Zuckerberg says AI safety is each lab's own responsibility, distancing himself from Amodei's call for coordination

Meta CEO Mark Zuckerberg has weighed in on the debate over whether to slow down AI development, arguing that labs should each take individual responsibility for safety by drawing on external evaluations and advisory input.

Zuckerberg posted on X, formerly Twitter, Wednesday that "every AI lab has the responsibility and incentive to move at whatever pace is needed to train their models safely, and the ability to take their own measures to ensure this."

He said Meta had delayed the release of Muse by several months to focus on safety and security. "We did it as a normal part of our work simply because it was clearly the right thing to do for people and for us," he wrote. "I'm proud of the security foundation we've built."

He added that engaging independent evaluators and advisers is industry best practice, noting that Meta Superintelligence Labs already does this across multiple areas "because it helps us produce better outcomes." He went on to say that other labs "can just do this too" and that the field as a whole would benefit from "a larger and more diverse ecosystem of evaluators."

Zuckerberg also argued that devoting the overwhelming majority of compute to serving people — rather than racing toward recursive self-improvement — is "one of the best ways to develop this technology safely," adding that Meta has made that commitment and other labs can do the same.

Zuckerberg was the last of the five heads of the top AI companies to publicly state his position on the AI slowdown debate.

The discussion was set off on Saturday when Anthropic CEO Dario Amodei proposed that leading AI companies voluntarily coordinate on safety standards and moderate their development pace. OpenAI CEO Sam Altman, xAI CEO Elon Musk and Google DeepMind CEO Demis Hassabis all publicly endorsed the proposal.

Zuckerberg's remarks — framing the issue as each lab taking responsibility through external evaluation rather than collective coordination — were widely read as a rejection of Amodei's proposal. The stance reaffirms the position he laid out in an essay titled "A Future for Everyone" published last month.

In that essay, Zuckerberg wrote that "people won't want to use agents that act against their interests or don't do what they ask, so labs have a strong incentive to make their models more aligned." He acknowledged the argument that capability development should be slowed until safety catches up, but contended that "trust and alignment are becoming the most important capabilities that distinguish agents and models."

He also noted that legal liability for model-related harms gives labs an additional incentive to prioritize safety.