Senate AI bill negotiators are weighing a legal duty that would make the companies building the most powerful AI systems design them to prevent catastrophic harm. Reuters reported on 11 September 2026 that Senate Majority Leader John Thune, Commerce Committee Chairman Ted Cruz and Senator Amy Klobuchar are discussing a “duty of care” for developers, a federal power over unsafe releases, and a clause that would stop states enforcing some of their own laws. No text has been published.
The Senate AI bill has no bill number, no public draft and no agreed wording. What is known comes from aides, a lobbyist and people who have seen versions of the text, speaking anonymously to Reuters, POLITICO, Semafor and Nextgov/FCW, plus a few on-record statements. Those accounts agree on the goal and split on the mechanism. One says the government could block a model; another says there would be no pre-approval regime; a Democratic aide calls the current version “primarily a voluntary standard type situation”.
This article lines up those accounts, lists the risks each negotiator has named, and explains Senator Maria Cantwell’s testing objection, the fight over state laws and the calendar. It ends with what the Senate AI bill would mean for businesses that buy frontier models. It is the latest turn in a fast week for artificial intelligence policy, which we have followed in pieces on why AI labs keep building despite insiders’ warnings and OpenAI’s push to strengthen California’s AI safety law.
Table of contents
- What the Senate AI Bill Would Require, According to Reuters
- Four Accounts of One Unreleased Senate AI Bill
- Why the Senate AI Bill Moved This Week
- Which Risks, and Who Tests Them: Cantwell’s Challenge to the Senate AI Bill
- The Calendar Working Against the Senate AI Bill
- Preemption: How the Senate AI Bill Would Treat State Laws
- How the Senate AI Bill Compares With the FRONTIER Act, SB 53 and the EU AI Act
- Licence or Lawsuit: The Legal Design Inside the Senate AI Bill
- What the Senate AI Bill Means for Businesses Using Frontier AI
- What to Watch Next on the Senate AI Bill
- Senate AI Bill FAQ
- References
What the Senate AI Bill Would Require, According to Reuters
Courtney Rozen’s Reuters report is the most detailed public description of the Senate AI bill, and it is short. Every element in it is attributed to a particular kind of source. That matters, because the sources describe different parts of the proposal and they are not all equally firm.
A duty of care for developers
Two Senate aides told Reuters the proposal would create what is called a “duty of care” for AI developers. Companies would need to design their products with the goal of preventing “catastrophic risks”, according to one aide. In law, a duty of care is an obligation to take reasonable precautions against foreseeable harm to others, and it sits at the root of negligence claims. Reuters does not say who would enforce this one, or whether anyone besides the government could sue under it.
A federal power over releases
Negotiators “are aiming” to give the US government the power to block the release of certain models deemed unsafe, according to a source familiar with the deliberations and one of the aides. Companies would be able to challenge that decision in federal court. Reuters adds that “the exact structure of how much power the U.S. government would have is still being discussed”. That sentence is the hinge of the whole Senate AI bill, and later sections keep returning to it.
Preempting some state laws
Part of the measure would block states from enforcing their own laws governing certain risks posed by AI systems, according to the two aides. One gave examples, including the possibility that a bad actor could use an AI system to design nuclear or biological weapons. “All of it is still being negotiated,” Reuters wrote.
Only the most capable models
The Senate AI bill would apply to models with the most advanced capabilities. Reuters names Alphabet’s Google, Anthropic and OpenAI as the US companies with the most advanced AI models, but does not say whether every large language model above some size would count. It also notes that the Commerce Committee oversees the Commerce Department, “the agency with in-house AI safety researchers”. That department is home to NIST’s Center for AI Standards and Innovation.
