMind’s Demis Hassabis in July, and the labs originally wanted a proper public-private partnership with actual federal oversight. That plan stalled when a draft White House executive order was shelved and the administration told them to go and “reach industry consensus” first. So this is not the labs heroically choosing self-rule; it is the labs reaching for self-rule after the government waved them off and left the room. Given a choice between real oversight and marking their own papers, they got handed the second one by default.
Why the model should worry you
Hassabis has said SAFA would be modelled on FINRA, the body that regulates Wall Street brokers. Pause on that. FINRA is the go-to example whenever anyone wants to argue that industry self-regulation is a polite fiction that lets an industry set rules it can live with and discipline its own members with a feather. Choosing it as the template is either refreshingly honest or completely oblivious, and neither is comforting. The people who spent this month telling the UN that AI might end civilisation are now proposing to be graded on that risk by a club they founded.
To be fair, and there is a fair case, an industry body with real teeth is better than nothing, and the labs are courting serious names to run it, reportedly including Condoleezza Rice and figures from both the Biden and Trump orbits. Voluntary standards, incident reporting and shared auditing frameworks are genuinely useful if they have force. The entire question is force. A regulator the regulated can leave, ignore, or defund the moment it becomes inconvenient is a press release with a logo, and until SAFA can do something a member does not want done, that is exactly what it is.
The takeaway
Watch for one thing only: can SAFA ever compel a member to do something it would rather not, and punish it if it refuses? If yes, this is a real, if imperfect, step while government dithers. If no, it is the AI industry building a very expensive mirror and calling it a watchdog. The context makes the cynical read hard to shake: this only exists because Washington passed on the job, and an industry that has spent the year insisting its product is uniquely dangerous has now appointed itself the referee of that danger. Mark your own homework long enough and you will always pass. (Sources: TheStreet, BankInfoSecurity, GuruFocus, September 2026.)
Related: the slowdown pact that became a lawsuit.