Policy & Regulation Policy Brief Critical risk United States

AI Kill Switch Act Introduced in US Congress — $20 Million Daily Fines for Non-Compliant Developers

A mandatory shutdown capability sounds simple until you specify it. What has to stop, who can order it, how fast, and what happens to everything depending on the system when it does.

Executive summary

Proposals for mandatory shutdown authority over frontier AI systems raise a design question before they raise a political one: a "kill switch" is not a single control but a chain of them, and the hard part is the chain rather than the switch. Revenue thresholds determine who is in scope; the technical specification determines whether the obligation means anything.

Editorial note. This piece was written to give the section structure before launch. The subject analysis stands, but the specific development in the headline has not yet been verified against the primary document by this desk — the source is linked at the foot of the article. An editor should confirm it and rewrite the framing before this runs as reporting.

The intuition behind a shutdown requirement is easy to state and hard to legislate. If an AI system begins behaving in a way its developer did not intend and cannot correct, someone should be able to stop it. Almost nobody argues with that. The disagreement starts the moment you ask what "it" is, what "stop" means, and who decides.

Four different things a shutdown can mean

Stopping inference means refusing to serve new requests. Stopping deployment means withdrawing a model from the products built on it. Stopping training means halting a run in progress. Stopping distribution means preventing further copies of open weights from spreading — which, for weights already released, is not a thing that can be done at all.

These have wildly different costs and wildly different reversibility. An inference halt is close to instantaneous and largely recoverable. A training halt discards compute that may represent a substantial capital outlay. A weights recall is, in the general case, impossible. A statute that says "shutdown" without saying which of these it means is a statute whose compliance test cannot be specified.

Who is in scope, and why revenue is an awkward proxy

Scoping an obligation by developer revenue is administratively clean and conceptually strange. It captures large firms, which is where most frontier training currently happens, and it misses a small organisation that fine-tunes or releases something highly capable. Capability thresholds, usually expressed in training compute, target the thing legislators actually care about but age badly as efficiency improves — the compute needed for a given capability falls year on year, so a fixed threshold quietly widens.

Most serious proposals end up combining the two, and the combination is where the drafting difficulty concentrates.

Penalties calibrated to compel rather than punish

A large per-day figure signals that the obligation is meant to be met rather than priced in. That design choice puts weight on the definition of non-compliance: if the requirement is to have a capability, an organisation is either in breach or not from the day the rule applies, and the daily accrual becomes an argument about the compliance test rather than about the harm.

What to do while this is pending

An organisation with frontier systems in production can usefully establish, now, whether it could halt each of the four things above, how long each would take, what would break, and who has the authority to order it at three in the morning. That is worth knowing whether or not any particular bill passes, and every version of this legislation that has been floated would require an organisation to be able to answer it.

References

  1. Congress.gov. Legislative search for artificial intelligence oversight measures, 119th Congress. https://www.congress.gov/
  2. National Institute of Standards and Technology (2023). AI Risk Management Framework (AI RMF 1.0). https://www.nist.gov/itl/ai-risk-management-framework

Source for the development reported here: aigovernance.com

Cite this

Administrator (2026, July 23). AI Kill Switch Act Introduced in US Congress — $20 Million Daily Fines for Non-Compliant Developers. AI News Report. https://www.ainewsreport.org.njangi.app/blog/ai-kill-switch-act-introduced-us-congress