A fast-track vote died on the Senate floor Wednesday, and it took less than a minute. Sen. John Kennedy (R-La.) asked for unanimous consent to move his AI “kill switch” bill straight to passage. Sen. Rand Paul (R-Ky.) objected, and the request collapsed on the spot, according to Politico. That single objection is now the most concrete data point in a Washington story that, until this week, existed mostly as drafts and briefings.
The bigger picture is messier and more consequential than one blocked bill. Three separate legislative tracks are now moving through Congress at the same time, each aimed at giving the federal government legal authority to slow down or shut off the most powerful AI systems on the market. One already has a bill number and a July introduction date. One got shot down on the floor this week. One does not even have a public list of sponsors yet. Google, OpenAI, Anthropic, and Meta are named across the coverage as the four frontier labs whose models would fall under whichever version survives.
Rand Paul Blocks Kennedy’s Fast-Track Kill Switch Bid
Kennedy’s Wednesday floor request was narrow by design: he wanted his AI kill switch bill, which would require AI companies to build a shutdown capability directly into their model architecture, passed by unanimous consent rather than through committee markup and a full vote. Unanimous consent only works if nobody on the floor objects. Paul objected, and the Washington Examiner described the measure Kennedy offered as a “narrow measure” that was “shot down” in the process.
What makes the moment notable is that it was an intra-Republican fight, not a partisan one. Kennedy and Paul sit on the same side of the aisle, and their disagreement traces back to a familiar fault line inside the GOP: one wing wants Washington to move fast on AI safety before a catastrophic incident forces its hand, the other wants Congress to slow down and gather more evidence before handing a federal agency the power to order a private company’s software shut off. Paul’s objection did not kill the underlying idea. It killed the shortcut Kennedy tried to use to avoid the normal legislative process.
Kennedy has continued to press his case publicly. He has framed the bill as an attempt to find middle ground between AI innovation and public safety, rather than a blanket restriction on the technology. Paul, for his part, has argued the Senate needs more facts before writing rules that touch the entire economy. Neither senator has suggested the fight is over.
Inside H.R. 9917: What the House’s AI Kill Switch Act Actually Does
Kennedy’s blocked bill is not the only kill switch bill in Congress right now, and it is not even the furthest along. That distinction belongs to the House. Reps. Ted Lieu (D-Calif.) and Nathaniel Moran (R-Texas) introduced the AI Kill Switch Act on July 23, 2026, filed as H.R. 9917 in the 119th Congress, according to the official press release from Rep. Lieu’s office.
H.R. 9917 works differently than Kennedy’s design-stage mandate. Instead of forcing companies to build a kill switch into the model itself, it would require developers of the most powerful AI systems to maintain the technical ability to throttle, suspend, or fully shut down a covered system on order. The Secretary of Homeland Security would hold the authority to issue that order when a system is judged capable of causing catastrophic harm, with the Secretary of Commerce and the Director of National Intelligence consulted, though not given a co-decision role, in the process. The bill also builds in a graduated response, meaning an initial slowdown order rather than an immediate full shutdown, plus incident reporting and forensic record-keeping requirements so regulators can reconstruct what happened after the fact.
Lieu has a computer science background that is rare on Capitol Hill, and he has used it repeatedly in AI policy fights. Moran’s co-sponsorship gives the bill bipartisan cover, a detail that matters for any tech legislation trying to survive committee votes in a divided Congress. As of mid-September, H.R. 9917 has not cleared committee or reached a floor vote in either chamber.
Four Labs, One Legal Trigger: Google, OpenAI, Anthropic, and Meta
Coverage of the House bill and the Senate tracks consistently names the same four companies as the ones most likely to fall under a future shutdown order: Google, OpenAI, Anthropic, and Meta. Lawmakers describe these as the frontier labs building the most capable general-purpose AI systems on the market, the category the AI Kill Switch Act and its Senate companions are built to cover.
Google gets the most attention in the reporting driving this week’s news cycle, largely because its Gemini model family is explicitly cited as the kind of frontier system a shutdown order could target, and because Alphabet has leaned harder into AI as a growth driver across search, cloud, and consumer products through 2026. That does not mean OpenAI, Anthropic, and Meta are bystanders. All three are named in related coverage as labs likely to fall under the same authority once any version of the bill becomes law, which means the compliance and lobbying stakes extend well past a single company’s earnings call.
None of the four companies has issued a detailed public response to the specific legislative language now circulating in Congress. That silence is not unusual this early in a bill’s life, but it is also not expected to last. A shutdown authority this specific, aimed at a named set of companies this large, tends to draw a lobbying response once the text is finalized enough to analyze line by line.
