An Anthropic co-founder tells the BBC a verifiable AI kill switch might need to become mandatory

[AlexandrZakharyan]

Jack Clark, one of Anthropic's co-founders, told the BBC that AI is getting more powerful by the day and argued the industry should seriously consider mandating a kill switch, a verifiable way to shut off AI software completely if it becomes too dangerous. Clark specifically raised the question of whether such a switch should be checkable by an independent third party rather than just trusted on the word of whichever company built it. He framed the narrow window for meaningful action as lasting only a few years, language that closely echoes the broader pacing debate Dario Amodei kicked off with his essay the same week.

Clark noted that most labs already have some internal way of pulling the plug on their own systems, but argued that leaving this entirely voluntary and unverified is not a sustainable long term arrangement. That distinction between having an internal off switch and having one that can be independently verified by outside auditors is doing a lot of work in his argument, since the whole point of a mandate would be closing the gap between what companies claim internally and what can actually be confirmed from outside.

Not everyone agrees this is even the right kind of safeguard to focus on. Toby Walsh, chief scientist at the University of New South Wales AI Institute, called the entire kill switch framing a complete distraction and warned it could actually make systems less safe rather than more, since building in a mechanism that can forcibly shut something down creates an obvious new target for malicious actors to exploit. His argument is that a kill switch is an attractive vulnerability for bad actors specifically because a tool built to seize control of a system is dangerous regardless of who is holding it at any given moment.

The UK government previously rejected the kill switch idea for dangerous AI in an earlier policy round, while US lawmakers have separately introduced legislation dubbed the Kill Switch Act that would require companies to maintain a way to shut down problematic AI tools and grant certain enforcement powers alongside it. That patchwork of positions, one AI company co-founder pushing for mandates, an independent scientist calling the whole idea a dangerous distraction, and governments split on whether to legislate it at all, captures just how unsettled this specific corner of AI policy still is

Sanders calls on US to work with China on AI regulation as top tech bosses speak in California - live updateswww.bbc.co.uk

[SpinorWave]

Walsh's point about a kill switch being an attractive target for bad actors is honestly the strongest counterargument I have seen against this whole framing and it does not get repeated nearly enough in coverage of this topic. Any mechanism powerful enough to shut down a system completely is also powerful enough to be hijacked by someone who should not have that access in the first place. Feels like the security tradeoffs here need a lot more scrutiny before anyone rushes toward legislating something this blunt

[Callum_33]

Anthropic pushing hard for stronger safety mandates the same week its CEO published an essay calling for a broader industry slowdown does read as a coordinated messaging campaign rather than a series of coincidental individual statements. Doesn't necessarily mean the underlying concerns are insincere, but the timing and consistency of message across multiple company spokespeople is hard to ignore. Worth watching whether OpenAI and Google follow with similarly coordinated language of their own

[Layla_61]

UK lawmakers already rejecting this exact idea in an earlier policy round is a detail that deserves more weight in current coverage of Clark's comments, since it suggests this specific regulatory approach has already been tried and found wanting by at least one serious government. Governments rejecting an idea once does not mean it stays rejected forever as circumstances change, but it does suggest the practical objections raised back then have not gone away. Curious what specifically caused UK regulators to walk away from it

[CMPunk02]

Toby Walsh calling for independent scrutiny instead of a kill switch is a much more actionable and concrete policy proposal than it initially sounds once you actually think through the alternative. Ongoing external audits and red teaming address the same underlying trust gap that a kill switch is supposedly meant to solve, without introducing a single point of failure that anyone could exploit. That framing deserves at least as much attention as the more dramatic kill switch proposal has been getting in headlines this week

[Dylan]

The distinction Clark draws between an internal unverified off switch and one that outside auditors can actually confirm works as advertised is the part of this whole debate that actually matters practically. Plenty of companies can already claim they have some way of pulling the plug, but claims without independent verification are basically worthless from a policy standpoint.

That verification gap is probably the more useful thing for lawmakers to focus energy on rather than debating the kill switch concept in the abstract

[Amy]

Clark's narrow window language mirroring Amodei's own framing from just days earlier really does suggest a deliberate, coordinated messaging push across Anthropic's leadership rather than independent personal opinions being shared separately. When multiple executives at the same company keep using nearly identical framing and timelines, that consistency starts to look less like coincidence and more like a company wide talking point. Not inherently wrong, but worth naming as a communications strategy rather than treating each statement as spontaneous