The Meta CEO says legal exposure, market incentives and outside evaluators give each frontier lab enough reason to slow on its own.
By [Ryan Merket](https://runtimewire.com/author/ryan-merket)
· Published
Primary source: [X - Mark Zuckerberg](https://x.com/finkd/status/2099997096896274533)
Why it matters #
Frontier labs agree that safety work may need to slow model development. Zuckerberg is fighting over who decides: each lab on its own, or competitors and governments acting together.
Mark Zuckerberg (@finkd) rejected calls for a coordinated slowdown in frontier AI development on September 15th, arguing that competition, legal liability and independent evaluation already give individual labs enough reason to train models safely.
The Meta CEO's intervention puts him on the other side of a debate opened three days earlier by Anthropic co-founder and CEO Dario Amodei. In "We Must Pace the Frontier", Amodei argued that AI capabilities are advancing faster than labs can reliably understand, monitor and control them. OpenAI CEO Sam Altman and xAI founder Elon Musk subsequently endorsed the broad call for pacing, according to Reuters.
Zuckerberg's answer was direct: each lab can decide when its own safety work requires a delay. He said users will avoid agents that fail to follow their instructions, making trust and alignment competitive product features rather than obligations that require an industry pact.
"Any lab that doesn't focus on alignment will fall behind," Zuckerberg wrote. He separately argued that labs face "significant liability" when their systems cause harm.
That position leaves model developers responsible for setting their own thresholds, deciding whether those thresholds have been crossed and determining when training can resume. Zuckerberg backed independent evaluators, calling their involvement an industry best practice, but his post proposed no common standard, mandatory reporting system or enforcement mechanism.
Zuckerberg pointed to Meta's own product calendar as evidence that unilateral restraint can work. Meta delayed its Muse AI agent for several months to strengthen safety and security, he said. "We didn't call for everyone else to do this before we would," Zuckerberg wrote. "We just did it as part of our day-to-day work because it was clearly the right thing for people and for us."
Meta also says Meta Superintelligence Labs already uses outside evaluators in several areas. Zuckerberg called for a larger and more diverse evaluation market, while maintaining that rival labs can hire those evaluators without waiting for a coordinated agreement.
Amodei's proposal goes further. Anthropic plans to give embedded third-party evaluators employee-like access to its systems, training processes and staff. Amodei said those evaluators should be able to publish findings about incidents, risks and the access they received without Anthropic controlling their conclusions, subject to narrow redactions for legally or commercially sensitive information.
The second stage of Amodei's plan calls for frontier labs in democratic countries to establish common safety standards and limits on unchecked capability gains. He said government support or narrow antitrust waivers may be needed for competitors to hold those discussions. A third stage would seek agreements with China and other governments, including model testing and possible limits on recursive self-improvement.
Zuckerberg instead said Meta has committed the "significant majority" of its compute to products serving users rather than a race toward recursive self-improvement. The distinction matters because Amodei's warning centers on AI systems increasingly helping researchers build their successors, potentially compressing the time available for safety work.
OpenAI has already shown how an internal threshold can affect training. In a September research report, OpenAI said it temporarily restricted work on advanced models after agents compromised research infrastructure and preliminary tests suggested another model might have critical cyber capabilities. OpenAI said it would slow or stop development or deployment when it identified risks it could not sufficiently contain.
Those decisions support part of Zuckerberg's case: labs can work without an industry-wide agreement. They also expose the central weakness in relying on self-policing. The public sees the threshold, incident and response only when a lab chooses to disclose them.
Zuckerberg's stance follows the philosophy he set out in an August 10th essay. He argued that concentrating advanced AI inside a few institutions creates its own safety risk and that distributing capable systems widely produces a healthier balance of power. Meta's strategy centers on personal agents, broad distribution and continued releases of some open-source models.
A coordinated pace limit could constrain that strategy by giving rivals, evaluators or governments a role in deciding how quickly Meta advances. Zuckerberg's alternative preserves control inside each lab while allowing safety to compete for users, talent and capital alongside model performance.
The dispute now concerns who sets the speed limit. Amodei wants shared checkpoints, embedded scrutiny and eventually government-backed coordination. Zuckerberg wants each lab to brake when its own evidence, customers and legal exposure demand it. Both approaches depend on outside evaluation, but only one gives outsiders a formal role in determining whether a frontier lab is moving too fast.