Anthropic, the company behind Claude, has been quietly writing the law it wants to be regulated by — and then slipping it across the table to American lawmakers.
That is the picture that emerged on October 16, 2026, when Politico, citing three people with knowledge of the effort, reported that Anthropic representatives had been pitching Capitol Hill on a draft bill to regulate frontier AI safety and security. Reuters confirmed the story the same day. According to the report, the draft would give the U.S. government something the tech industry has long resisted: the power to block the release — or halt the use — of an AI model it deems too dangerous.
Not everyone agrees with Anthropic’s approach. But before the debate begins, it is worth asking a simpler question: what does the bill actually say? The draft itself has not been published, so the best picture we have comes from Politico’s reporting, Reuters’ confirmation and Anthropic’s own statements. Here is what is confirmed, what is reported and what remains unknown.
An illustration representing artificial general intelligence, by David S. Soriano / Wikimedia Commons (freely licensed)
The headline: a government off-switch, with a very high bar
The most eye-catching detail is the threshold for government intervention. According to Politico, the bill would let the government block a model’s release or use only if it poses a foreseeable, imminent risk of either:
- More than $1 billion in damages, or
- Killing or seriously injuring more than 50 people.
Those are not small numbers. They describe a catastrophic scenario — the kind of “loss of control” moment AI safety researchers talk about — not an everyday product failure. The line is deliberately sharp: this would not be a tool for policing chatbots that say the wrong thing. It would be a brake for models that could cause genuine national-scale harm.
The draft also narrows the kinds of risk that count. According to the report, the danger must stem from one of three specific sources:
- The creation or release of a biological weapon — the scenario that worries security officials most, since AI could put expert-level biology knowledge in the hands of people who never had it.
- The model evading its developer’s controls — the AI doing things its creators did not intend and cannot stop.
- A cyberattack — an AI system that meaningfully upgrades offensive hacking capabilities.
Only if all conditions are met — a foreseeable, imminent risk, above those casualty or damage thresholds, from one of those three sources — does the government’s power kick in.
How the shutdown would actually work
The draft does not hand the president, the FBI or any single agency a button labeled “turn off the AI.” Instead, it builds a multi-step process that is slow, deliberate and checked by a judge.
Here is the chain, as reported by Politico:
Step one: independent evaluators. The bill would create a category of independent evaluators who register with the Department of Commerce and periodically evaluate AI labs and their frontier models. Labs could pay these evaluators — but the payment could not be tied to the evaluation’s outcome, a detail meant to prevent labs from buying a clean bill of health.
Step two: a 72-hour report. If evaluators determine there is what the bill calls “imminent enumerated risk” — a defined, serious danger from one of the three trigger categories — they must report it to Commerce within 72 hours of finishing their assessment.
Step three: Commerce goes to court. Commerce could then ask a federal judge — specifically in the U.S. District Court for the District of Columbia — to order the suspension of the model.
And here is the detail that surprised even some Washington watchers: the draft explicitly prohibits any other U.S. government department, agency or officer from suspending a model’s development, deployment or internal use. There would be exactly one legal path to a shutdown — evaluators, then Commerce, then a single federal court — and no shortcut through any other part of the government.
Anthropic appears to be threading a needle: give the government enough power to stop a truly dangerous deployment, while making sure no agency can quietly seize control of AI development through regulatory creep.
Why Anthropic wrote this — in its own words
Anthropic is not pretending the draft came out of nowhere. A company spokesperson told Reuters the bill “represents proposals outlined in Anthropic’s ‘Advanced AI Framework’ released earlier this year” — a policy paper laying out the company’s view of AI regulation.
According to that spokesperson, the bill would grant the government the “ability to block or deter dangerous deployments of the most powerful models,” while requiring developers to test frontier AI models, be transparent about their findings and have them independently evaluated.
That framing matters. Anthropic is positioning itself as the lab that wants binding rules — a stance it has taken publicly for years. “Voluntary commitments are not enough” has become something of a company motto.
Why now: the regulatory era is already shifting
The draft did not arrive in a vacuum. Three days before Politico’s report, Anthropic and OpenAI representatives appeared before the UK parliament’s Business, Innovation, Science and Trade Committee and told MPs that voluntary safety pledges no longer cut it. Anthropic’s head of Europe, Liam Booth-Smith, was blunt: “The rules should be binding.” OpenAI’s Tom Duff Gordon backed a statutory framework, starting with mandatory incident reporting. Google DeepMind and Meta declined to back new legislation — a public split between the four leading labs.
