· AI
Google, OpenAI, and Anthropic's New Safety Body Locks Out Their Biggest Rivals
Self-regulation only beats a government mandate when it is not just the three biggest labs writing a rulebook that keeps everyone else out of the room.
Google, OpenAI, and Anthropic are building a private safety authority for frontier AI, and the companies with the most to lose from it are not in the room where it is being designed. Reporting on the plan, confirmed by the three labs to Bloomberg, CNBC, and TechCrunch, describes the Standards Authority for Frontier AI, or SAFA: a body modeled in part on FINRA, the organization that polices Wall Street brokers, meant to set common testing benchmarks, certify third-party auditors, and handle incident reporting for the largest AI models. A separate account describes the same effort as designed to get ahead of government rules by setting shared testing and audit standards first. The idea traces back to a proposal Google DeepMind chief Demis Hassabis floated in July. Target launch is late 2026 or early 2027, with names like Condoleezza Rice and former White House AI adviser Sriram Krishnan floated for leadership.
I want less government rule-writing in this industry, not more, so my instinct on a story like this should be to cheer it. I am not cheering, and the reason is who gets a seat.
The steelman for SAFA is real and I have made a version of it myself. A body the labs run and fund can move faster than any legislature, and it avoids exactly the kind of outcome I wrote about here three days ago: a state bill that starts with real teeth and gets negotiated down to a penalty a $500 million company can treat as a line item. Chris Lehane, OpenAI's chief global affairs officer, made this case directly in a September 9 post arguing the policy window is open and industry needs to act before Washington writes something worse. If the choice really is SAFA or a clumsy federal mandate, SAFA is the better bet on the merits. Faster iteration, technically literate standards, no compliance theater written by staffers who have never trained a model.
But that is not the actual choice on the table, because SAFA is not open to the industry. Meta, xAI, and Nvidia rejected it publicly at Dreamforce on September 15. Mark Zuckerberg's argument was that labs already face strong liability incentives and existing law is enough. Jensen Huang went further, calling additional AI-specific rules unnecessary and arguing makers should simply be held responsible for what they ship. Cohere's chief executive, Aidan Gomez, was blunter about the structure itself, calling the proposed body "a cartel by any other name." That is not an argument against AI safety standards existing. It is an argument that the three companies with the largest compute budgets should not also be the three companies deciding who counts as safe enough to compete with them.
The timing makes the cartel framing harder to wave off, not easier. On September 12, Anthropic CEO Dario Amodei published an essay arguing the industry should deliberately slow capability gains by a year or two, framed around the idea of pacing the frontier rather than racing it. Within hours, rivals including Sam Altman and Hassabis publicly endorsed it. Six days later, four subscribers to the labs' own AI products sued Anthropic, OpenAI, Google, and xAI in federal court in San Francisco, arguing that a public agreement among competitors to slow product improvement is an agreement to restrict output under Section 1 of the Sherman Act. Worth noting: xAI's Elon Musk endorsed the pacing idea the same week his company publicly rejected joining SAFA. That split is the tell. The objection from the holdouts is not to safety coordination in general. It is to a structure where three incumbents write the rules, certify the auditors, and decide whose model clears the bar, while a fourth major frontier developer is on the outside looking in.
Whether that lawsuit wins is a separate question from whether the underlying concern is sound. A safety certification from a body its own subjects also run is not independent verification, it is a seal of approval the largest labs control the price of admission to. Smaller labs and the open-weight developers who have done more for AI affordability this year than any policy body, mine included when I recommend them to clients, would either pay to meet standards calibrated to what Google, OpenAI, and Anthropic already do well, or compete without the seal at all.
The practical takeaway for a business owner evaluating AI vendors: do not treat a SAFA badge, if it ships, as a substitute for your own evaluation of a model's cost, performance, and fit for your workload. A seal written by three sellers about themselves is marketing with better production values, not proof. If you want a second opinion on a model or vendor decision that is not filtered through whoever is selling it to you, that is a conversation worth having at mojoaiservices.com/#contact.
Sources
References used in this article. Links also appear alongside the relevant claims.
Tell us what's on your mind.
You don't need a polished brief to reach out. A two-line email about what's bugging you is plenty; we'll tell you straight if we're the right fit, and what we'd tackle first.
We'll scope the work around your workflow, goals, and timeline before quoting anything, so you know what's included before committing.