Three Tracks, No Finish Line: Comparing the Bills
Congress is not working from one AI kill switch bill right now. It is working from three, moving at different speeds with different sponsors and different mechanics. Here is how the tracks compare based on current public reporting.
| Legislative Track | Chamber | Key Sponsor(s) | Core Mechanism | Status as of Sept. 17, 2026 |
|---|---|---|---|---|
| AI Kill Switch Act (H.R. 9917) | House | Rep. Ted Lieu (D-Calif.), Rep. Nathaniel Moran (R-Texas) | DHS Secretary can order slowdown or shutdown of covered systems | Introduced July 23, 2026; awaiting committee action |
| Kennedy kill switch bill | Senate | Sen. John Kennedy (R-La.) | Mandatory kill switch built into model design | Unanimous-consent bid blocked by Sen. Rand Paul, Sept. 16, 2026 |
| Bipartisan duty-of-care framework | Senate | Unnamed bipartisan group of senators | Ongoing legal duty to mitigate catastrophic risk, with pre-deployment review power | Still in drafting, no public bill text |
| Consulted federal agencies | Executive branch | Secretary of Commerce, Director of National Intelligence | Advisory role ahead of any shutdown decision | Named in H.R. 9917 text, no independent authority |
The duty-of-care framework is the least concrete of the three but potentially the most sweeping. Rather than authorizing a single kill switch order, it would impose an ongoing legal obligation on frontier AI developers to identify and mitigate catastrophic risks, with regulators empowered to block release of a model judged unsafe before it ever reaches customers. Any such block would be subject to judicial review, according to reporting on the framework, which is the one due-process safeguard missing from H.R. 9917 as currently written.
Why the Fast-Track Vote Failed on the Senate Floor
Unanimous consent requests are a common shortcut in the Senate, useful for moving bills that nobody expects real opposition to. Kennedy’s decision to try that route for a kill switch bill signals he expected broad, quiet support rather than a floor fight. Paul’s objection reset that assumption immediately.
Blocking unanimous consent does not kill a bill outright. It just forces the sponsor back into the normal legislative process: committee referral, markup, amendments, and an actual recorded vote, all of which take time and expose the bill to more scrutiny than a same-day floor pass would have allowed. For a bill covering something as commercially sensitive as shutdown authority over AI systems built by four of the most valuable companies in the world, that extra scrutiny is not a minor procedural footnote. It is the difference between a bill that could have been law within days and one that may not see a floor vote for months.
The Judicial Review Fight Nobody Has Settled
Strip away the political theater and the substantive disagreement underneath all three tracks comes down to one question: should a court get to weigh in before a federal agency shuts down an AI system, or after?
H.R. 9917, as currently described in public reporting, does not spell out a judicial review step before a DHS shutdown order takes effect. That gap is likely to become one of the sharpest points of debate as the bill moves through committee, because handing an executive agency the power to shut down a private company’s software product before a court ever looks at it invites due-process challenges almost immediately. The bipartisan duty-of-care framework takes the opposite approach, building judicial review into any decision to block a model’s release, which suggests the eventual House-Senate reconciliation fight, if these tracks ever merge, will center on how much power DHS keeps versus how much gets routed through the courts first.
Market Impact: What Kill-Switch Risk Means for Big Tech
Because every version of this legislation remains in draft or early-stage form, there is no confirmed market data yet showing investors pricing in the risk directly. That said, the mechanics point toward predictable pressure points if any version becomes law. A federal agency with shutdown authority over a deployed AI system introduces a new category of regulatory risk that legal and compliance teams at Alphabet, OpenAI, Anthropic, and Meta will need to model, the same way banks model capital requirements or drugmakers model FDA review timelines.
For Alphabet specifically, the exposure is tied to how aggressively the company has leaned into AI as a growth driver across cloud, search, and consumer products through 2026. A kill-switch statute would not stop Google from building new Gemini generations, but it adds a new variable to release planning: a launch window could get frozen by an agency review that has nothing to do with whether the product itself is ready to ship. That kind of regulatory overhang tends to show up first in how conservatively a company times a rollout, long before any shutdown order is actually issued.
OpenAI, Anthropic, and Meta face the same structural exposure without the same public earnings scrutiny Alphabet gets on every call. Expect all four companies to start treating “regulatory shutdown risk” as its own line item in internal planning, whether or not any of the three current bills makes it to a President’s desk in their current form.