The timing was dramatic. At the same hearing, the head of Britain’s AI Security Institute admitted it tests frontier models only when companies volunteer them and has no power to stop a release it deems dangerous. And around the same time, MI5’s director general, Ken McCallum, used his annual threat update to warn that AI’s potential to change the stakes on biological weapons is the misuse case he weighs most heavily.
Stateside, Anthropic has been pushing stricter rules in Massachusetts, where it backs requiring frontier labs to undergo independent catastrophic-risk evaluations every few months — a stance that puts it at odds with OpenAI and Google, which favor lighter annual audits.
The pattern is consistent: Anthropic believes the voluntary era of AI safety is ending, and it would rather help write the mandatory rules than have them written for it.
What this could mean for ordinary users
If a version of this bill ever became law — a very big “if” — the practical effects on everyday users would likely be subtle. Because the thresholds are so extreme, the shutdown power would almost certainly never touch the AI tools people use daily: chatbots, writing assistants, image generators. The bill targets frontier models — the most powerful systems labs are building — and catastrophic scenarios like bioweapon assistance or a model escaping its creators’ control.
The more meaningful effects would fall on developers. Mandatory testing, transparency about findings and independent evaluations would add cost and friction to frontier development — which is, in Anthropic’s view, partly the point. The company has argued for “pacing” the AI race so safety evaluation can keep up with capability gains, and this bill would encode some of that pacing into law.
The open question is whether compliance would be real or theater. If evaluators are paid by the labs they evaluate — even with payment decoupled from outcomes — critics will question their independence. And since the bill text has not been released, the real definitions of “frontier model,” “foreseeable” and “imminent” — where the actual fights will happen — remain unknown.
The skeptics’ case
Three lines of criticism deserve to be taken seriously.
First, the draft comes from the very industry it would regulate. When a company writes the rules for its own oversight, reasonable people ask whether the thresholds sit where they do because safety demands it — or because Anthropic is confident its own models would pass. Bars this high may never be triggered, which critics will read as a feature, not a bug.
Second, the ban on other agencies suspending models cuts both ways. Anthropic presents it as protection against overreach. But it also means that if the single permitted channel — evaluators, Commerce, one D.C. court — fails or stalls, there is no backup.
Third, a draft is not a law. Quietly pitching lawmakers is how many bills begin; most never become legislation. Congress has a long history of letting AI regulation bills die quietly, and there is no guarantee this one moves at all.
All of that said, the draft represents something genuinely new: an AI lab asking, in writing, for the government to be able to pull the plug on its most powerful creations. In an industry whose default posture has been “trust us,” that is a bet worth watching.
Frequently asked questions
What is the Anthropic AI safety bill?
A draft bill that Anthropic representatives have quietly pitched to U.S. lawmakers, first reported by Politico on October 16, 2026 and confirmed by Reuters. It would regulate frontier AI safety, including a narrow government power to block or suspend dangerously risky AI models.
What is the “AI kill switch” in the bill?
The reported draft would let the government block a model’s release or use if it poses a foreseeable, imminent risk of more than $1 billion in damages or of killing or seriously injuring more than 50 people — stemming from a bioweapon, the model evading its developers’ controls, or a cyberattack. Only a federal judge in Washington, D.C., could order the suspension, after evaluators report the risk to the Commerce Department.
Is this an actual law?
No. It is a draft shared with lawmakers — not introduced, voted on or signed. Treat it as a proposal, not a done deal.
Why would an AI company ask to be regulated?
Anthropic says voluntary commitments are not enough. Strategically, helping write the rules lets the company shape them — and high intervention thresholds mean the rules would rarely touch its own products. Critics see self-interest; supporters see responsibility.
Would this affect the chatbots I use?
Almost certainly not directly. The bill targets frontier models in catastrophic scenarios. Testing and evaluation requirements would fall on developers, not everyday users.
Where did the details come from?
Politico’s reporting (three people familiar with the effort), confirmed by Reuters, plus Anthropic’s own statements about its Advanced AI Framework and its testimony before the UK parliament on October 13, 2026.
Sources
- Reuters: “Anthropic quietly pitches its own AI safety bill to lawmakers, Politico reports” (Oct 16, 2026)
- WebProNews: “Anthropic Pitches Draft Bill to Regulate Frontier AI Safety”
- The Hindu BusinessLine: “Anthropic quietly pitches its own AI safety bill to lawmakers, Politico reports”
Featured image: “Artificial General Intelligence Illustration” by David S. Soriano, via Wikimedia Commons (freely licensed).
The conversation
Post a Comment