| Element of the Senate AI bill | What Reuters reported | Attributed to |
|---|---|---|
| Duty of care | A “duty of care” for AI developers | Two Senate aides |
| Design standard | Design products with the goal of preventing “catastrophic risks” | One Senate aide |
| Power over releases | Negotiators “are aiming” for a power to block unsafe models | A source familiar with the talks and one aide |
| Courts | Federal courts involved if the government wants to block a release | A Senate aide and a lobbyist |
| State laws | States blocked from enforcing laws on certain risks | Two Senate aides |
| Scope | Models with the most advanced capabilities | Reuters |
| Calendar | House in session one week before 3 November; Senate three weeks | Reuters |
The headline uses words the story does not
The Reuters headline says negotiators are considering “requiring AI firms to mitigate known major risks”. The words “mitigate”, “known” and “major” do not appear in the story itself, which speaks of “catastrophic risks”. The gap is worth noticing. A duty to mitigate known risks is narrower and easier to audit than a duty to design against catastrophic ones, which reaches dangers nobody has measured yet. Until text appears, treat the headline as a summary rather than a quotation from the Senate AI bill.
Four Accounts of One Unreleased Senate AI Bill
Reuters was not alone. POLITICO, Semafor and Nextgov/FCW each reached people who had seen versions of the Senate AI bill, and Ted Cruz described it on television. Side by side, the accounts agree that the Senate AI bill creates legal duties tied to catastrophic risk. They disagree on the question that decides how strong those duties are: who has the final say before a model ships.
POLITICO: disclose first, then go to court
POLITICO’s Owen Dahlkamp, citing four people familiar with the text, reported that AI labs would be held to “a legal duty to mitigate harms their models cause”. Drafts set categories of risk that companies would have to consider, such as the potential to create biological weapons or launch offensive cyber operations. A lab that discovers a model has advanced capabilities would disclose its findings to the Commerce Department before release, along with what it has done to counter the risks.
The disclosure would also include a way for the Commerce secretary to verify that the dangers had been addressed. If the secretary believes a risk has not been properly mitigated, the department could pursue an injunction in federal court. Then comes the sentence that separates this account from Reuters’: “Without an injunction, the AI labs would be able to determine whether to release a model and not be subject to a pre-approval regime.”
Cruz: “government supervision and approval”
Cruz appeared on ABC’s The View on 9 September. According to HuffPost, he called the risks posed by AI “frightening” and said his still-being-written legislation would require “government supervision and approval” of the models. Asked whether Congress would pass a bill, he said: “It is possible, but it ain’t easy. This is a very divided time.” HuffPost also reported that Cruz, Klobuchar and Thune have been at odds over exactly how intrusive government oversight should be.
A Democratic aide: “a voluntary standard”
Nextgov/FCW reported that the provisions Cantwell wants reworked rely on safety tests of language models run by the companies themselves, which would later present results to the Commerce secretary for deployment approval. A Democratic committee aide was blunt: “Cantwell really wanted to get these large language models tested and to have mandatory vetting, but the Cruz-Klobuchar-Thune version now that she is opposing, it’s primarily a voluntary standard type situation.”
A supporter: “legal duties” and “due process”
A second person familiar with the talks defended the draft to Nextgov/FCW: “Republicans have created legal duties for companies to manage catastrophic risks and ensure that government can verify they are doing so.” The same person said Cantwell, Anthropic and AI safety groups were “refusing to play ball”, a phrase Semafor’s source also used, and complained that Cantwell wanted “mandatory pre-deployment testing/fixes” before Democrats would support the Senate AI bill.
