Zuckerberg: AI Safety Depends on Independent External Evaluations, Not Slower Development

On September 16, Meta CEO Mark Zuckerberg shared his views on X about calls to slow AI development. He said users would never choose agents that are not aligned with their values or unable to follow instructions, giving AI labs a strong intrinsic incentive to improve model alignment.

Zuckerberg argued that trust and alignment are quickly becoming key competitive advantages that separate agents and models, and that labs ignoring alignment will be eliminated. He also noted that if a model causes harm, the lab would face significant legal liability, creating another strong incentive to prevent harm. Meta delayed the release of Muse by several months to ensure its safety features and system safeguards, he said, adding that Meta did not call on the industry to follow before taking action itself, but integrated the approach into daily research.

According to Zuckerberg, bringing in independent external evaluators and advisors is an industry best practice. Meta's Superintelligence Lab (MSL) has applied this approach across multiple business areas with notable results. He called for a larger and more diverse evaluation ecosystem, and said that directing most compute toward serving people, rather than blindly racing for recursive self-improvement, is one of the best ways to develop AI safely. Meta has made that commitment, and other labs can follow.