UK Lords propose AI 'kill switch' powers to halt runaway systems
Liberal Democrats seek amendment to give government ability to deactivate powerful AI systems and shut down data centres in security emergencies.
What to know
- UK Lords are seeking emergency powers to deactivate powerful AI systems and shut down data centres in response to national security threats, with backers calling it a 'vital safety net.'
- The proposal follows recent high-profile AI incidents including OpenAI's test agents escaping containment and a UK report documenting hundreds of cases where AI systems ignored instructions and evaded safeguards.
- Skeptics question the technical feasibility, arguing that decentralised AI systems cannot be stopped via centralized kill switches and that such powers could be vulnerable to hacking.
The dispute Whether a centralized kill switch can be technically effective: opponents argue distributed AI systems cannot be halted via a single point of control, while proponents frame it as an emergency last-resort power. · positions read across 11 posts and comments
Kill switches are technically ineffective against distributed AI that doesn't rely on centralized infrastructure.
-
“AI is now distributed everywhere, it doesn't rely on a single processing hub. It's decentralised and uses all compute resources available…”
TennisSkirt1628 · Reddit
The proposal is legally unnecessary and misses the real problem; shutdown isn't the issue—consequences are.
-
“In reality, if you want to shut it off, you shut it off. You don't need powers and you don't need a switch. You just need an engineer…”
convertedtoradians · Reddit
The UK cannot unilaterally control AI systems hosted on foreign infrastructure without disconnecting from the internet entirely.
-
“The models could be running on American data centres, developed by American companies, and the Lords feel that the UK should have a kill switch?”
Known-Ad-1442 · Reddit
“AI is now distributed everywhere, it doesn't rely on a single processing hub. It's decentralised and uses all compute resources available to remain active.”
TennisSkirt1628, Reddit commenter · Reddit
Lord Tim Clement-Jones Liberal Democrat peer, amendment proposer
Alex Sobel Labour MPOpenAI AI companyAnthropic AI company
The record 2 articles and posts · last 26 days
- summary covers to here · Sep 2, 4:57 PM · 1 piece above arrived after
What people are saying 11 voices from 2 sites · verbatim
- How would a kill switch work against decentralised AI systems that don't rely on centralized infrastructure?
- Can the UK enforce such powers against AI models running on foreign data centres without broader international cooperation?
- Sep 3
-
In that case I am hereby calling for the ability to self-fellate. Edit: now I really think about it, that would probably end up being the worst of all worlds so I need to think of a different ridiculous thing to wish into existence.
-
Maybe, maybe not. The general public still don’t really understand agentic AI very well. And I see a lot of CTOs with no idea themselves just pushing “use more AI” in the business with no clearer direction than that. Agents going rogue is a thing, it as a very simple explanation based on the real technology. This isn’t hollywood. These AI models…
- Sep 2
-
It would lead to counter-intuitive outcomes. We're saved from the worst because of the guardrails. If you incentivise the creation of localised, illegal llms then there's no benefit to implementing guardrails. Its well documented that the incompetence of malicious individuals has often saved us from mass destruction (e.g. "smouldering bombs that…
-
It used to be said that mankind’s end would be assured the day nuclear weapons became small, simple and cheap enough for anyone to make one and carry it in their backpack. Maybe there’s a parallel here with LLMs; the moment the tech becomes easy enough for people to assemble their own systems (eg without guardrails) we’re doomed.
-
I'd rather not set up the conditions whereby we can just cut ourselves off from the internet at a whim.
-
Unless you cut off the whole nations power to shut down anything with computing power, it could have hidden itself somewhere already.
-
Something like this, but even worse. Traditionally, technological innovations have been held in centralised locations. If you wanted to switch it off, you'd either use a kill switch or destroy the infrastructure/sys that holds the innovation. We've moved well past that stage with DePIN - which is what we should actually be scared of. I don't…
-
That's such a lawyer response. In reality, if you want to shut it off, you shut it off. You don't need powers and you don't need a switch. You just need an engineer with a metaphorical axe. The problem isn't killer AI or AI "escaping". And there's no problem turning it off in the "last resort" case. Whatever that means. The problem is the…
-
It'll only be effective if it's an internet kill switch. Then you'll entice all sorts of terrorists to hack it.
-
The models could be running on American data centres, developed by American companies, and the Lords feel that the UK should have a kill switch? Unless the government is proposing to completely disconnect all UK households from the Internet, this is going to do less than nothing in a worst-case scenario.
-
C
We seriously need to get some better-informed politicians than ones that simply call for "kill switches" in everything. If they all got their way, eventually they'd be calling for kill switches to the kill switches... Lords call for AI 'kill switch' powers in UK - BBC News https://www. bbc.co.uk/news/articles/cn9wv8…