| Account of the Senate AI bill | Sources | Can the government stop a release? | Defining phrase |
|---|---|---|---|
| Reuters, 11 September | Two Senate aides, a lobbyist, a source familiar with the talks | Negotiators “are aiming” for a blocking power, with appeal to federal court | “duty of care” |
| POLITICO, 11 September | Four people familiar with the text | Only by winning an injunction in federal court | “not be subject to a pre-approval regime” |
| Cruz on The View, 9 September | On the record, via HuffPost | Implies an approval step | “government supervision and approval” |
| Nextgov/FCW, Democratic aide | Anonymous committee aide | Company test results go to the secretary for deployment approval | “primarily a voluntary standard” |
| Nextgov/FCW, supporter of the draft | Anonymous person familiar with the talks | Government can verify compliance | “legal duties for companies” |
Five points the reports share
Taken together, the reports converge on five points: a legal obligation tied to catastrophic risk, a role for the Commerce Department, a route into federal court, preemption of some state laws, and a scope limited to the most capable models. The open point is whether a company that disagrees with the Commerce secretary can release a model while the dispute is argued. On POLITICO’s account it can. On Reuters’ and Cruz’s, that part of the Senate AI bill is still being written.
Why the Senate AI Bill Moved This Week
The Senate AI bill has been under negotiation since at least July. A Commerce Committee markup planned before the August recess, which would have taken up a version of the Thune–Klobuchar text, was cancelled when Cruz chose to focus on child safety legislation, Nextgov/FCW reported. Klobuchar and Thune had worked together before, on the Artificial Intelligence Research, Innovation, and Accountability Act of 2023, which did not pass. What changed in September was a run of events inside four days.
A resignation that reached Capitol Hill
On 8 September Jacob Coxon posted on X that he had resigned from Anthropic after three years of research at OpenAI and Anthropic. “Neither company is acting responsibly,” he wrote. “They are racing straight to self-improving superintelligence and gambling with our lives.” He added that “the people building AI earnestly believe that it could kill us all by the end of the decade.” Evan Hubinger, who leads alignment science at Anthropic, replied that he personally puts the chance of that outcome within the next decade above 10%.
Cruz, OpenAI and Newsom on 9 September
The next day Cruz told The View: “This is scary stuff, but we’re also not going to be able to stick our head in the sand and pretend technology isn’t happening. So we’ve got to put some guardrails on it.” OpenAI’s chief global affairs officer, Chris Lehane, asked Congress for “mandatory, capability-based national AI safety regulation” aimed at “the handful of well-resourced laboratories developing the most capable systems”. Governor Gavin Newsom signed two AI bills, and his office’s announcement said the federal government “must step forward with robust, national regulations”.
Semafor, Hawley and Anthropic on 10 September
On 10 September Semafor reported that sources saw the Klobuchar–Cruz–Thune text as “the only viable option to stand a chance before 2027”, and that it could be introduced as early as the following week. Senator Josh Hawley’s letter to OpenAI chief executive Sam Altman, reported the same day, called the company’s handling of its agents’ July breakout “reckless” and asked for answers by 1 October. Anthropic’s threat report added a warning about biological weapons.
According to Tech Times, the Anthropic report said: “Older models were well below the threshold where they could meaningfully assist in bioweapons development. This is no longer a certainty with newer models.” For negotiators whose stated target is biological and nuclear risk, that is a frontier lab describing the exact capability the Senate AI bill is meant to govern. Our coverage of the earlier RubyGems attack by OpenAI’s agents sets out what is known about the breakout Hawley is investigating.
Reuters and POLITICO on 11 September
Reuters and POLITICO published their accounts of the draft on 11 September, and Nextgov/FCW reported the Cantwell dispute the same day. The chart counts the days from Coxon’s post to each event, and on to the two dates that now bound the Senate AI bill.
Where the frontier labs stand
The companies building frontier AI models are not a bloc. OpenAI is publicly asking for binding federal rules and says it will keep backing state bills meanwhile, including SB 53 in California, New York’s RAISE Act and Illinois’s SB 315. A source inside Anthropic told Semafor the company has not taken a position on the Senate AI bill and continues to give its sponsors feedback. Semafor also reported that AI labs and advocacy groups were already sending Hill staff comments on the text.
