Trump Dismisses AI Safety Concerns as Hoax
The story frames Trump’s dismissal of AI safety as a mere political posture, a president being a president, as if the substance of the matter were simply his temperament. But look at what is actually at stake: the knowledge commons of how these systems fail, the shared understanding of their risks, and - in Jack Clark’s proposal - a mechanism, the “kill switch,” that presumes someone must hold the lever. The framing skips the question of who gets to decide what counts as a hoax, and who is positioned to hold the switch once the danger is conceded to be real.
Consider the two men named here. Trump calls safety concern a hoax, which is not a technical judgment but a political one - it clears the field of regulation, and a cleared field is exactly the terrain on which the largest, best-capitalised labs move fastest. Clark, co-founder of Anthropic, answers not by widening the circle of those who assess and govern risk but by proposing a mandatory kill switch, a single point of control installed presumably in the architecture of the frontier models themselves, which is to say in the hands of the companies, like his own, that build them. I do not doubt Clark’s sincerity. But sincerity is not the same as structure, and the structure on offer is this: either no oversight at all, which favours whoever moves fastest and largest, or oversight administered as a switch built into proprietary systems, which favours whoever already holds the wiring. Both paths run through concentration. Neither one asks whether the underlying capability, the weights and the understanding of how to make them safe, might instead be held and audited in common.
This is the old trick dressed in new copper. We are told, on one side, that danger is not real and therefore no fence is needed; we are told, on the other, that danger is so real that only a fence - a specific fence, with a specific gatekeeper - will do. What neither side proposes is what mutual aid has always proposed for genuinely dangerous shared resources: distributed scrutiny, open publication of failure modes, red-teaming done by many hands rather than sanctioned by one boardroom. The history of software security teaches this in miniature - the vulnerabilities caught fastest are the ones exposed to the most eyes, not the ones locked in a vault awaiting a single keyholder’s judgment. A kill switch owned by one firm is not a safety mechanism; it is a title deed with an emergency clause.
Picture the actual mechanics of the switch Clark imagines: a command, issued by whom, verified by whom, triggered on what threshold, disputed by no one because no one outside the company has visibility into the model’s internals to contest the call. That is not oversight. That is a landlord’s spare key, presented as a fire escape.
And here is what the hoax framing actually accomplishes, whatever Trump intends by it: it does not remove the switch from consideration, it only removes the public argument for who should hold it. If there is no hoax and no urgency, there is no political pressure to build oversight as a shared, contestable, publicly auditable process. The field is left open for the kill switch to be installed quietly, as a private feature of private systems, unlegislated, undebated, decided in a lab in San Francisco rather than in any forum where the people who will live under its trigger have a voice. The showman’s denial and the safety advocate’s proposed remedy are not opposites; they are sequential moves in the same enclosure, the first clearing the commons of scrutiny, the second fencing what remains.
I have watched this pattern before, in forests and in fisheries and in the guilds that once held knowledge as a trade secret against the villages that needed it. The claim was always that the resource was too dangerous, too complex, too valuable to be held by the many. The claim was always convenient for the few who ended up holding it. Somewhere a fire brigade decides for itself when to let the house burn, and calls this prudence.