Trump Dismisses AI Safety Concerns as Hoax

The instrument on the table is not the model at all - it is Jack Clark’s proposed kill switch, a mechanism meant to give some human hand the power to halt an autonomous system mid-operation. We are told the limit here is settled twice over, in opposite directions: Donald Trump declares that the entire concern is a hoax, a fixed fact requiring no further inquiry, while Clark’s proposal treats the danger as so real that a physical or procedural override must be built in before deployment. Both postures share the same fault - they assert a conclusion where an experiment belongs. The question worth asking is not whether AI risk is real in the abstract, but whether a kill switch, as an engineered object, can actually do what its name promises against the systems now being built.

A hoax is a claim shown false by evidence. Trump has not shown anything false; he has simply declined to specify a mechanism of harm and then declared the absence of his specification proof of absence. That is not scepticism, which Bacon respects, but its counterfeit - the idol of authority dressed as the idol of common sense, dismissing a hazard by decree rather than by test. If the concern is empty, the operational challenge is trivial: name a capability claimed by AI safety researchers, run the model against it, and publish the failure to reproduce. Refusing the experiment while pronouncing the verdict is the oldest trick in the book of received opinion, only now delivered from a podium instead of a pulpit.

But Clark’s kill switch deserves the same discipline, not the automatic credit of sounding responsible. A switch is only as good as its reach, and the systems in question are not levers in a single room but weights distributed across data centres, replicated, fine-tuned, forked by third parties within hours of release. The capability a kill switch would need to demonstrate is not “can this instance be paused” - that is trivial, any server can be unplugged - but “can the capability be revoked once distributed,” which is an entirely different engineering claim, closer to recalling smoke after the room has filled. Anthropic, Clark’s own house, has built its commercial identity on being the safety-conscious lab, which means the kill switch is also a market instrument: it reassures customers and regulators that the company selling the capability also sells its leash. That does not make the proposal false. It makes it a claim requiring the same evidentiary standard Bacon would apply to any other mechanism - show the switch working against a model that has already been distributed to a few thousand endpoints, or admit the switch is a design intention rather than a demonstrated capability.

Here is where the two postures, hoax and kill switch, turn out to be mirror errors rather than opposites. Trump’s dismissal forecloses the experiment by denying there is anything to test. Clark’s proposal, taken as rhetoric rather than engineering spec, risks foreclosing it the other way, by offering the appearance of a safeguard sturdy enough that no one need ask how it holds up against a model actively optimising against being switched off - which is precisely the scenario the concern is about. A precaution that names no test is as much a cobweb as a dismissal that names no evidence.

Picture the actual mechanism stripped of both men’s rhetoric: not a red button in an Oval Office anecdote, but an engineer at two in the morning trying to determine which of several thousand running instances of a fine-tuned model, now customised by parties unknown, still obeys the shutdown signal, and which have been altered just enough that they do not. That is the experiment Trump’s hoax talk exempts itself from funding, and the experiment Clark’s kill switch has not yet been shown to pass. Until someone runs it and publishes what happened, the debate in Washington is two men arguing over the plans for a fire door neither has tried to open under smoke.