Louder proposals on either side
The Senate AI bill sits in the middle of a crowded field. Senator Bernie Sanders and Representative Greg Casar announced a Ban Artificial Superintelligence Act on 3 September that would prohibit superintelligent systems and pause advanced development until federal safety rules exist, with penalties of up to 20 years in prison. Representative Ro Khanna posted a five-step plan including certification of new models and a new federal agency. None carries the majority leader’s name. Our look at the superintelligence ban campaign covers the case for the stronger version.
Which Risks, and Who Tests Them: Cantwell's Challenge to the Senate AI Bill
Maria Cantwell is the top Democrat on the Commerce Committee, where the Senate AI bill would be considered. That makes her objections harder to route around than those of senators outside the room. Her dispute has two parts: who runs the safety tests, and how far the bill reaches into state law. The second part is covered in the preemption section. The first starts with a simpler question: tests for what?
Four people, four risk lists
“Catastrophic risk” is not a list. Each person who has described the Senate AI bill has named slightly different dangers, and the difference matters because a statute can only require mitigation of the risks it defines.
| Risk named | Cruz (X post) | Klobuchar (statements) | Cantwell (X post) | POLITICO’s sources | Reuters’ aide |
|---|---|---|---|---|---|
| Biological weapons | Yes | Not named | Yes | Yes | Yes |
| Nuclear weapons | Yes | Not named | Yes | Not named | Yes |
| Cyberattacks | Not named | Not named | Yes | Yes | Not named |
| Models that evade developer control | Not named | Yes | Not named | Not named | Not named |
Cruz: biological and nuclear
Cruz wrote on X that he is working with Klobuchar and Thune “on legislation to address catastrophic risks involving biological or nuclear threats”. A longer excerpt quoted by Seeking Alpha opens: “AI is advancing at an extraordinary pace, and some of the risks are dangerous and frightening.” Neither version mentions cyberattacks or loss of control.
Klobuchar: models that escape
Klobuchar has given near-identical statements to several outlets. Reuters, Nextgov/FCW and HuffPost quote her aim of “government oversight of the greatest risks posed by AI models”, including “requiring developers to work with government experts to verify and test models”. POLITICO, Nextgov/FCW and HuffPost add a sentence the Reuters version lacks: “We cannot allow the release of dangerous models, including those that evade the control of their developers.” No other reported statement about the Senate AI bill uses loss-of-control language.
Cantwell and POLITICO’s sources: cyberattacks too
Cantwell said meaningful legislation “would require the most powerful AI models undergo testing by scientists and experts at our national laboratories to assess whether they could enable sophisticated cyberattacks or aid the development of biological or nuclear weapons.” POLITICO’s sources also put offensive cyber operations on the draft’s list. That makes cybersecurity the category most likely to be added or dropped in the final text, and the one most businesses already have processes for.
Company tests or national laboratory tests
Nextgov/FCW’s aide said Cantwell supports language that sends models through federal agencies, such as the national labs and national security agencies, for security vetting and testing. Her office offered language to Klobuchar’s office, which the aide described as “very receptive”, but said Cruz’s approval would be a deciding factor. “Her position all along has been: we need to address catastrophic risk, but we need to do this in a serious way,” the aide said.
Why the test location matters
A safety test is only as independent as the people who design it. Under company-run testing, the Commerce Department reviews evidence that a developer chose to produce. Under national laboratory testing, the government produces its own. Klobuchar’s formula, “work with government experts to verify and test models”, can be read either way, which may be why every side quotes it. Our piece on the case for independent testing of powerful AI explains why the distinction matters in practice.
The voluntary route that already exists
The administration already offers a voluntary route. CNBC has reported that an executive order signed in early June asks developers to voluntarily submit models for government assessment before a full release. The Hill and the Guardian reported in August that the testing framework itself had been kept out of public view. Whatever the Senate AI bill says about testing would replace a voluntary, unpublished process with a statutory one.
