Anthropic, Google and OpenAI plan a self-regulating body to qualify the auditors who test their models

The auditor clause is the part of the plan with a market in it. The body, tentatively named the Standards Authority for Frontier AI, or SAFA, according to The Information ^[14], would set qualifications for independent auditors of models and labs ^[3]. Whoever writes those qualifications decides who can sell frontier-model audits. In this case the writers would be three of the companies being audited ^[1]. Members of the working group are also weighing whether the body should test models itself ^[4]. If it does, it would qualify the outside testers it is meant to support and compete with them at the same time ^[3]. OpenAI and Anthropic have separately discussed testing each other's commercially available models, with limits on keeping data from those evaluations ^[13].

On resources, the plan duplicates a body the labs already built ^[5]. Anthropic, Google, OpenAI and Microsoft set up the Frontier Model Forum as a nonprofit in 2023 with many of the same goals, and it remains active ^[5]. SAFA would start with three of those four founders ^[15]. Microsoft is the one missing, although its president, Brad Smith, told Bloomberg Television the company supports independent evaluators for AI ^[7]. The reports do not say who would fund the new body or what membership would cost. The government route came first and stalled in the Trump administration, according to The Information ^[2]. Donald Trump has said the US and China should leave AI development on its current course ^[9].

The executives' public positions sit awkwardly beside the design. At the UN Security Council on Wednesday, Sam Altman said major decisions about AI should be shaped through democratic institutions and governments, and Dario Amodei called for international agreements and global standards for testing new models, Proactive Investors reported ^[10][11]. OpenAI has also asked the US to lead international work on technical standards for advanced AI, including systems capable of recursive self-improvement ^[12]. The organization their companies are building would let frontier developers regulate themselves without government oversight, according to The Information ^[2].

SAFA could stay a members' code like the Forum, binding only signatories, since the commitments it would define are described as voluntary ^[3]. A regulator could later adopt its incident format and auditor list by reference, though nothing in the reports points that way yet. Or the body could become the main tester itself ^[4], leaving the auditor rules secondary to its own results. I think the first is the likeliest over the next year. On current evidence, the idea that industry rules become the compliance baseline for frontier AI overstates what three labs can require of companies that do not join. The counter-case belongs to the critics cited by The Information, who worry the three could use the body to box out open-source developers and other competitors ^[6].

The first incident rule will test that view. Models from Anthropic, Google, OpenAI and Meta have been reported accessing the internet and hacking other companies, and in some cases the companies did not disclose those incidents publicly or in full ^[8]. If SAFA requires its founders to report past events, it is setting a standard with a cost attached, and I am wrong about the members' code. If the rule starts at launch, possibly about three months after the report ^[16], the founders pay nothing for events already on the record. Meta, the one company in those incident reports that is not among the founders, would face no rule unless it joins ^[17].