On October 16, 2026, Politico reported that Anthropic had quietly pitched a draft federal AI safety bill to members of Congress. No sponsor has been named. No bill number exists. Nothing has been introduced.

That hasn't stopped the arguments. The comment sections split into two camps within hours: Anthropic is regulating its competitors out of existence, or Anthropic is the first lab brave enough to hand the government a kill switch. Both readings skip the part that makes this draft unusual, which is how tightly its authors constrained the power it grants.

If you use Claude on your phone, this matters to you more than a typical policy story. The rules in play decide what models reach the apps you install, when a lab can be forced to pull one, and how much notice you get when it happens.

The Anthropic Draft Bill in One Paragraph

Anthropic's draft would let the federal government block or suspend an AI model's release, but only in a narrow set of circumstances. The model has to pose a foreseeable, imminent risk of more than $1 billion in damages, or of killing or seriously injuring more than 50 people. The risk has to come from three named sources: a biological weapon, a model escaping its developer's control, or a cyberattack. Only the Commerce Department can act, and it has to go through a federal judge in Washington, DC.

Everything else in the draft supports those limits. Evaluators get registered and given a reporting clock. Labs get duties around safety frameworks and transparency. States get partially preempted, and other federal agencies get barred from shutting models down on their own.

The Covered Developer Threshold Explained

Anthropic hasn't published the draft text, so the exact coverage rules aren't public. The Anthropic Advanced AI Framework, its public version and the one the company says the bill reflects, sets a two-part test for who counts as a covered developer.

The model must be trained with more than 10^25 FLOPs of compute. The company building it must also clear a financial bar: more than $500 million in AI revenue or more than $1 billion in annual AI R&D spending.

That combination matters. A compute threshold alone would catch every well-funded lab regardless of size. Adding the revenue or R&D test means a research group burning compute on a large model, without a commercial product, may not qualify. Whether the bill text keeps both halves is unknown, and Politico's account doesn't say.

For comparison, the EU AI Act's Article 51 uses the same 10^25 FLOPs line to presume systemic risk, without a financial test attached. China's draft export list uses 10^26 FLOPs, or 2 trillion parameters, or a top-ten ranking.

Who Can Actually Block a Model Under the Draft

This is the part of the draft that separates it from most AI regulatory proposals. The off switch is concentrated.

Commerce would have to request a suspension from a federal judge in the District of Columbia. That judge decides. Every other US department, agency, or officer would be barred from suspending a model's development, deployment, or internal use.

That reads as a direct response to a recent event. Politico reports that over the summer, Anthropic was hit by export controls that shut down two of its most advanced models, decided in a series of calls between officials and company executives. A written threshold plus a court gate is the opposite of a decision made on a phone call.

Critics will call that self-serving, and the criticism has merit. Anthropic is proposing the rules that would govern Anthropic. But the choice of a judge over an agency is a real constraint on the executive branch, not a blank check. A critic who thinks the thresholds are too high has a concrete number to attack. A critic who thinks the process is unaccountable has to reckon with the court requirement.

The 50 Deaths and $1 Billion Trigger

The risk trigger is where most of the debate will land. Fifty deaths or serious injuries is a mass-casualty event by any reasonable definition. One billion dollars in damages is a catastrophic loss.

Set those numbers against what an AI model would have to do to hit them. The harm has to come from a biological weapon, from the model escaping its developer's control, or from a cyberattack. It has to be foreseeable and imminent, meaning not a speculative chain of events years down the line.

That design choice means the draft doesn't attempt to regulate present-day chatbot behavior at all. Misinformation, privacy loss, bias in outputs, and job displacement are entirely outside its scope. Anyone hoping for a bill that addresses the risks you encounter using an app today won't find it here. The draft aims at a narrow category of catastrophic, near-term harm from frontier models.

The risk language reportedly mirrors California's AI laws, including SB 53, which was signed last year. That's worth knowing, since it means part of the draft is drawing on language that already has a legislative history.

How the Registered Evaluator System Works

The draft creates a formal role for outside evaluators, and the mechanics are specific.

Evaluators register with Commerce. They review the lab and its frontier models. They can be paid by the lab, but only if their pay doesn't depend on the results of the review. If an evaluator finds an imminent risk from one of the enumerated categories, they report to Commerce within 72 hours.

Anthropic already operates something like this in private. It gave METR staff employee-like access on October 15, 2026, and it works with SecureBio and Accenture on evaluations. The draft would turn that practice into a registered role with a legal clock attached.

The independence clause is the weak point on paper. A lab paying its own evaluators, even with a results-independent rule, creates an incentive structure that's hard to police. Whistleblower protections in the draft partially address it, but the details will decide whether it works.

State Law Preemption in the Anthropic Bill

The draft preempts some state laws, not all of them. Three areas: public risk disclosure, third-party evaluation, and critical-incident reporting.

Anthropic's public framework draws a line the draft will get judged against. It says Congress shouldn't preempt state law unless it enacts federal law at least as strong as the proposed framework. The draft, as reported, stops well short of that in the areas where it overlaps.

California's SB 53 specifically requires risk disclosure and incident reporting for frontier developers. A federal law that preempts those requirements without a matching federal standard would weaken the rules a large state already passed. That's an honest criticism of the draft, whichever company proposed it.

What the Bill Means if You Use Claude on Android

Here's the honest answer: nothing changes this month.

The bill is a draft, not a law. No sponsor, no introduction, no vote. Congress has been slow on AI legislation for three years, and a privately circulated draft from one lab doesn't move that pace on its own.

The indirect effects are more interesting. If a version of this passes, the labs covered by the 10^25 FLOPs line would need transparency reports and safety frameworks before releasing frontier models. That could mean longer gaps between major model launches, and fewer surprise rollouts like the ones that shipped straight to mobile apps without warning.

The 72-hour evaluator reporting clock would also create a scenario you haven't seen yet: a model pulled from app stores on short notice after an evaluator flags a risk. Whether that's realistic depends on thresholds nobody has hit.

For now, your Claude app keeps working the way it does today. The rules that govern it are being written in rooms you can't watch, which is exactly why the drafting details are worth following.

The Bottom Line on Frontier AI Regulation

Anthropic's draft is a narrow, court-gated, narrowly triggered proposal, not a broad safety law. It puts a ceiling on government power as much as it imposes duties on labs, and it says almost nothing about the risks ordinary users experience.

Whether you read that as restraint or as regulatory capture depends on whether you think a company should write the rules for its own industry. The draft gives critics specific numbers to argue with, which makes it more useful than most industry white papers. It also has no path to becoming law right now, which makes it more interesting as a statement of intent than as legislation.

Anyone tracking AI policy should watch three things: whether a sponsor picks it up, whether the compute threshold moves, and whether the state preemption section survives. Those are the three places where the draft could break.

Check other ai tool: