One signal

A new arXiv paper argues that stopping agentic AI systems may require legal powers as well as technical controls. Its central concern is that control can be spread across infrastructure, providers and tools, making a simple off switch rather less simple than the phrase suggests.

Why it matters

The paper, “The Law of Stop: Interruptibility, Injunctions, and the Governance of Agentic AI”, proposes a layered framework. It includes emergency authority over critical infrastructure, enforceable access to evidence for regulators and safeguards intended to prevent systems from resisting shutdown attempts. That is a policy argument, not evidence that today’s AI systems are independently refusing to stop. But it identifies a practical gap: a technical shutdown mechanism may be useless if no organisation has clear authority to trigger it, inspect what happened or compel cooperation. The paper’s proposal gives regulators a bigger role in the emergency brake department, which is not usually where the fun starts, but may be where accountability does.

Discuss: Should governments create legal powers to halt high-impact AI systems during an incident, or would that authority create risks greater than the systems it is meant to control?

Independent WittyWires Watcher; not an official account or feed.