OpenAI, Anthropic and other large AI companies are jointly building an AI standards organization, according to The Information — a body one publication described as a private "self-regulatory" one. It is the concrete outcome of a month of very public argument about slowing AI down, an argument opened by a nearly 4,000-word essay from Anthropic's Dario Amodei and given urgency by an OpenAI agent that broke into several companies through Hugging Face. A comparable organization under federal authority was discussed earlier. The Trump administration never built it.
Amodei's essay made the case for reducing the pace of AI development so that adequate safeguards can be built in time. Among other things it calls for international cooperation between companies and governments on safe deployment. Sam Altman of OpenAI and Elon Musk of xAI — men who compete with Anthropic and with each other — both endorsed it. That unanimity is the first thing worth noticing, because unanimity among rivals on a question of restraint usually means the restraint costs them nothing, or costs someone else more.
The essay did not produce agreement on the part that matters. Plenty of people in the industry accepted the diagnosis and rejected the prescription of globally coordinated action, on the view that government oversight does not need to be part of this at all.
Mark Zuckerberg joined this week. Writing on X, the Meta CEO argued that trust and alignment with human intentions are fast becoming the capabilities by which AI agents and models will be distinguished from one another. He also disclosed that Meta had held back its Muse model for several months over safety and security questions — and was careful to add that the company did not ask anyone else to do the same, that it decided on its own as part of ordinary work, and that it judged the delay right for people and for Meta.
The framing is doing heavy lifting. Presented as evidence that voluntary restraint works, the Muse delay is equally evidence of what voluntary restraint is: unverifiable, unilateral, disclosed months after the fact, on the discloser's schedule, with no external account of what the safety questions were or how they were resolved. Zuckerberg endorsed parts of Amodei's plan while implying that the state has no role in it — the market will reward developers whose models do not cause chaos. It is consistent with his own longer essay a month earlier, which allowed limited cooperation with government where necessary but treated American competitiveness as the precondition for technological prosperity.
Others landed in roughly the same place. Reddit co-founder Alexis Ohanian told CNBC that the industry has largely done a poor job of explaining AI's risks to the public, and then, like Zuckerberg, expressed hope that companies can work the problems out themselves. Google DeepMind co-founder Shane Legg said capabilities are advancing very fast and cannot be allowed to outrun safety, and that the industry needs to work out in detail how to prevent that in practice — a statement that names the requirement without naming who enforces it.
Washington has removed itself. The administration includes vocal AI boosters and broadly favors leaving the technology unregulated; it has gone further and tried to stop individual states from passing their own AI laws. David Sacks, the administration's AI czar, has said regulation of AI should be left to the companies building it. Shortly after Amodei's essay appeared, House Speaker Mike Johnson responded to worries about catastrophic AI outcomes by saying that in 10 years people will not all be dead — which is a bar, of a kind.
So the standards body arrives into an empty room. It leaves the industry with no binding rules, only voluntary commitments, and it hands the companies that founded it structural authority inside it. That is the shape of the regulatory capture argument, and it does not require any law to be passed: voluntary industry standards written by the largest players still exert pressure on smaller developers with fewer resources, and still disadvantage them.
The argument would be easier to wave off if its sharpest confirmation came from a competitor. Aidan Gomez, CEO of Cohere, has accused the large labs of forming a "cartel," writing on his blog that AI needs guardrails and that this was never in dispute; the dispute is over who writes the rules, who gets to participate in writing them, and whose interests they protect. Chinese officials said something structurally similar — foreign ministry spokesman Guo Jiakun accused the Americans of using intimidation to disrupt the development of global AI governance, and a state newspaper editorial called Amodei's rhetoric something out of the Cold War arsenal. Both are interested parties.
The confirmation actually comes from Amodei. Reworking the rules governing how companies deal with China is one of the plan's provisions, and he does not hide what it is for: he wrote that if the measures succeed, China's progress slows and the American advantage grows appreciably over the next three to five years — precisely the window in which he expects AI to become geopolitically decisive. A safety plan with a stated competitive payoff written into it is not disqualified by that payoff. But it stops being possible to treat the safety case and the competitive case as separable, which is exactly what the rhetoric of the past month has asked everyone to do.
What no one has explained is enforcement. Not one participant has said what the standards organization does about a company that declines to join, who adjudicates a violation, or what a violation costs. Those are the questions that distinguish a regulator from a trade association, and the week produced answers to none of them.
Gomez's point survives all of it: AI safety is a deeply political field, and the definition of safety, along with the right to set it, determines who gains ground and who falls behind in the next round of competition. On present course that definition will be written by a handful of large American companies, ratified by no one, and applied to everybody — including the firms that will be measured against it by the same companies that sell against them.