Dario Amodei is no longer arguing only for better AI safety work. The Anthropic chief executive is now calling for frontier labs to slow the rate at which their models become more capable, while giving independent evaluators far deeper access to the companies building them.
The distinction matters because “slow down AI” can mean almost anything.
Amodei’s proposal, published in an essay titled We Must Pace the Frontier, is not a call to stop model training. He argues for slowing capability gains enough to give alignment, interpretability, operational safeguards and testing more time to catch up.
Anthropic is also putting one part of that proposal on the table immediately: permanent embedded third-party evaluators with access closer to internal risk teams than the short external model tests the industry has often relied on.
The harder parts require everyone else to cooperate.
Anthropic is making the evaluator commitment itself
The most concrete part of Amodei’s three-stage framework is also the part Anthropic can control on its own.
He says external evaluators should receive ongoing, employee-like access to Anthropic’s systems, including office access, company devices, relevant workspaces and permissions broadly comparable with internal risk teams. The evaluators would be able to report on safety practices, incidents and model alignment during training, not only inspect a finished model shortly before release.
Anthropic also says those reviewers should be able to publish key findings without the company exercising normal editorial control, subject to narrow security, legal and confidentiality limits.
This goes further than simply promising more transparency.
A frontier lab can publish a safety framework and still decide what outsiders get to see. Embedded evaluators create a separate set of eyes with a better chance of checking whether the lab is following its own rules when commercial pressure is high.
OpenAI has already argued publicly that trusted third-party evaluations are an important part of frontier-model safety. Reuters reported that Sam Altman backed Amodei’s broader pacing argument and said OpenAI would match the independent-evaluator commitment. Elon Musk also expressed support.
Support is useful. It is not the same thing as an enforceable system.
Actual pacing needs coordination
Amodei’s second step is much more difficult: frontier AI companies in democratic countries agreeing common safety standards and limits on unchecked capability growth.
He explicitly acknowledges the legal problem. Competitors discussing how quickly they will develop technology can run into antitrust concerns, which is why he argues for government mediation or a narrow route that allows safety coordination.
Voluntary coordination also has a basic incentive problem.
If one company slows a training programme while a rival keeps pushing, the cautious company risks losing customers, talent, investor confidence and technical leadership. The pressure becomes even sharper when companies are spending extraordinary sums on chips, data centres and research teams built around the assumption that capability gains will continue quickly.
Amodei’s answer is partly regulatory: common rules can remove the penalty for moving more carefully because every major frontier lab faces the same constraint.
No such binding framework exists today.
Altman and Musk backing the principle does not create an industry speed limit. It does not bind Meta, Google DeepMind, Chinese labs or future entrants. It also leaves unresolved what “slow” would mean in practice: fewer training runs, limits tied to model behaviour, restrictions on AI-assisted AI research, capability checkpoints, or something else.
Amodei himself presents several possible mechanisms rather than one settled standard.
The international problem is harder again
The third step is global coordination, particularly between the United States and China.
Amodei argues that any meaningful long-term pacing system eventually has to account for frontier development outside US companies. He is also clear about the conflict: slowing American labs while a geopolitical rival does not slow creates a strategic incentive to defect.
His preferred framework therefore combines slower development with measures intended to preserve a US and allied technology lead, including stronger controls on advanced chips and efforts to prevent model theft or unauthorised distillation.
This is a very different proposition from a simple global pause.
The plan is trying to balance two goals that pull against each other: reduce the speed of frontier capability growth, while avoiding a situation where the side that moves more cautiously gives away a decisive technical advantage.
The more international the proposal becomes, the more verification matters and the less voluntary goodwill can carry it.
Markets reacted faster than policy did
Investors did not wait for any of this to become binding.
Reuters reported sharp falls across AI-linked semiconductor and technology shares on Monday as markets digested the calls for slower frontier development. Asian chipmakers and technology groups were among the names hit, with the weakness carrying into European technology shares.
The reaction is understandable. A genuine slowdown in frontier-model development could change assumptions around data-centre construction, accelerator demand and capital spending.
One trading day cannot tell us whether that will happen.
There is no industry moratorium. Anthropic’s embedded-evaluator commitment, by itself, neither reduces chip purchases nor halts training. Wider pacing remains a proposal whose strongest parts depend on cooperation not yet in place.
Treating the market sell-off as proof that long-term AI demand has weakened would run well ahead of the evidence.
The real change is inside the labs
The interesting shift is not that AI executives are warning about risk. Frontier labs have published safety frameworks, evaluation plans and policy proposals for years.
The change is that one of the leading labs is now saying capability growth itself may need to be managed, and is attaching that argument to a concrete oversight mechanism it says it will implement regardless of whether competitors follow.
Embedded evaluators are the part to watch first because they can exist before governments agree a new regime or rivals agree a common pace.
If they receive the access Anthropic is promising and can publish uncomfortable findings, the arrangement could make a frontier lab’s internal safety claims more verifiable than they are today.
The bigger pacing project remains much less certain.
Voluntary support from other chief executives can start a conversation. Turning that conversation into common limits without collapsing under competition law, commercial incentives or geopolitical rivalry is the real test of whether “pacing the frontier” becomes an operating constraint rather than another safety principle the industry agrees with in public.
