The biggest AI labs are not only competing. They have also been talking about safety together.
TechCrunch's Rebecca Bellan reports that Chris Lehane, OpenAI's global policy chief, told reporters Tuesday that OpenAI has been working with Anthropic and Google DeepMind on AI safety for weeks. Bloomberg first reported the talks. Lehane has been in Washington working with lawmakers on catastrophic AI risks.
The timing matters. The conversations follow a Saturday essay from Anthropic CEO Dario Amodei that urged the industry to slow the pace of frontier AI. Sam Altman, Demis Hassabis, and Elon Musk backed Amodei's letter. Altman also said OpenAI would join Anthropic in embedding third-party evaluators.
That kind of coordination raises an old legal worry. Some people, including Altman, noted that talks could create antitrust risk if companies work together in ways that suppress competition. Amodei proposed a narrow government waiver so firms could cooperate on safety without that chill. Lehane said the firms do not need one.
According to The Information, cited by TechCrunch, the three companies are working to create an AI industry standards body. Altman reportedly told staff that such a body would need to happen without U.S. government support. Hassabis called in July for a U.S. standards body or AI watchdog that could screen advanced models and help coordinate slowdowns when needed.
Politics is pulling the other way. President Trump dismissed safety concerns as a hoax and argued that a slowdown would help China. Advisor David Sacks said existential risk fears are overblown. Against that backdrop, Lehane said OpenAI supports a FRONTIER Act provision that would force top labs to allow independent verification organizations.
I read this as an industry trying to build shared guardrails while the political climate is skeptical of slowdowns. Independent evaluators, a standards body, and forced verification sound dry. For regular people, the stake is simpler: the tools that will sit inside banking apps, hospitals, and workplaces need someone other than the builder checking what they can do.
None of this freezes progress by itself. Talks are not laws. A standards body without government teeth may struggle. Still, when OpenAI, Anthropic, and Google DeepMind spend weeks on the same safety problem, that is a signal. The labs know frontier systems create shared risk, and they are at least practicing how to respond together.
If you care about AI as a user or a builder, watch two follow-ups. First, whether third-party evaluators actually get real access, not just press-friendly demos. Second, whether any standards body publishes clear rules for when a model is too risky to ship. Safety talk is cheap. Verified testing and honest slowdowns when needed are the parts that help people.
