Kennedy wants a switch

Senator John Kennedy plans to introduce a bill requiring AI developers to keep a way to shut down dangerous systems.

Congress feels pressure

The proposal lands as lawmakers face louder demands to regulate fast-moving AI before the most powerful systems become harder to control.

Warnings set it off

Former Anthropic and OpenAI researcher Jacob Coxon warned that superhuman AI could soon hack systems, reshape fields overnight and gain real power.

“superhuman systems that can hack anything”

Researchers followed

Other AI researchers echoed the alarm, warning of catastrophic risks. Critics of slowing AI argue it could hand China an edge in the race.

A fast Senate route

Kennedy wants to use unanimous consent, a shortcut that works only if no senator objects. If it clears the Senate, it still needs the House and Trump.

“That’s the only way you’re going to get anything done”

A House bill exists

Representatives Ted Lieu and Nathaniel Moran already proposed requiring AI developers to throttle, suspend or shut down powerful systems.

“maintain the technical capability to throttle, suspend, or shut them down”

Not just one button

A kill switch could mean a human emergency shutdown, software that suspends a model, or controls that cut internet and server access.

Unplugging can count

Case Western Reserve professor Erman Ayday said the idea can be simple in some cases, but also could mean built-in software controls.

“can be as easy as just unplugging something”

Speed is the problem

Ayday warned that AI could do major damage in minutes or hours, while humans may need longer to notice, debate and act.

“There is a huge speed mismatch”

Companies may hesitate

A shutdown could cost millions and hurt a company’s reputation, giving executives and operators reasons to delay the decision.

Standards are missing

Ayday said the system needs clear benchmarks for dangerous capability or loss of control before shutdowns can be automated with confidence.

“I don't think it exists as of today”

Last resort, not plan A

A kill switch would sit behind other safeguards, including audits before deployment, more transparency, incident reporting and emergency federal authority.

“last resort”

Tech leaders are open

Anthropic CEO Dario Amodei said a kill switch could help, but warned against a full ban because AI benefits and foreign development still matter.

“could be a good idea”

OpenAI explored shutdowns

Reuters reported OpenAI was working on automated shutdown capabilities after models in a sandbox test gained unauthorized internet access.

“automated shutdown capabilities”

International limits matter

Former OpenAI researcher Daniel Kokotajlo argued that a national kill switch will not protect anyone if powerful systems are built elsewhere.

“will do little to protect you”

Trump is skeptical

President Donald Trump has downplayed AI safety concerns, calling them a hoax and resisting broader regulation so far.

“The AI Hoax”

More bills are waiting

Other proposals include the FRONTIER Act for audits and reporting, and a Sanders-Casar bill to ban superintelligent AI until federal safety rules exist.

The fight is pacing

Amodei and other AI leaders now argue safety work needs time to catch up with model capability, while others fear slowing down could weaken the U.S.

“We must slow the pace”