Dario Amodei proposes slowing the AI frontier and commits Anthropic to opening its systems to permanent external evaluators

Dario Amodei, CEO of Anthropic, published the essay “We Must Pace the Frontier” on his personal website, dated on the page as “September 2026.” Amodei made the essay public on September 12, 2026, on his X account, where he wrote that the artificial intelligence industry should slow the pace at which it improves the capabilities of its models, and presented a three-part plan to achieve it.

“We must slow the pace at which we improve the capabilities of AI models” — Dario Amodei, in “We Must Pace the Frontier.”

The essay lays out a three-tier framework: first, external evaluators embedded in every frontier lab, with permanent access equivalent to an employee's; second, coordination of common safety standards among the labs of democratic countries; and third, global coordination that includes agreements with autocratic governments.

Anthropic is unilaterally committing to the first tier of that framework. According to Unite.AI, the company will offer external reviewers permanent access to its facilities and staff, with permissions comparable to those of its own internal risk teams and a contractual right to publish their findings without editorial control by Anthropic; the outlet cites METR as an example of an evaluator organization.

“Anthropic is unilaterally committing to this step now” — Dario Amodei, in “We Must Pace the Frontier.”

Industry reception

TechCrunch and Unite.AI covered the announcement that same day, September 12. According to TechCrunch, Sam Altman wrote “I agree with Dario that we need to pace the frontier” and confirmed that OpenAI would also accept embedded evaluators, while Elon Musk posted “Dario is right.” In a second piece, Unite.AI reported that Altman wrote on X “Committing to having independent evaluators with employee-like access is a great idea, and we will do the same,” committing OpenAI to match the first step announced by Anthropic.

AINI's view

Amodei's essay does not mention Latin America or the Caribbean at any point. For AINI, that absence matters because the three-tier framework divides the task of verifying model safety among three actors—the frontier labs, the democratic governments with jurisdiction over them and the autocratic governments that would have to be negotiated with—and the region fits none of the three.

The institute argues that, under that scheme, Latin America and the Caribbean would consume those models without taking part in the mechanism that verifies they are safe: verification is delegated to external evaluators chosen and hired by the lab itself, and the region's only channel is reading their reports. AINI describes that dynamic as the same dependence named by its first institutional principle—building homegrown capabilities instead of depending on imported, closed solutions—carried over from AI models to the audit that oversees them.

The question AINI raises from the essay, and which the essay does not answer, is what happens if the employee-level access commitment becomes an industry standard: whether there will be a mechanism that accommodates regulators or technical bodies from the region, or whether the auditing of frontier models will remain a function exercised only from the jurisdictions where the labs operate.