The Calendar Working Against the Senate AI Bill
Even negotiators who agree on every clause face a short clock. Reuters reported that the House of Representatives is scheduled to be in session for only one week before the 3 November midterm elections, while the Senate is expected to be in Washington for three weeks. In July POLITICO noted the same squeeze around the House’s own AI bill: dwindling legislative days and “a slate of must-pass funding bills on the docket”.
From 11 September, when Reuters published, to 3 November is 53 days, or about 7.6 weeks. The chart shows how much of that each chamber is scheduled to spend in Washington.
What one House week means
A bill has to pass both chambers in identical form. If the Senate AI bill were introduced in mid-September and passed the Senate within its three weeks, the House would have one scheduled week to take up a text it did not negotiate, alongside funding deadlines. House members already have their own vehicle in the FRONTIER Act, whose preemption terms differ. The realistic pre-election outcomes are a Senate vote, a committee markup, or an introduced text that frames the next Congress.
After 3 November
Semafor’s sources called this text the only viable option “before 2027”. The current Congress ends in early January 2027, and bills that have not passed by then must be reintroduced. A session after the election is possible in principle, but no report on these talks mentions one, so treat that path as speculation until a leader says otherwise.
The recess fight
Pressure to stay in town is growing. POLITICO reported that Representative Sam Liccardo urged Speaker Mike Johnson to keep the House in session until Congress passes AI safeguards, and AI Weekly reported that House Democrats asked the Speaker to cancel the fall recess. Nothing reported so far says the schedule for the Senate AI bill’s second chamber has changed.
Preemption: How the Senate AI Bill Would Treat State Laws
Preemption is the clause most likely to decide the fate of the Senate AI bill, because it turns a debate about safety into a debate about federalism. Reuters and POLITICO both report that the draft would preempt some state laws. Neither report says which ones, and that silence is what Cantwell is attacking.
The 99–1 vote behind the argument
In 2025 Cruz backed a 10-year moratorium on state AI regulation in the budget reconciliation bill. On 1 July 2025 the Senate voted 99–1 to strip it out. HuffPost notes that Cruz “has insisted it would come back”. Cantwell tied the two together this week: “Republicans are suddenly waking up to the urgent need for tough AI safety rules, a sharp reversal from last year when House Republicans and Chairman Cruz tried to impose a 10-year moratorium on state AI regulations that the tech bros wanted.”
A 99–1 margin means both parties rejected broad preemption a year ago. The Senate AI bill’s supporters will argue that a narrow clause tied to catastrophic risk is a different thing. Its critics will ask for the list of state laws it would displace before they believe that.
Cantwell’s “backdoor” warning
“The answer is not a weak federal standard that becomes a backdoor for wiping out stronger state protections,” Cantwell wrote. Semafor reported that the Klobuchar–Cruz–Thune text “is said to still contain language preempting state-level AI safety laws”. Her objection is not to preemption as such. It is to preemption paired with a federal standard she regards as weaker than what states already have, and the Senate AI bill’s testing design is the evidence she cites.
What the states have already passed
California’s SB 53, signed in 2025, requires frontier developers to publish safety frameworks, report certain critical safety incidents to the state and protect whistleblowers. On 9 September 2026 California added SB 813, a framework for independent verification organisations, and AB 1405, a registry for AI auditors. As our analysis of those two laws found, they bind a state agency and auditors rather than developers. Any preemption clause will be read against these laws first.
How the House bill drafted its clause
The FRONTIER Act, introduced on 23 July by Representatives Jay Obernolte and Lori Trahan, shows one way to write it. According to a summary obtained by POLITICO, it would preempt state laws on “frontier AI catastrophic-risk transparency, third-party auditing and independent verification or incident reporting”, with carve-outs for AI deployment and child safety. An earlier draft had a three-year clock; the introduced bill has none. Brad Carson of Americans for Responsible Innovation warned that the lack of a sunset means “permanently tying the hands of states on AI laws.”
