AI Kill Switch Act (H.R. 9917)

BadIn progressH.R. 9917 (119th)United StatesUnited States
Referred to committeeHigh threat
  1. Introduced
  2. Passed House
  3. Passed Senate
  4. Became Law

Would force the largest commercial AI providers to build centralized controls that can identify and cut off users, restrict use patterns, or shut models down—requirements fundamentally at odds with open weights that people can run privately beyond a provider’s control.

Latest update

Referred to committee

This bill is still moving

This bill is still moving, so there's still time to push back.

Contact Lawmakers

What it does

H.R. 9917 would require covered AI companies to maintain the technical ability to stop inference, terminate user access, suspend an account, user, or use pattern, restrict capabilities, throttle compute, and shut down a covered system. It initially applies to companies earning at least $500 million a year from covered technology and to systems developed with computing power costing more than $100 million; personal, academic, and noncommercial uses are exempt. The Homeland Security secretary could order these controls used after determining that a covered incident occurred. An appeal would not pause the order. The bill also mandates incident reports, preservation of model weights and telemetry, compliance audits, and penalties reaching $20 million per day. DHS would redefine the covered companies and systems annually and is specifically directed to consider how model weights are made available.

What’s at stake

Open-weight models can be downloaded, modified, and run locally without calling a provider’s server. Once weights are released, the original developer cannot reliably terminate a user’s access, stop inference, monitor use patterns, collect telemetry, or remotely shut every copy down. The bill therefore rewards closed, hosted systems and creates a strong incentive for covered commercial developers to withhold weights. Its user-, account-, and use-pattern controls also push providers toward persistent identification and activity monitoring, weakening the privacy that local AI can provide. The initial thresholds are high, but DHS must revisit them every year, creating a route for this control system to spread.

Our take

A mandatory kill switch creates an architecture of centralized control. It keeps AI observable and revocable by providers so the government can intervene. Genuinely open weights and private local computing place that control with users. We oppose requirements that make lawful users identifiable, their activity monitorable, and their access remotely terminable. Specific unlawful conduct should carry specific liability, with emergency powers bounded by courts and general-purpose AI kept free from government-directed control layers.

Timeline

Referred to committee

Latest

Introduced

Sponsor

Ted Lieu (D-CA)

Cosponsors (1)

Nathaniel Moran (R-TX)

Committee

Homeland Security Committee

Bill TrackerLast updated July 24, 2026