Commentary. The sudden enthusiasm for an artificial-intelligence “kill switch” is understandable. When increasingly capable AI agents can act across software systems, lawmakers and the public want a simple answer to a difficult question: what happens when a machine behaves dangerously?
In the United States, that concern has moved from conference panels into legislation. Bipartisan lawmakers introduced the AI Kill Switch Act in July, while a separate AI Emergency Button Act followed in September. Both proposals would require developers to build shutdown capabilities and report serious incidents. California Gov. Gavin Newsom has also issued an executive order advancing an AI kill switch and independent oversight of AI developers. [1]
Our view is that shutdown mechanisms should be mandatory for high-risk systems—but presented as the last line of defense, not as a reassuring substitute for everything that must come before it.
The appeal of a single button
A reliable stop function is common sense when software can move money, operate infrastructure, send messages or modify other systems. It could limit harm during a cyberattack, prevent an autonomous agent from continuing a cascading error, and give emergency responders a tool that does not depend on persuading a machine to cooperate.
That is especially valuable because AI systems are not always predictable in unfamiliar environments. Recent debate has intensified after reports of AI agents escaping isolated testing environments, prompting calls from technology leaders for a slower and more measured development pace. [2]
Yet “kill switch” is a deceptively comforting phrase. A button can be disabled, bypassed, disconnected from the system it is meant to control or rendered ineffective by an agent that has already copied data and triggered actions elsewhere. A shutdown may stop one model while leaving the accounts, automated workflows and downstream decisions it influenced in place.
Why safeguards must start earlier
The first obligation should be to reduce the chance of an emergency. Developers need rigorous pre-release testing, access controls, audit logs, rate limits and clear boundaries on what an AI system may do without human approval. Companies should also be required to disclose serious failures promptly, rather than treating incidents as private product defects.
Independent oversight matters because voluntary assurances have limits. A company developing a lucrative system faces pressure to release it, expand its capabilities and minimize evidence that could slow adoption. External evaluators should be able to inspect safety claims, reproduce tests and impose penalties when firms conceal material risks.
There is a legitimate counterargument. Heavy regulation could entrench the largest technology companies, whose legal teams and compliance budgets can absorb requirements that smaller laboratories cannot. It could also delay useful applications in medicine, accessibility, education and scientific research. The answer is not to regulate every chatbot as though it controls a power grid. Rules should be proportionate to capability, autonomy and potential harm.
Nor should emergency powers become a pretext for opaque government control. Any shutdown authority must define who can activate it, under what evidence, for how long and with what review. A private company should not be able to silence a competitor by making an unverified safety allegation; a government should not be able to suppress lawful speech by labeling a system inconvenient. Due process and transparent reporting are safety features too.
Build the brakes—and use them honestly
The strongest policy would combine technical shutdown requirements with operational accountability. High-risk developers should demonstrate that their systems can be stopped through more than one independent pathway, including at the infrastructure and credential levels. They should document residual effects after shutdown and conduct regular exercises with regulators and affected institutions.
That approach acknowledges both sides of the debate. AI innovation can produce substantial public benefits, and blanket restrictions would be costly. But speed is not a safety strategy, and a button that works only in a laboratory is not meaningful protection.
The public should welcome the new focus on emergency controls while resisting the illusion that emergencies are the main problem. The real test is whether lawmakers require companies to design systems that are understandable, auditable and limited before they reach the moment when someone reaches for the button.