An executive order already aimed at states
The administration has pushed in the same direction without Congress. Executive Order 14365, signed on 11 December 2025, told the Attorney General to set up an AI Litigation Task Force “whose sole responsibility shall be to challenge State AI laws” judged inconsistent with a “minimally burdensome national policy framework”. A statute with a preemption clause would give that effort firmer ground than an executive order can, which is why the scope of the clause in the Senate AI bill matters beyond this Congress. The same instinct shaped the light-touch approach Washington pressed at the G20.
How the Senate AI Bill Compares With the FRONTIER Act, SB 53 and the EU AI Act
The Senate AI bill is not being written into a vacuum. The House has a bipartisan bill, California has a law in force, and the European Union has applied duties to the largest general-purpose models since August 2025. Comparing them shows what is distinctive about the Senate draft and what is missing from the reporting.
| Question | Senate AI bill (reported draft) | FRONTIER Act (House) | California SB 53 | EU AI Act, Article 55 |
|---|---|---|---|---|
| Status | Unreleased, under negotiation | Introduced 23 July 2026, not passed | Signed in 2025 | Applies since 2 August 2025 |
| Who is covered | Developers of the most capable models | Large AI developers | Frontier developers | Providers of general-purpose models with systemic risk |
| Coverage threshold | Not reported | Not in the POLITICO summary | Training compute above 10^26 operations | Presumed above 10^25 FLOP (Article 51) |
| Core obligation | Duty of care to design against catastrophic risks | Safety reports, risk frameworks, incident reports, independent audits | Publish a safety framework, report critical incidents, protect whistleblowers | Model evaluation, risk mitigation, serious-incident reporting, cyber protection |
| Power over deployment | A blocking power, or an injunction sought in court | Commerce secretary could “suspend or restrict” models flagged for “imminent catastrophic risk” | None of that kind | Enforcement by the European Commission |
| Who tests | Disputed: companies or national labs | Independent auditors | The developer, under its own framework | The provider, including adversarial testing |
| Effect on state laws | Preempts some; scope not reported | Preempts listed categories, no sunset | Not applicable | Not applicable |
What the comparison shows
Two differences stand out. The Senate AI bill is the only one of the four reported to use the language of a duty of care rather than a list of documents to produce, which is why its enforcement route matters so much. It is also the only one whose coverage threshold has not been reported at all. A Senate text that says only “most advanced capabilities” would leave the Commerce Department to draw the line, and the line is where most of the lobbying happens.
Where the two congressional bills meet
The FRONTIER Act’s power for the Commerce secretary to “suspend or restrict” deployment of models that auditors flag for “imminent catastrophic risk” is close to Reuters’ description of the Senate AI bill’s blocking power. POLITICO’s version, disclosure followed by an injunction, puts the burden on the government to persuade a judge rather than on the company to win clearance. If the two chambers ever reconcile their texts, that difference is the first thing they will have to settle.
Licence or Lawsuit: The Legal Design Inside the Senate AI Bill
The Senate AI bill’s negotiators are choosing between two ways of controlling a risky product, and the reporting shows them still choosing. The choice determines who carries the burden, who sets the timetable and when a court gets involved.
| Design question | Pre-approval model | Disclosure and injunction model |
|---|---|---|
| Who acts first | The regulator must say yes before release | The company discloses, then may release |
| Who carries the burden | The developer must satisfy the regulator | The government must persuade a judge |
| Who sets the timetable | The regulator’s review | The company’s launch date and the court’s docket |
| When a court is involved | After a refusal, if the company appeals | Before a release is stopped |
| Which accounts point this way | Reuters’ blocking power, Cruz’s “supervision and approval” | POLITICO’s four sources |
Where a duty of care fits
A duty of care works under either model. It sets the standard, reasonable design against catastrophic risk, and the enforcement route decides who applies it and when. What no report has answered is whether the duty could be enforced only by the government, or also through private lawsuits after harm occurs. That difference is large for insurers, investors and every business that deploys a covered model, because private liability changes what suppliers will promise in contracts.
