As AI systems grow more capable, an old question has become urgent: if a powerful model starts causing serious harm, who has the authority to stop it? On 23 July 2026, two US lawmakers introduced legislation aimed squarely at that question — a bill widely referred to as the AI Kill Switch Act.

What the bill proposes

The legislation, introduced by Representatives Ted Lieu of California and Nathaniel Moran of Texas, would grant the Department of Homeland Security authority to require major AI companies to shut down, throttle, or suspend models the agency judges to be dangerous. Reporting on the bill describes penalties of up to $20 million per day for companies that fail to comply.

Two things are worth stating plainly. First, this is an introduced bill, not law — most bills never pass, and those that do are usually changed substantially along the way. Second, the specific mechanics, including how "dangerous" would be defined and what due process would apply, are exactly the details likely to be contested. Treat the current text as an opening position rather than a settled framework.

Why it is bipartisan

The bill's sponsorship by one Democrat and one Republican is notable in itself. AI safety has been one of the few technology areas where concern crosses party lines, even when the preferred remedies differ. That does not guarantee passage, but it does suggest the underlying worry — that no clear emergency mechanism currently exists — is broadly shared.

The case for an emergency stop

Supporters make a straightforward argument. Other high-risk industries already have emergency powers: regulators can ground aircraft, recall vehicles, halt drug trials, and shut down financial trading. AI has no equivalent. If a widely deployed model were found to be enabling serious harm at scale, the current response would depend largely on the goodwill and speed of the company that built it.

The case against

Critics raise substantial objections, and they are not merely industry self-interest:

  • Definitional vagueness. "Dangerous" is doing enormous work in a bill with severe penalties attached. Unclear standards invite inconsistent enforcement.
  • Technical practicality. Open-weight models, once released, cannot be recalled. A shutdown order reaches companies, not copies already downloaded worldwide — a limitation rooted in the open-model reality that now shapes the industry.
  • Competitive effects. Rules applying only to US firms may shift development elsewhere without reducing global risk.
  • Speech and due process. Government authority to switch off a widely used information system raises serious constitutional questions.

Background: the wider regulatory moment

This bill did not appear in isolation. It arrives amid intense scrutiny of AI's concentration of power — from the enormous infrastructure commitments reshaping the industry to disputes over how models are trained and whose data they use. Governments worldwide are moving from broad principles toward specific enforcement powers, and the practical question everywhere is the same: what does meaningful oversight look like for a technology that ships as software and improves every few months?

Why it matters

Whether or not this particular bill becomes law, it marks a shift in the conversation. The debate has moved past whether AI should be regulated toward the far harder problem of designing mechanisms that actually work — powerful enough to matter, precise enough to avoid abuse, and technically realistic given how AI is built and distributed.

For anyone building with AI, the practical signal is that compliance obligations are coming. Knowing which models you depend on, where they run, and how quickly you could switch is becoming ordinary risk management.

Key takeaways

  • The AI Kill Switch Act was introduced on 23 July 2026 by Reps. Ted Lieu and Nathaniel Moran.
  • It would let DHS require major AI firms to shut down, throttle, or suspend models judged dangerous, with reported penalties up to $20 million per day.
  • It is an introduced bill, not law, and its key definitions remain unsettled.
  • Supporters cite the absence of emergency powers that exist in aviation, medicine, and finance.
  • Critics point to vague definitions, the impossibility of recalling open-weight models, and constitutional concerns.

The bottom line

An emergency stop for AI is easy to describe and genuinely hard to design. This bill is best read as the start of a serious argument about how oversight should work — one that will shape how AI is built and deployed regardless of this particular bill's fate.