Zuckerberg rejects coordinated AI slowdown as industry leaders split

Meta CEO Mark Zuckerberg pushed back on calls for a coordinated AI slowdown in September 15 comments, arguing that each lab already has its own reasons to pace development safely.

The comments, covered by The Rundown, came days after Anthropic CEO Dario Amodei called for outside scrutiny and coordination over the pace of AI development.

Meta’s case for setting its own pace

Zuckerberg said Meta delayed its new Muse agent for several months for safety reasons without asking rivals to match. He argued that users want agents that follow instructions, making reliable behavior a selling point. Labs that skip that work “will fall behind,” he said.

Meta announced Muse on September 8, describing an agent that can send emails and book travel through a dedicated virtual computer and browser. The company says users decide what it can access.

Zuckerberg also said Meta already works with independent evaluators. He said the company directs most of its computing power toward serving people and criticized the race to build AI that repeatedly improves itself.

What Amodei proposed

Amodei’s “We Must Pace the Frontier” essay calls for embedded outside evaluators, coordination among democracies, and global coordination to pace continued AI development.

Anthropic has committed on its own to the evaluator plan. Under the proposal, reviewers would get access similar to what employees have and the right to publish unfavorable findings, subject to confidentiality restrictions. That gives the debate a concrete question about how much access and freedom outside reviewers should receive.

Why it matters

The slowdown scorecard is split, with important differences within each group. Sam Altman and Elon Musk endorsed Amodei’s September 12 call. Demis Hassabis has backed the direction of the proposal, with details still unresolved. Altman supports pacing alongside continued progress.

Zuckerberg favors company discretion. Nvidia CEO Jensen Huang argues that speed and safety can coexist, while Donald Trump opposes efforts to obstruct AI and data center development. China’s Foreign Ministry has opposed the slowdown proposal. Those positions differ in scope, but each complicates efforts to build a broad agreement.

The split resembles a Prisoner’s Dilemma. A lab that slows while rivals press ahead risks giving them an advantage. Governments face a similar concern about foreign competitors. Amodei’s own proposal ties pacing among democracies to their technological lead over China. Shared concern about safety leaves each participant needing confidence that competitors will accept restraint too.

For people considering agents such as Muse, safety includes control over what the software can do through their accounts. Users and buyers should ask which actions require approval, how to revoke access, and what record remains after an unexpected action. Those answers would help them assess the safeguards behind a company’s safety claims.

Narrower agreements on outside review may still be possible. Both Zuckerberg and Amodei support outside scrutiny, though they describe different arrangements. Any common standard would need to spell out what reviewers can inspect, whether they can publish adverse findings, and what happens when they find a serious problem.

A shared pace would require major competitors to agree on what gets slowed and how to check compliance. Delaying a product’s release gives a lab time to address safety before users get access; development elsewhere can continue during that hold. Zuckerberg’s account of the Muse delay illustrates how individual safety decisions can fit inside an industry that keeps advancing. With major players resisting coordination, acceleration appears to be the default.