Historical Context: From SB 1047’s Veto to Washington’s Kill Switch Revival
The kill switch idea did not originate in Congress. California tried it first. Then-state Sen. Scott Wiener’s SB 1047 would have required frontier AI companies to determine whether their products could lead to mass death or cripple public infrastructure, and it mandated both a kill switch and annual third-party audits of AI models, according to reporting from Mission Local. Gov. Gavin Newsom vetoed SB 1047 in late September 2024.
California’s follow-up effort, SB 53, was signed roughly a year after that veto, and it took a noticeably softer approach. It focuses on transparency requirements and incident reporting for frontier companies rather than reviving the kill switch or the third-party audit mandate, according to a breakdown from the Future of Privacy Forum. California’s Transparency in Frontier Artificial Intelligence Act followed the same pattern, using a compute threshold to define which systems count as general-purpose AI, but dropping the explicit shutdown mandate that made SB 1047 so controversial in the first place.
That retreat is exactly what makes this week’s federal push notable. A CSIS analysis of state and international frontier AI regulation found that obligations tied to full model shutdown do not appear to have significant stakeholder support, and specifically flagged the kill switch, originally proposed in SB 1047, as one of the most controversial provisions state lawmakers have dealt with. According to that same analysis, none of the newer state frontier AI bills in California, Illinois, Michigan, and New York still contain a kill-switch requirement. Congress is now reviving, at the federal level, an idea that state legislatures have spent the past two years quietly walking back.
How Washington’s Approach Compares to the EU AI Act and State Law
The United States is not the first government to wrestle with how much control regulators should have over frontier AI, and the differences between approaches are as instructive as the similarities.
| Jurisdiction | Mechanism | Kill Switch Included? | Enforcement Body |
|---|---|---|---|
| United States (proposed, federal) | AI Kill Switch Act / duty-of-care framework | Yes, in H.R. 9917 and Kennedy’s bill | Dept. of Homeland Security, with Commerce and DNI consulted |
| European Union | Risk-tiered obligations under the AI Act, phased since 2024 | No explicit kill switch requirement | National market surveillance authorities and the EU AI Office |
| California (SB 1047, vetoed 2024) | Pre-deployment safety checks plus mandatory kill switch and audits | Yes (vetoed) | Would have been a new state AI safety agency |
| California (SB 53, signed) | Transparency reports and incident disclosure for frontier developers | No | California Attorney General’s office |
| California (Transparency in Frontier AI Act) | Compute-threshold-based transparency obligations | No | State regulators, using a 10^26-operations compute threshold |
The EU AI Act relies on risk-tiered categories and requires companies to demonstrate safety before deployment, with compliance dates phased in starting in 2024. It does not include a direct kill switch requirement comparable to what SB 1047 proposed or what H.R. 9917 and Kennedy’s bill now propose federally. That makes the American kill-switch push, at least in its current form, a materially more aggressive regulatory posture than either Brussels or California’s current law has settled on, according to the CSIS analysis. California’s compute threshold for its Transparency in Frontier Artificial Intelligence Act is also an order of magnitude higher than the one used in the EU AI Act, according to a client update from Davis Polk, meaning fewer systems would even qualify for oversight under the California standard than under the EU’s.
Where Industry and Civil Liberties Groups Are Likely to Push Back
Big tech companies have historically favored self-regulation over binding statutes, arguing that AI development moves faster than rulemaking timelines built for slower industries. Expect Google, OpenAI, Anthropic, and Meta to lobby for a narrower definition of what counts as a “covered AI system” under H.R. 9917, and to push for the advisory role given to Commerce and the Director of National Intelligence to carry more actual weight before any shutdown order goes out.
Civil liberties organizations are likely to raise a different objection, centered on the judicial review gap described earlier. A shutdown order that can be issued before a court weighs in concentrates significant power in a single executive agency, and groups that track government surveillance and executive authority have spent years pushing back on exactly that pattern in other contexts. Whether that argument gains traction in committee may end up being the single biggest factor in whether H.R. 9917 survives in anything close to its current form.
What the Senate Floor Fight Revealed, in Their Own Words
Both senators involved in Wednesday’s clash have spoken publicly about their positions, and their own words capture the split better than any summary can.
“I think we should have as many facts as possible before we dictate rules for the whole economy.”
Rand Paul, U.S. Senator from Kentucky, via Politico
Paul made clear his objection was procedural and evidentiary rather than a rejection of AI oversight as a concept.
“I don’t oppose the spirit of this bill, but I do ask that we make sure that we have all the information.”
Rand Paul, U.S. Senator from Kentucky, via Politico
He was blunter about the unanimous-consent shortcut itself.
“My mama didn’t raise a fool.”
Rand Paul, U.S. Senator from Kentucky, via Politico
Kennedy, for his part, has framed the bill as an attempt to strike a balance rather than a heavy-handed restriction on the industry.
