Sriram Krishnan (centre) on a panel at TechCrunch Disrupt NY in May 2017. Image: TechCrunch / Wikimedia Commons, CC BY 2.0, cropped
With Washington refusing to regulate frontier AI, three of its biggest makers are planning to do it themselves. Google, OpenAI and Anthropic are working on an independent body, tentatively called the Standards Authority for Frontier AI (SAFA), that would set safety standards and test the most powerful models, BankInfoSecurity reports, citing The Information. It could launch by the end of this year or early 2027.
What SAFA would do
The idea is concrete testing, not a code of ethics. According to the reports, SAFA would:
- develop common standards for assessing frontier AI models, and run benchmark tests against them;
- support third-party safety testing before models are released;
- set rules for reporting AI safety incidents.
It’s modelled on FINRA, the self-regulatory body that polices Wall Street brokers under government oversight. Because it would be a self-regulatory organisation, it wouldn’t need approval from Congress or the President.
Who could run it
The three companies have approached Sriram Krishnan, until recently the White House’s senior policy adviser on AI, to be chief executive. They have also approached former Biden science adviser Arati Prabhakar, former Secretary of State Condoleezza Rice and venture capitalist David Friedberg for roles, and have sounded out METR founder Beth Barnes and AI safety researcher Paul Christiano as scientific advisers. None of them has publicly confirmed involvement, and the companies haven’t announced the plan themselves, so treat the details as reported rather than official.
Picking Krishnan would be a shrewd political move. The Trump administration has fiercely resisted outside control of AI, and a body led by one of its own former AI advisers would be far harder for the White House to dismiss.
Where the idea came from
It started with Demis Hassabis. In a July 14 essay, Google DeepMind’s then-CEO called for a U.S.-led, industry-funded “Frontier AI Standards Body” on the FINRA model. “Initially, Frontier Labs would voluntarily share models with the Standards Body for review up to 30 days before release,” he wrote, TechCrunch reported at the time. He also warned that funding would “need to be substantial and likely mostly come from industry.”
Separately, Proactive reports that OpenAI and Anthropic have discussed letting each other test their commercial models for vulnerabilities, with limits on what data each could keep.
Why now
The timing is no accident. The news broke the day after Sam Altman and Dario Amodei told the UN Security Council that AI needs global testing standards and incident reporting, and heard the U.S. reject any “global governance of AI” in the same session. With governments either unwilling or too slow to act, the labs are building the referee themselves.
Who’s missing is telling too. Meta, xAI and Microsoft aren’t named in the reports. A standards body only works if the whole frontier signs up, and right now it covers three labs.
There’s also an obvious question: can a watchdog paid for by the companies it watches really be independent? FINRA faces the same criticism on Wall Street, and SAFA will have to answer it with who sits on its board and who pays the bills.
Sources: The Information, via BankInfoSecurity and Proactive; TechCrunch
Related