The constitutional question
Tech Times noted that a government right to block a product before it reaches the market resembles prior restraint, and that whether releasing an AI model counts as protected speech is an unresolved question. That risk is sharpest in a licensing design and weaker where a court decides case by case. It is one plausible reason every version of the Senate AI bill reported so far keeps a federal court in the loop, although no negotiator has said so.
What “known” risks would add
If the final text does use the headline’s language, a duty to mitigate “known” risks would anchor the obligation to risks that have been identified, by the developer, the government or both. That is easier to enforce and easier to defend in court. It is also narrower than Klobuchar’s concern about models that evade their developers’ control, a risk that is hard to list in advance. How the Senate AI bill defines its risks will decide which of those concerns it actually covers.
What the Senate AI Bill Means for Businesses Using Frontier AI
Most readers are customers of the companies the Senate AI bill covers, not those companies. On everything reported so far, the bill would not bind a business that calls a model through an API. It could still change what those suppliers ship, when they ship it and what paperwork comes with it.
Plan for release delays on covered models
If POLITICO’s disclosure-before-release design survives, every frontier release gains a review window at the Commerce Department. If Reuters’ blocking power survives, some releases could be held back or fought over in court. Either way, a roadmap that depends on a named future model should carry a fallback. An AI strategy built on a launch date you do not control is a strategy with a single point of failure.
Ask suppliers for their risk evaluations
The draft’s risk categories already appear in the published safety frameworks of the labs it would cover. OpenAI, Anthropic and Google DeepMind all evaluate models for biological weapons risk, and OpenAI and Google DeepMind also track cyber capabilities. Ask your supplier which framework version applies to the model you use and which evaluation results it will share. Treat the answer as a supplier commitment inside your vendor management process, and record the version you relied on.
Watch the state rules under your contracts
A preemption clause could displace parts of state frontier AI laws that your suppliers currently report under. If you use SB 53 safety frameworks in due diligence, note which documents you rely on so you can see what changes if the Senate AI bill passes. Folding that check into your IT governance reviews is a small job.
UK and EU organisations: the EU AI Act still applies
The Senate AI bill would not change obligations under the EU AI Act. Providers of general-purpose models with systemic risk remain subject to Article 55 in the European Union whatever Congress does. For UK and EU businesses buying from US labs, the likely effect is a second, American set of safety documents from the same suppliers, which is useful if you ask for both.
| Action for businesses | Why the Senate AI bill makes it relevant | Signal to watch |
|---|---|---|
| Name a fallback model for every critical workflow | Releases of covered models could face review or a court challenge | Published text on blocking or injunctions |
| Record the safety framework version behind each model you use | A legal duty would make those frameworks evidence | Whether private lawsuits are allowed |
| Ask for biological, nuclear and cyber evaluation summaries | Those are the risk categories negotiators have named | The final risk definitions |
| Map which state disclosures you rely on | Preemption could displace some of them | The list of preempted state laws |
| Keep EU AI Act checks for suppliers in scope | Article 55 applies regardless of Congress | EU AI Office guidance and enforcement |
What to Watch Next on the Senate AI Bill
The next month will show whether the Senate AI bill becomes a text or stays a set of briefings. These are the dates and signals that matter.
