Mark Zuckerberg says AI labs can slow down on their own when safety demands it

Meta CEO Mark Zuckerberg has stated that artificial intelligence labs do not require a unified agreement to temporarily slow down their development when safety is a concern. Speaking on X, Zuckerberg argued that market pressure will make ensuring AI systems align with users' intentions a competitive advantage. Every lab, he said, has both the responsibility and incentive to proceed at the pace necessary to train models safely.

Zuckerberg emphasized that trust and alignment are becoming the most crucial capabilities that will differentiate agents and models, warning that any lab neglecting alignment risks falling behind. He highlighted Meta's decision to delay its Muse model for several months to improve safety and security, noting that this move was not prompted by calls for industry-wide action.

Zuckerberg's remarks come as the debate intensifies over whether AI development should be slowed through coordinated safeguards. While some, like Anthropic CEO Dario Amodei, have advocated for government regulation and mechanisms enabling labs to pause development if red flags appear, others, such as Nvidia CEO Jensen Huang, support independent evaluations and individual liability for AI companies.

Meta Chief AI Officer Alexandr Wang echoed Zuckerberg's stance, emphasizing the need for robust governance, external evaluators, and independent oversight of model launches. Wang pointed out that rushing recursive self-improvement poses significant risks of losing control over powerful models. Meta has committed the majority of its computing power to serving users rather than engaging in a race to recursive self-improvement, a strategy other labs could adopt, according to Wang.