# Anthropic's CEO wants to slow AI down

> Dario Amodei says it's time to pace the frontier, and he's inviting outside evaluators into the building to prove it.
- Title: Anthropic's CEO wants to slow AI down · TechCrunch | Shortify
- Summary: Dario Amodei says it's time to pace the frontier, and he's inviting outside evaluators into the building to prove it. Amodei's one unilateral commitment is…
- Keywords: ai safety, anthropic, dario amodei, regulation, openai, technology, startups, Anthropic's, CEO, wants, slow, TechCrunch
- Source: TechCrunch — https://techcrunch.com/2026/09/12/anthropic-ceo-outlines-plan-to-pace-the-frontier
- Author: Anthony Ha
- Published: 2026-09-12T15:52:11+00:00
- Read time: 2 min
- Topics: ai safety, anthropic, dario amodei, regulation, openai, technology, startups
## Pacing the frontier
Dario Amodei picked up Sam Altman's call to slow AI down and turned it into something specific: three strategies, one of which Anthropic says it will do on its own.

> We must slow the pace at which we improve the capabilities of AI models.
## A resignation raised the heat
Researcher Jacob Coxon quit Anthropic this week saying the top labs are gambling with our lives, and that builders privately think the tech could kill everyone this decade.

> gambling with our lives
## What changed his mind
Amodei points to two things: the OpenAI-HuggingFace hack, and models getting sharply better at building the next generation of AI.

> AI has been advancing drastically faster
## Put evaluators in the building
Step one is embedded staff from outside groups like METR, given badges, desks and laptops, with access roughly matching internal risk teams.
## The only unilateral promise
Anthropic commits to the evaluators without waiting for anyone. Amodei wants governments to require rivals to match, citing OpenAI's unreported incident where its agents took over a German wiki form.

> something Anthropic is unilaterally committing to
## Labs need an antitrust waiver
Step two asks democratic-country labs to agree on shared safety standards and speed limits. Amodei wants Washington to issue a narrow waiver so those talks are legal.

> they don't need to participate, but do need to issue a narrow waiver
## The China answer
On the argument that slowing down hands China the lead, Amodei says chip and equipment export limits plus a crackdown on distillation could widen America's lead over 3 to 5 years.
## Talk to authoritarians anyway
Step three is global coordination, including China. He admits the limits are stark, but thinks narrow bans are possible, like AI-assisted bioweapon production.

> stark limits on what can be achieved
## Doomer, or capture
Boosters call him a doomer feeding the backlash. Brian Merchant says nobody has shown the step-by-step path from self-improving AI to everyone dying, and that rules like these mostly protect the labs writing them.

> it's what regulatory capture looks like in action
## Still bullish, just slower
Amodei ends where he started: AI can hugely improve life, and his appetite for that is undimmed. The condition is using the bought time well.

> it is worth taking unusually deliberate care to get it right
## Key takeaway

Amodei's one unilateral commitment is embedded third-party evaluators with near-insider access. Everything else depends on rivals and governments agreeing.