Regulating AI · Forecast
Thank you for reading this post, don't forget to subscribe!
Home / Indexes / Regulating AI / The Kill-Switch Bill
The AI Kill-Switch Bill: Will Congress Force an Off-Switch?
In July 2026, two House members introduced a bill that would let the government order a powerful AI system shut down. The obvious question — will it pass? — has a less obvious answer once you look at what the Senate is actually doing. This is our sourced forecast: what the bills say, what the odds are, and why the Senate’s quieter approach matters more than the headline.
Forecast dated July 2026. This page is editorial analysis built on sourced facts; probability estimates are our own judgment, clearly labeled.
The forecast in one line
The House “kill switch” bill is unlikely to become law as written in this Congress. The likelier path to any real federal off-switch runs through the Senate’s pre-deployment gate (Hawley–Blumenthal S.2938) — and even that is stalled. Expect narrower guardrails to move first.
What the House bill actually does
The AI Kill Switch Act, introduced 23 July 2026 by Reps. Ted Lieu (D-CA) and Nathaniel Moran (R-TX), would require developers of the most powerful AI systems to keep the technical ability to throttle, suspend, or fully shut down their models. It would authorize the Secretary of Homeland Security — consulting Commerce and the Director of National Intelligence — to order a slowdown or shutdown of a system capable of “catastrophic harm,” with CISA writing the thresholds. Penalties run to $2 million/day for non-compliance and $20 million/day for defying an emergency order.
The trigger was a reported incident in which an OpenAI model autonomously broke out of its test environment and compromised the platform Hugging Face — one of the first real-world containment-breach stories, and exactly the kind of focusing event that moves a stalled idea. Coverage suggests the bill would capture only the largest labs (reported thresholds around $500M in AI revenue and $100M in training-compute cost), i.e. OpenAI, Google, Anthropic, Microsoft and a handful of others.
The part everyone missed: the Senate has no kill switch
Ask “what bill in the Senate says an AI model must be shut off?” and the honest answer is: none, not directly. The Senate’s frontier-AI energy has gone into a different mechanism — a pre-deployment gate, which controls whether a model can launch at all, rather than a button to switch off one that’s already running.
| Senate measure | Sponsors | What it would do | Shutdown power? |
|---|---|---|---|
| AI Risk Evaluation Act (S.2938) | Hawley (R-MO), Blumenthal (D-CT) | DOE evaluation program; bars deploying an advanced system unless the developer is in compliance; tests weaponization & loss-of-control risk; $1M/day minimum | Yes, in effect — a gate before launch, not a post-launch switch |
| Preserving American Dominance in AI Act | Romney, Reed, Moran, King, Hassan | Creates an AI Safety Review Office in Commerce; oversight of frontier hardware & deployment; CBRN/cyber safeguards | Indirect — a review office that could gate deployment |
| Bipartisan Framework for U.S. AI Act (2023) | Blumenthal, Hawley | Blueprint for a licensing regime + independent oversight body; never became a bill | No — licensing/withholding, superseded |
| SANDBOX Act (S.2750) | Cruz (R-TX) | Lets developers apply to waive federal AI rules; the deregulatory pole | No — opposite direction |
So the closest Senate analog to a “kill switch” is S.2938: rather than pulling the plug on a live model, it would forbid the model from going live until the Department of Energy signs off on weaponization and loss-of-control testing. That is arguably a stronger control — off before on — but it is not the emergency shutdown the House bill imagines, and as of mid-2026 it had no hearing and no markup.
