Meta CEO Mark Zuckerberg has said alignment and trust are emerging as the key differentiators among AI labs, in a follow-up to the AI manifesto he published last month.
Writing online, Zuckerberg said users would resist adopting AI agents that do not act in line with their intentions, arguing this creates a natural incentive for developers to prioritise alignment. He said labs that fail to focus on alignment risk falling behind competitors, pushing back against calls within the industry to pause capability development until safety research catches up.
On liability and safety delays
Zuckerberg said companies also face significant liability exposure if their models cause harm, giving them further reason to guard against misuse. He cited Meta's decision to delay the rollout of its AI model Muse by several months to strengthen safety and security measures, saying the company did not ask other labs to follow suit before acting. He described the move as part of Meta's routine work rather than an industry mandate.
Independent evaluators
He said engaging outside evaluators and advisors is standard practice in the industry, noting that Meta Superintelligence Labs (MSL) already works with external reviewers in several areas. Zuckerberg said other labs could adopt similar practices and called for a larger, more diverse ecosystem of independent evaluators across the sector.
Compute allocation
On the debate over resource allocation, Zuckerberg said Meta has committed the majority of its computing capacity to serving users rather than pursuing recursive self-improvement of its own systems, and said this approach represents one of the safer paths for developing the technology. He said other labs have the option to make similar commitments.
Zuckerberg's remarks come at a time when AI safety conversations are on the rise, and AI leaders are calling for 'pacing the frontier'. Researchers are reportedly quitting with fear, claiming that 'AI would kill us all.' Zuckerberg reiterated that maintaining an appropriate balance of power is central to building AI safely, describing it as an outcome within the industry's control.
