Meta CEO Mark Zuckerberg has said artificial intelligence companies do not need an industry-wide agreement to slow development when safety concerns emerge.
In a post on X on Tuesday, Zuckerberg said individual AI labs have both a responsibility and a business incentive to adjust the pace of development when necessary to ensure their models can be trained safely.
“Every lab has the responsibility and incentive to move at the pace required to train its models safely,” Zuckerberg said.
He added that trust and alignment, which refers to ensuring AI systems follow users’ intentions, are increasingly becoming important factors that could differentiate AI models and agents.
“Any lab that doesn’t focus on alignment will fall behind,” Zuckerberg said.
Meta delayed Muse model
Zuckerberg pointed to Meta’s own decision to delay its Muse AI model for several months while the company worked to strengthen its safety and security measures.
“We didn’t call for everyone else to do this before we would. We just did it,” he said.
His comments come amid an ongoing debate over whether development of increasingly capable AI systems should be slowed through coordinated safeguards that can be independently verified.
Some AI companies and researchers have called for stronger external oversight and mechanisms that would allow leading laboratories to pause development if significant warning signs emerge.
Other technology leaders have instead backed approaches involving independent evaluations and accountability for individual AI companies.
Meta AI chief calls for oversight
Meta Chief AI Officer Alexandr Wang also argued that AI companies should strengthen governance and oversight around model development and launches.
Wang said laboratories should have strong governance systems, external evaluators and independent oversight.
He also warned about the risks associated with “recursive self-improvement”, a process in which AI systems could potentially contribute to improving subsequent versions of themselves.
“Racing on recursive self-improvement is one of the riskiest pathways for potential loss of control to powerful models,” Wang said.
Wang added that Meta was directing most of its computing resources towards serving users rather than competing in a race focused on recursive self-improvement.
He said AI development carries significant responsibility and should take place alongside appropriate checks and balances.
