OpenAI, Anthropic and Google have been discussing AI safety with each other for several weeks, OpenAI's Lehane confirmed on Tuesday, attaching a timeline to cooperation the companies had previously only hinted at. The talks follow an essay Amodei published on Saturday calling on the industry to slow the development of frontier AI together in order to reduce catastrophic risks. The appeal was publicly backed by Sam Altman of OpenAI, Demis Hassabis of Google and Elon Musk of SpaceXAI. Altman said OpenAI would join Anthropic in bringing third-party evaluators into the companies to monitor the safety of their systems.
The legal argument underneath those talks is the more revealing part. Several executives, Altman among them, have warned that coordinating this way could breach antitrust law if regulators decide it restricts competition. Amodei's proposed answer is a narrow exemption from the law that would permit companies to coordinate on safety. Lehane's answer is that companies need no such permission.
That is a genuine split rather than a shade of emphasis. One company wants written cover before it acts; the other says the activity is already lawful, and that asking for an exemption concedes it might not be. Whichever view prevails determines how much of this coordination ever happens in public.
The talks confirm earlier reporting on cooperation between the labs. In a Fortune interview late last week, Altman hinted that he was in closed-door discussions with other AI company heads. This week, The Information reported that OpenAI, Anthropic and Google DeepMind are jointly building an organization to develop AI standards. According to the publication's source, Altman told employees the organization should come into being without support from the US government.
That last detail is worth holding onto. In July, Hassabis called on the US to create a new standards body that would act as an oversight structure for AI, testing the world's most advanced models and coordinating a slowdown in the industry if risks began to rise. Two months later, three of the largest labs are assembling something with a comparable remit and deliberately keeping Washington out of it.
The reason is not hard to locate. President Donald Trump has dismissed AI safety concerns as fiction and opposed tighter regulation, arguing that any slowdown would hand China an advantage in the technology race. His AI adviser, David Sacks, holds the same position, treating existential-risk concerns as exaggerated. Sacks has invested in the technology industry for years and is personally connected to a significant part of it.
Read together, the shape of this is clearer than any single announcement. The labs are not choosing self-regulation over government regulation. They are building the only structure available, because the government they would otherwise hand oversight to has said the risk is imaginary. Hassabis asked for a referee in July and is now helping construct a private one. That is not a change of conviction; it is what happens when the referee declines to appear.
OpenAI's stance on legislation is a partial exception. At the same Tuesday appearance, Lehane said the company supports a provision of the FRONTIER Act that would oblige leading labs working on frontier models to admit "independent verification organizations" into their companies to oversee the safety of model development. So OpenAI backs a statutory requirement to let outside inspectors in, while its chief executive tells staff the standards body should stay clear of the government. The two are compatible — verification is not standard-setting — but only just. The company is asking to be legally compelled to open its doors while reserving the right to write the rules the inspectors would measure it against.
Nothing said publicly defines what slowing down means. Amodei's essay asks for a joint deceleration and Hassabis's July proposal would coordinate one if risks grew, but neither the threshold that triggers it, nor the measurement that establishes the threshold, nor the consequence for a member that ignores it has been described. A standards organization created by the three companies it exists to constrain, with no government backing and no stated penalty, has exactly the authority its members choose to grant it in any given quarter. Musk endorsed the call for a slowdown, and his company is not among the three building the body.
TechCrunch asked Anthropic, Google and OpenAI for comment. The more telling gap sits in what has already been said. An essay calling for a slowdown, executives endorsing it, evaluators being invited in, a standards body being assembled — and in none of it has any company named a single thing it is actually slowing. What is being built is the machinery for deciding to stop. Nobody has described stopping.