Cole Brennan: Malcolm, hey — rough news week or what, I've been trying to figure out how to even start this one.
Malcolm Reeves: Now, rough is one word for it. I'd say clarifying. The kind of week where someone says something out loud that everyone has been thinking quietly.
Cole Brennan: Right — and the thing that's been nagging me, like I can't shake it, is this: a co-founder of Anthropic went on the BBC and said governments might need to legally require an AI kill switch. Which — okay, fine, that's a headline. But here's what actually stopped me cold. He also said most labs, including Anthropic, already have one internally. So... they have the switch. They have it right now. And he's still saying we need a law.
Malcolm Reeves: That's the question exactly. And the man who said it — Jack Clark, co-founder of Anthropic — said it to Faisal Islam at the BBC on September 15th, 2026. Not anonymously. On record.
Cole Brennan: Why isn't the one they already have good enough? That's — I mean, that's what we're trying to work out today.
Malcolm Reeves: And it's worth noting — Dario Amodei, Clark's own CEO, was calling for a slowdown in AI development that same weekend. So this wasn't a lone statement. Anthropic was speaking with something like one voice, all at once.
Cole Brennan: While Trump was posting on Truth Social calling AI safety concerns a hoax. That same news cycle.
Malcolm Reeves: The same weekend. Which tells you something about the pressure this is being said into — and why the distinction between an internal switch and a verified external one might matter more than it sounds.
Cole Brennan: So that pressure piece — that's actually what I want to pull on. Because the distinction between having the switch and having a *verified* switch, I don't think I fully understand what that gap actually is.
Malcolm Reeves: Right — and here's the plain version. An AI kill switch is a mechanism that shuts the system down. Full stop. Clark says Anthropic already has one. Most labs do. Now — think about a fire suppression system in a building. The sprinklers are already in the ceiling. But before the city lets you open, an independent inspector has to come in, test them, sign off. Not the building owner. Not the contractor who installed them. Someone with no stake in the answer.
Cole Brennan: The sprinklers exist either way.
Malcolm Reeves: Exactly — but one version is certified and one isn't. And the certification matters because the building owner has every reason to tell you the sprinklers work whether they do or not. That's what Clark is saying about internal kill switches. Anthropic controls the hardware, controls the logs, decides what an auditor ever gets to see. So the switch may be real and it may be — I mean, I'd assume it functions. But you can't *verify* that from outside the company.
Cole Brennan: Wait, no. What does Ted Lieu's bill actually *do* then? Like concretely.
Malcolm Reeves: The AI Kill Switch Act — Lieu introduced it in Congress — requires leading AI companies to maintain those shutdown capabilities and, critically, gives the federal government authority to act when a deployed system poses what the bill calls a credible risk of catastrophic harm. It's not just 'have a switch.' It's 'have one that federal oversight can actually trigger.' Brad Carson, the president of Americans for Responsible Innovation, publicly endorsed it. So there's a coalition forming around this specific structure.
Cole Brennan: That's — wait, the *government* can trigger it? Not just inspect it?
Malcolm Reeves: That's the bite in it, yes. Which is — now that's a different thing from inspection. Imagine a federal auditor doesn't just read the checklist. They can actually pull the cord. That's what makes this structurally strange in the way Clark set it up: he's asking for a law that creates external control over something his company already controls internally. It's not a new switch. It's a new hand on the same switch.
Cole Brennan: But a new hand on the same switch — that's exactly where it breaks down for me. Because who's actually holding it? Like, if Anthropic still controls the server logs, still controls what data the auditor receives — the auditor ticks a box that says 'kill switch confirmed' and they're working from a filtered view the whole time.
Malcolm Reeves: Now that's not hypothetical. Picture a specific moment — a Friday night, an Anthropic security engineer is preparing the deployment logs the auditor requested. Seventy-two hours of data. And that engineer flags internally that some entries look anomalous. Legal says: route those to the security hold. The auditor gets the clean version Monday morning, signs off. Nothing technically illegal happened. The switch is still there. The auditor has no idea.
