Mark Zuckerberg: Labs should self-regulate

Mark Zuckerberg argues that AI labs must self-regulate because they already have both the responsibility and strong market and legal incentives to train models safely. He frames trust and alignment as rapidly becoming the most important competitive capabilities, saying users will reject agents that are misaligned or unhelpful and that labs that ignore alignment will fall behind. He also emphasizes liability risk: companies face significant exposure if models cause harm, which creates a direct corporate incentive to prevent dangerous behavior rather than rely on external mandates.

He points to concrete practices Meta adopted as examples: delaying Muse for months to strengthen safety and security, building what he calls robust security foundations, and routinely engaging independent evaluators and advisors. He calls for a larger, more diverse ecosystem of evaluators and urges labs to commit the significant majority of compute toward serving people instead of racing toward recursive self-improvement. He closes by asserting that maintaining the right balance of power among developers and institutions is achievable and essential to building a positive future, placing that responsibility squarely within industry’s control.