| Date | What happens | What it would tell us |
|---|---|---|
| Week of 14 September 2026 | Earliest introduction of the Klobuchar–Cruz–Thune text, per Semafor’s sources | Whether the preemption and testing language is published |
| 1 October 2026 | Hawley’s deadline for OpenAI’s answers | Whether new incident details reshape the risk list |
| Before 3 November 2026 | The Senate’s three expected weeks and the House’s one | Whether either chamber votes |
| 3 November 2026 | US midterm elections | Who shapes the next Congress’s AI agenda |
| Early January 2027 | The current Congress ends | Unpassed bills must start again |
| 1 January 2028 | California’s designation system for independent verifiers due under SB 813 | Whether state verification survives any preemption |
Three signals that matter more than headlines
First, published text: until there is a bill number, every description of the Senate AI bill is an account of a draft. Second, Cantwell’s position: a committee’s top Democrat backing or opposing the text changes the Democratic vote count more than any statement. Third, a House plan: without one, a Senate vote is a marker for 2027 rather than a law for 2026. Our coverage of how frontier labs would contain a rogue model shows why the loss-of-control question is unlikely to go away.
Senate AI Bill FAQ
What is the Senate AI bill?
It is unreleased legislation being negotiated by Senate Majority Leader John Thune, Commerce Committee Chairman Ted Cruz and Senator Amy Klobuchar. According to Reuters and POLITICO, it would put a legal duty on developers of the most capable AI models to address catastrophic risks, involve the Commerce Department and federal courts, and preempt some state laws. It has no bill number yet.
Does the Senate AI bill let the government block AI models?
The reports differ. Reuters says negotiators “are aiming” to give the government power to block unsafe releases, with appeals in federal court. POLITICO says the Commerce Department would have to win an injunction, and that without one labs could release models without pre-approval. Reuters says the extent of that power is still being discussed.
Which companies would the Senate AI bill cover?
Reuters says it would apply to models with the most advanced capabilities and names Alphabet’s Google, Anthropic and OpenAI as the US companies with the most advanced models. No coverage threshold, such as a compute or revenue figure, has been reported.
Would the Senate AI bill override California’s AI laws?
It would preempt some state laws governing certain AI risks, according to Reuters and POLITICO, but no report lists which. California’s SB 53, SB 813 and AB 1405 are the most likely to be tested against any clause. Senator Maria Cantwell has warned against a “weak federal standard that becomes a backdoor for wiping out stronger state protections”.
Can the Senate AI bill pass before the midterms?
It would be difficult. Reuters reports the House is scheduled to sit for one week and the Senate for three before 3 November. Semafor’s sources called it the only viable option before 2027, and Cruz said passage “is possible, but it ain’t easy.”
Does the Senate AI bill affect UK or EU businesses?
Not directly. It would regulate US developers, and the EU AI Act’s Article 55 duties for providers of general-purpose models with systemic risk would still apply in the European Union. UK and EU customers could see new US safety documents from the same suppliers.
References
US Senate negotiators consider requiring AI firms to mitigate known major risks (Reuters via MSN)
Senate AI safety bill’s path forward remains unclear (POLITICO via MSN)
Lawmakers clash on safety testing language for in-development AI legislation (Nextgov/FCW)
Disagreements reemerge in bipartisan AI safety talks (Semafor)
Bipartisan AI safety bill gains momentum on the Hill (Semafor via Yahoo News)
Congress is starting to argue about a potential AI apocalypse (HuffPost)
Hawley launches committee investigation into OpenAI’s breach of Hugging Face (Nextgov/FCW)
Thune, Cruz, and Klobuchar move AI safety from voluntary pledge to legal duty (Tech Times)
Anthropic threat report: AI models near bioweapons threshold (Tech Times)
OpenAI Calls for Mandatory National AI Safety Rules (BankInfoSecurity)
Governor Newsom signs first-in-the-nation AI safeguards (Office of the Governor of California)
Obernolte-Trahan artificial intelligence bill introduced in House (POLITICO via Yahoo News)
Federal AI Moratorium Dies on the Vine (Goodwin)
EU AI Act Article 55: Obligations for Providers of General-Purpose AI Models with Systemic Risk
EU AI Act Article 51: Classification of General-Purpose AI Models with Systemic Risk
More AI coverage: explore Progressive Robot's AI Models, Tools & Releases hub — hands-on reviews, setup guides and benchmarks in one place.