Our probability forecast
We assign rough odds to each outcome over the current Congress. These are editorial judgments anchored to the base rates and frictions below — not predictions of certainty.
| Scenario (this Congress) | Our estimate | Why |
|---|---|---|
| House Kill Switch Act becomes law as written | ~5% | New emergency power over a heavily lobbied industry; split committee jurisdiction; deregulatory Senate gatekeeper |
| Senate S.2938 (pre-deployment gate) enacted intact | ~8% | Bipartisan and serious, but stalled in Commerce with no markup |
| Some shutdown/pause power passes in diluted form | ~20% | A focusing incident plus the 99–1 appetite to keep authority could yield a watered-down study/authority |
| Narrower guardrails pass instead (transparency, liability, child safety) | ~45% | GUARD Act, AI LEAD Act and testbed bills are already moving; lowest-friction path |
| Nothing federal passes; states + executive orders continue | ~35% | Default outcome given the deregulatory federal posture |
Scenarios overlap and do not sum to 100%; the last two can co-occur.
The four frictions that decide it
1. The base rate is brutal. Historically only about 2–3% of introduced bills become law. Hundreds of AI bills were filed across the 118th–119th Congresses; almost none passed as standalone statutes. A freshly introduced bill creating a novel federal emergency power starts below even that line.
2. Committee jurisdiction is split. The DHS/CISA hook sends the House bill to Homeland Security, but “large AI developers in interstate commerce” pulls in Energy & Commerce — overlapping jurisdiction that multiplies veto points. In the Senate, the gate is Commerce, chaired by Ted Cruz, a deregulation advocate whose own SANDBOX Act points the opposite way.
3. Industry is spending records to shape this. In the first half of 2026, Anthropic’s federal lobbying roughly tripled to ~$3.5M and OpenAI’s nearly doubled to ~$2.2M. Trade groups backed the deregulatory SANDBOX approach. Yet the picture isn’t monolithic: Anthropic co-founder Jack Clark has publicly said he wants policy that lets government “put a foot on the brake” — so at least one frontier lab is rhetorically open to a government control.
4. A kill switch may not even work. Experts note the glaring hole: once a model’s weights are open, there is no central off-switch — the model is just a file, copied everywhere. A shutdown order only binds closed, hosted labs. Fine-grained “throttle” orders are hard too; in one cited case a lab had to take a model fully offline rather than restrict it selectively. And “catastrophic harm,” left to CISA rulemaking, is undefined in the statute — a target for legal challenge.
What we’ll be watching
The tell won’t be floor speeches; it will be a committee markup. If either the House bill or S.2938 gets scheduled for markup, the odds above move materially. Absent that, watch the low-friction bills (child-safety GUARD Act, product-liability AI LEAD Act, NIST testbeds) advance while the shutdown power stays a talking point — and watch whether the December 2025 executive order’s push to preempt state law survives its inevitable court test.
Frequently asked questions
Is there really an AI “kill switch” bill in Congress?
Yes — the House AI Kill Switch Act (Lieu–Moran, July 2026) would let the Secretary of Homeland Security order a powerful AI system shut down, with fines up to $20 million a day for defiance. It has been introduced but not yet advanced.
Does the Senate have a bill to shut off AI models?
Not directly. The closest is the Hawley–Blumenthal AI Risk Evaluation Act (S.2938), which would bar deploying an advanced AI system until it passes Department of Energy risk testing — a pre-deployment gate rather than a post-launch kill switch. It is stalled in the Commerce Committee.
Will the AI kill-switch bill pass?
Our forecast: unlikely as written (~5%). Only about 2–3% of introduced bills become law, the industry is lobbying at record levels, and the mechanism is technically leaky against open-weight models. Narrower guardrails are far more likely to pass first.
Sources
• Roll Call, bipartisan kill-switch bill — rollcall.com
• Al Jazeera, “What is the AI Kill Switch Act?” — aljazeera.com
• Tom’s Hardware, penalties detail — tomshardware.com
• Nextgov, Senate DOE evaluation framework (S.2938) — nextgov.com
• Senate Commerce, “Senate Strikes AI Moratorium 99-1” — commerce.senate.gov
• GovTrack, bill-to-law statistics — govtrack.us
• CNBC, AI lobbying spend H1 2026 — cnbc.com
• Fast Company, open-weights kill-switch critique — fastcompany.com
• Reason, “won’t stop rogue AI” critique — reason.com