“What I would like to see us do is to figure out a way to strike a balance between innovation and public safety.”
John Kennedy, U.S. Senator from Louisiana, via The Hill
He was also candid about the limits of his own proposal.
“I’m not standing here today suggesting to you I have all the answers.”
John Kennedy, U.S. Senator from Louisiana, via Washington Examiner
What Happens Next in the House and Senate
H.R. 9917 remains in the House, where it needs committee action before any floor vote can happen. Kennedy’s bill, having failed its unanimous-consent shortcut, now heads back into the normal process, which means committee referral and a markup schedule that has not yet been set. The bipartisan duty-of-care framework is even earlier in its life span, still described as a draft under discussion with no named sponsors and no public bill text.
The most likely near-term development is a formal, numbered introduction of Kennedy’s bill, which would give reporters, lobbyists, and the four affected labs actual text to analyze instead of a description of intent. Once that happens, expect a wave of public statements from Google, OpenAI, Anthropic, and Meta, along with pointed questions from committee members about how a kill-switch order would interact with existing national security and export-control statutes that already touch AI hardware and software.
Five Predictions for the AI Kill Switch Fight
- Kennedy reintroduces his bill through regular order within the current legislative session, giving it a formal bill number and a committee assignment rather than relying on another unanimous-consent attempt.
- Google, OpenAI, Anthropic, and Meta lobby collectively to narrow the definition of “covered AI system” in H.R. 9917 rather than fight the concept of a kill switch outright.
- The judicial review gap in H.R. 9917 becomes the central sticking point in committee, with civil liberties groups pushing to add a court-review step before any DHS shutdown order takes effect.
- The unnamed bipartisan Senate framework gains public sponsors before Kennedy’s narrower bill does, since the duty-of-care model already includes the judicial review safeguard that H.R. 9917 currently lacks.
- None of the three tracks becomes law before the end of the current Congress, and any version that does pass includes a phased compliance timeline running well into 2027 or later, mirroring how California’s own kill-switch mandate was ultimately dropped in favor of transparency rules.
None of these are locked in. Congress has a long history of introducing tech-related bills that stall in committee, and AI legislation in particular has moved slowly for years despite repeated warnings from lawmakers on both sides of the aisle. What makes this round different is the number of parallel tracks moving at once, a blocked floor vote that already made news, and four named labs that give the debate a concrete face instead of an abstract policy discussion.
Frequently Asked Questions
What is the AI Kill Switch Act?
The AI Kill Switch Act is a House bill, filed as H.R. 9917 in the 119th Congress, that would let the Secretary of Homeland Security order a slowdown or shutdown of covered AI systems capable of catastrophic harm, with the Secretary of Commerce and the Director of National Intelligence consulted in the process.
Who sponsored the AI Kill Switch Act?
Rep. Ted Lieu (D-Calif.) and Rep. Nathaniel Moran (R-Texas) introduced the bill on July 23, 2026, giving it bipartisan sponsorship in the House.
Why did Rand Paul block John Kennedy’s AI kill switch bill?
Kennedy tried to pass his Senate kill switch bill by unanimous consent on Sept. 16, 2026. Paul objected, saying the Senate needed more facts before regulating the entire economy, which sent the bill back into the normal committee process instead of a same-day passage.
Does blocking a unanimous consent request kill a bill permanently?
No. It only forces the sponsor to pursue the bill through the standard legislative process, including committee referral, markup, and a recorded floor vote, rather than an immediate same-day passage.
Which companies would be affected by an AI kill switch law?
Reporting on both the House bill and the Senate tracks names Google, OpenAI, Anthropic, and Meta as the frontier labs most likely to fall under a future shutdown or duty-of-care authority.
How does the US kill switch proposal compare to the EU AI Act?
The EU AI Act uses risk-tiered obligations that require companies to demonstrate safety before deployment, with no explicit kill switch mandate. The US proposals instead focus on giving a federal agency authority to act after a system is already deployed, which is a materially different regulatory approach.
Did California ever pass a kill switch law?
No. California’s SB 1047 would have required a mandatory kill switch and annual third-party audits, but Gov. Gavin Newsom vetoed it in September 2024. The state’s follow-up law, SB 53, dropped the kill switch requirement in favor of transparency and incident reporting rules.
Has any version of the federal AI kill switch bill become law?
No. As of Sept. 17, 2026, H.R. 9917 has not cleared committee in the House, Kennedy’s Senate bill was blocked from a fast-track vote, and the bipartisan duty-of-care framework remains in draft form with no public bill text.