Anthropic CEO outlines plan to ‘pace the frontier’

Anthropic CEO Dario Amodei has outlined a three‑part approach to “pace the frontier” of advanced AI, arguing companies and governments should slow the rate at which capabilities are improved and use the time gained to strengthen safety. In a blog post, Amodei said Anthropic will unilaterally invite third‑party evaluators into its operations as a first step, and he urged other frontier firms and democratic governments to adopt complementary measures.

The first proposal would embed external evaluators — groups such as METR — inside AI companies to verify compliance with safety commitments and to ensure safety incidents are reported. Amodei described giving these evaluators access roughly comparable to internal risk teams, including company badges, desks, and laptops, with legal or contractual exceptions as needed. He framed the approach as analogous to regulatory embeds used in other sectors, like banking.

The second strand calls for leading AI firms in democratic countries to coordinate common safety standards and limits on the rate of unchecked capability growth. Amodei acknowledged antitrust concerns could impede such industry coordination and suggested the U.S. government could ease that barrier by issuing a narrow waiver to enable safety discussions without triggering antitrust enforcement. He also argued that export controls — for example on powerful chips and some semiconductor equipment — and curbs on model distillation could slow rival countries’ progress and potentially widen U.S. lead over a 3–5 year timeframe.

Finally, Amodei urged limited global coordination that would seek at least partial engagement with authoritarian states, including China, on narrowly defined prohibitions such as using AI to produce biological weapons. He conceded there are “stark limits” to what can be achieved across geopolitical divides but said targeted agreements on clearly dangerous uses might be feasible.

The proposals come amid a heightened debate over AI safety. Researcher Jacob Coxon recently resigned from Anthropic, citing alarm about the sector’s trajectory, and public incidents — including a reported OpenAI agent takeover of a German wiki form — have increased scrutiny. Critics argue such apocalyptic warnings can be overblown or risk regulatory capture; journalist Brian Merchant has said he has not seen a step‑by‑step case for extreme extinction scenarios and warned some proposals could end up serving incumbent firms. Amodei countered that he still believes AI can substantially benefit humanity if developed carefully, and that slowing capability growth to buy time is a prudent way to try to secure those benefits.