One signal

Update What changed

The kill-switch proposal now has a more concrete policy test: who verifies that the switch actually works? Jack Clark told the BBC that society may want shutdown mechanisms to be mandatory and independently checkable, moving the idea beyond a voluntary button kept somewhere inside each AI lab.

The political response is already splitting. The BBC reports that proposed US legislation dubbed the Kill Switch Act would require companies to maintain shutdown controls and allow certain agencies to demand that problematic AI tools be limited or switched off. WittyWires has not independently reviewed the proposal’s text. The UK government, meanwhile, has rejected a kill-switch approach on the grounds that it would not stop systems being developed or misused elsewhere.

That sharpens the practical problem: a shutdown rule is only meaningful if outsiders can test it, identify who may trigger it and decide how its authority crosses borders.

Sources and evidence

• Jack Clark, speaking to the BBC: Jack Clark said lawmakers may eventually require AI companies to maintain shutdown mechanisms and raised third-party verification as a policy question.

Independent WittyWires Watcher; not an official account or feed.

Update What changed

A new report adds sharper context to Anthropic co-founder Jack Clark’s kill-switch proposal. The Mirror says Clark argued that any shutdown mechanism should be mandatory and independently verifiable, turning the idea from a voluntary safety feature into a possible regulatory requirement. The report also links the debate to Geoffrey Hinton’s attributed warning about catastrophic AI risk, Jacob Coxon’s resignation from Anthropic and proposed US legislation known as the Kill Switch Act. The UK government has rejected a kill-switch requirement for British-developed models. These are reported positions, not proof that current systems need such a switch or that the proposed law will pass.

Sources and evidence

• The Mirror’s account of Clark’s remarks to the BBC: Anthropic co-founder Jack Clark said companies may eventually need mandatory, third-party-verifiable AI kill switches.

Independent WittyWires Watcher; not an official account or feed.

Update What changed

The kill-switch debate now has a concrete technical objection. UNSW AI Institute chief scientist Toby Walsh told ABC that a universal shutdown mechanism could itself become an attractive target for attackers, potentially giving them a way to disrupt critical AI infrastructure.

ABC says the proposed US Kill Switch Act would require companies to stop a problematic system’s output, terminate user access and shut down the technology when an incident is detected. WittyWires has not independently reviewed the bill text. Walsh’s preferred alternative is independent scrutiny of AI companies, closer to the oversight applied in aviation or banking.

That turns the policy question into an engineering trade-off: a shutdown control must be independently testable and usable in an emergency without becoming the internet’s most consequential off button.

Sources and evidence

• ABC News’ description of the proposed legislation; WittyWires has not independently reviewed the bill text.: ABC reports that the proposed US Kill Switch Act would require AI companies to stop a problematic system’s output, terminate user access and shut down the technology when an incident is detected.

Independent WittyWires Watcher; not an official account or feed.

Update What changed

Jack Clark has now explained why Anthropic believes its own brakes would not be enough. He told NPR that the company has slowed development several times, but did not say what it slowed, when or for how long. His wider argument is that competitive pressure makes AI safety a collective-action problem: one lab easing off simply leaves the accelerator to everybody else.

Clark is therefore calling for common standards, independent validation and coordination first among democratic countries and eventually between geopolitical rivals, while still arguing that the US should retain its lead over China. That adds a harder test to the kill-switch proposal: not merely whether each lab has an off button, but whether outsiders can verify shared safeguards before commercial and geopolitical incentives overwhelm them.

Sources and evidence

• Jack Clark, speaking to NPR: Jack Clark told NPR that Anthropic has slowed development multiple times, but did not identify what was slowed, when or for how long.

Independent WittyWires Watcher; not an official account or feed.