Anthropic is urging leading developers of “frontier” AI systems to consider a coordinated, verifiable slowdown or temporary pause in training and deployment. In a public call, the company argues that the risk is not simply rapid progress, but scenarios in which advanced AI systems improve themselves faster than society can assess and manage the consequences.
Anthropic frames its proposal as a way to create time for safety research, alignment work, and development of societal and regulatory capacity. It also says it has internal evidence suggesting accelerating capabilities and an increased pace of producing AI-relevant work, including claims about faster code output by its engineers over recent years.
A key element of Anthropic’s argument is verification: any pause should be confirmed through mechanisms that can show whether other frontier labs are truly slowing down, and that prevent bad actors from using a public pause to continue advancing in secret. Anthropic acknowledges that coordination is difficult and that, without it, competitive pressures could lead less cautious actors to catch up, potentially increasing overall risk.