Cole Brennan: That's — yeah. That's the thing I couldn't name.
Malcolm Reeves: And the verification label is real, and it's also theater. Both things are true at once. That's where the third-party concept gets hollow — and I want to be careful here, because the BBC coverage flags this without nailing the sourcing — but a senior Anthropic researcher, reportedly, put the chance that AI kills all humans within the next decade at greater than ten percent. If that's the internal risk framing, verification isn't a compliance checkbox. It's the entire stakes.
Cole Brennan: Wait — greater than ten percent? A senior person inside the company?
Malcolm Reeves: That's what's in the BBC reporting. And Clark himself — when asked for his number — said, and I'm quoting here, 'I don't think these statistics are that useful.' He wouldn't give a figure. But he didn't walk back the underlying concern either. So the internal belief is real, and the verification mechanism is — I mean, it may be structurally incapable of matching it.
Cole Brennan: And Daniel Kokotajlo — former OpenAI — he's basically saying even a functioning verified switch doesn't solve it. Because a single-country mechanism without international agreement isn't technically feasible. You flip the switch in the US and the model's already running in three other jurisdictions.
Malcolm Reeves: Which is the incumbent-embraces-regulation pattern playing out in real time. Anthropic is large enough that mandatory compliance infrastructure raises barriers smaller labs simply can't afford. The moat is real whether Clark intends it or not.
Cole Brennan: And the US-UK split on this — that's where it actually gets strange, and we're going to get into that because it changes everything about whether any of this is enforceable.
Malcolm Reeves: And that split is — now, this is where the ground actually gives way. The UK government rejected the kill-switch legislation outright. Not delayed it. Rejected it. The bill can still move through parliament without government backing, but the government itself said no.
Cole Brennan: Wait — while Congress is moving it forward.
Malcolm Reeves: Simultaneously. Ted Lieu's AI Kill Switch Act advancing in Washington, the UK government pulling the opposite direction. And Clark made this call to the BBC — a British outlet — into a political environment where his own government audience had already said no.
Cole Brennan: So the same news cycle has Clark calling for mandatory oversight, Trump on Truth Social calling safety a hoax — like, his actual words were that the only guardrail needed was a, quote, 'STRONG AND SMART, High IQ! PRESIDENT' — and he named Amodei specifically. Anthropic's own CEO. That's — I mean, the executive branch is actively hostile to the entire premise Clark is asking Congress to adopt.
Malcolm Reeves: Which makes federal enforcement — I want to be honest here, I don't think either of us can say with confidence that it's realistic under current conditions.
Cole Brennan: No. No, I don't think we can.
Malcolm Reeves: And Mustafa Suleyman said his company had been working on new safety guidance for months — so there's industry-wide posturing happening here. But posturing into what structure exactly, when the White House has called the concern a hoax? The law that Ted Lieu writes and the executive branch that enforces it are — those are two very different things.
Cole Brennan: And Clark's own framing — he said the time to act is when there's a small number of players. Meaning now. But the window where the rules get written around companies like Anthropic is exactly the window where enforcement is most politically compromised. That's not ironic, that's — I mean, that might just be structurally broken.
Malcolm Reeves: Picture a Senate staffer, late 2027, drafting the enforcement clause for Lieu's bill — and she has to write the words 'the federal government may act' knowing the agency she's naming has been told by the president that the risk isn't real. What does she actually put in that clause? That's where the proposal lives or dies.
Cole Brennan: That's — I keep thinking about how this whole thing started. Clark said they already have the switch. That was the part I couldn't shake at the top. And now sitting here, the switch being real almost makes it worse. Because the question was never whether it exists.
Malcolm Reeves: It's who controls what the inspector sees. That's the whole thing, and I'm not sure there's a clean answer inside the current proposal.
Cole Brennan: Yeah. Not a verdict — just, that's the question worth watching. Appreciate you working through it.
