The race to build increasingly capable AI systems may be entering an unusual new phase: some of the companies leading it are now publicly arguing that development needs to slow down.
Anthropic CEO Dario Amodei has called for the industry to “pace the frontier,” proposing a series of measures intended to limit unchecked advances in the most powerful AI models. Perhaps more significantly, OpenAI CEO Sam Altman has publicly backed the idea, saying the issue has become a major topic of discussion inside OpenAI.
Amodei’s concern centres partly on how quickly AI capabilities are advancing, particularly the growing ability of AI systems to contribute to building their successors. His argument isn’t for stopping AI research altogether. Instead, he wants companies developing frontier models to buy more time for safety work as those systems become more capable.
The first proposal is unusually practical. Anthropic plans to allow independent evaluators from third-party organisations such as METR to work inside the company, with access broadly comparable to internal risk assessment teams. These evaluators would monitor whether safety commitments are actually being followed and whether serious incidents are properly reported.
Anthropic says it will adopt the approach voluntarily while calling for governments to require similar arrangements at other frontier AI companies. Altman described embedded evaluators as a good idea and said OpenAI intends to introduce a similar system, although details have yet to be announced.
Amodei also wants leading AI companies in democratic countries to establish common safety standards and agree on limits around unchecked development. That’s considerably more complicated than inviting external evaluators into an office. Competitors coordinating how quickly they develop technology raises obvious antitrust questions, so Amodei suggests the US government could enable narrowly defined safety discussions without necessarily participating directly.
Then there’s China, which has long complicated arguments for slowing American AI development. The fear is straightforward: if US companies deliberately reduce their pace while Chinese laboratories continue accelerating, safety-conscious companies could simply surrender their technological advantage.
Amodei argues that tighter restrictions on advanced chips, semiconductor manufacturing equipment and model distillation could slow China’s progress enough to give US companies room to adopt a more cautious approach. He also wants eventual international coordination, including limited cooperation with China around particularly dangerous applications such as using AI to assist biological weapons development.
The proposals won’t end the argument over AI risk. Critics have questioned whether catastrophic scenarios receive too much attention compared with harms already associated with AI, while others warn that safety regulations designed with today’s biggest AI companies could strengthen their positions against smaller competitors.
That’s the uncomfortable contradiction at the centre of the debate. Anthropic and OpenAI are simultaneously competing to build more powerful AI while arguing that increasingly powerful AI demands greater restraint. Agreeing that the frontier should be paced is relatively easy; deciding who sets that pace, who enforces it and whether competitors will actually accept the same limits will be much harder.
