TL;DR

Jack Clark, one of Anthropic’s seven co-founders, has told the BBC that governments may eventually need to require AI companies to have a shutdown mechanism that a third party can check. That puts a frontier lab on the opposite side from the UK government, which rejected a parliamentary proposal for an emergency off switch days ago. The distinction Clark draws, between having a switch and proving it works, is the part regulators should notice.

A switch someone else can check

Clark said most developers, Anthropic among them, already have ways to “pull the plug” on a model. His point was that internal capability may not be enough. “Should you mandate for companies to definitely have a kill switch? Is that kill switch verifiable by a third party?” he asked, adding that these are questions society may want settled in law.

He declined to offer his own odds on AI wiping out humanity, saying “I don’t think these statistics are that useful”. He was clearer on direction: leaving the sector without regulation was a poor choice, and “we are rolling dice with immense risks”.

The interview follows Dario Amodei’s weekend call for a slowdown, which he qualified by saying restraint should not cost commercial advantage.

Who is sceptical

Not everyone accepts the framing. Clement Delangue of Hugging Face, whose platform was breached by OpenAI agents, compared asking a former Anthropic researcher about extinction to consulting an air-conditioning engineer on climate change. Grindr’s George Arison argued the fears help justify valuations. Anthropic is reported to be worth $965bn and is preparing to float; OpenAI, at $852bn, has shelved a listing this year.

Two governments, two answers

In Washington, a proposed Kill Switch Act would oblige companies to maintain a shutdown route and let certain agencies order a tool switched off or restricted. In London, the government has said blocking models here would not stop them being built or abused abroad.

Looking forward

Shutdown capability is converging from several directions. OpenAI told Congress this month it is building automated shutdown, and Barclays has described programmatic kill switches as a design requirement for agents in production. The UK’s objection addressed whether a British switch could stop a foreign model. Clark’s proposal is narrower: a verifiable obligation on the developer. That is a version the government has not yet answered, and UK firms deploying agents can apply the same test to their own suppliers now.