Anthropic's draft federal AI safety bill sets a high bar for government intervention

The headline mechanism in Anthropic's draft is strikingly narrow. The government could only move to block a frontier model's release or use if it poses a foreseeable, imminent risk of killing or seriously injuring more than 50 people, or causing more than $1 billion in damages - and only if that risk traces back to one of three scenarios: a biological weapon, a model evading its developer's controls, or a large-scale cyberattack ^[1]. Even clearing that bar doesn't hand power to just any regulator. The Commerce Department is the only agency allowed to act, and it cannot suspend a model on its own authority - it must ask a federal judge in the U.S. District Court for the District of Columbia for an order, which the Commerce Secretary must justify by showing the least restrictive option available, and which automatically expires after 45 days ^[2].

That combination - a huge harm threshold, a single enforcement channel, and a short-lived remedy - is what turned the bill into an internet punchline once the Politico scoop spread. Readers fixated on the arbitrariness of drawing the line at exactly 50 deaths or $1 billion, as if 49 deaths or $999,999,999 in damage would be an acceptable cost of doing business. The independent evaluators meant to catch problems early don't change that calculus much either: they're required to flag an 'imminent enumerated risk' to Commerce within 72 hours of finishing an assessment, but nothing in the bill stops a lab from paying the evaluator it's being assessed by, as long as the pay isn't explicitly tied to the outcome ^[1].

Perhaps the most uncomfortable detail in the leaked text isn't in the bill itself but in the gap between Anthropic's public framing and what its own people reportedly believe. Internal commentary on the leaked draft, surfaced in reporting on the bill, included a blunt risk estimate from Anthropic's technical side that dwarfs the legislation's own thresholds: 'We give it a 10 percent chance it'll cause 8 billion' ^[3]. That figure - an order of magnitude above the bill's $1 billion trigger - sits awkwardly next to a policy team that chose to set the bar exactly where it did. Critics reading the bill's fine print have also flagged that without licensing requirements for the independent evaluators labs get to select and pay, the system risks a race to the bottom in auditor rigor rather than a genuine check on risk ^[4].

That tension is exactly what fueled the loudest public reaction to the story: the charge of regulatory capture. The sharpest version of that critique treats the bill less as a safety measure and more as a compliance moat - a framework whose costs are easiest for an incumbent like Anthropic to absorb and hardest for smaller, open-weight, or foreign competitors to match, while the company that wrote it gets to point to Washington and say it asked to be regulated. A more charitable reading, echoed in parts of the public reaction, credits Anthropic for proposing something concrete at all - but that reading still has to explain why the proposal's numbers land so far below what the company's own technical staff reportedly estimate the real risk to be.

Not every close reading of the draft lands on 'too weak.' A detailed legal analysis of the leaked text makes a counterintuitive argument: by creating an explicit statutory 'blocking order' - defined as an order suspending development, deployment, or internal use of a model found to present an imminent enumerated risk - the bill could actually increase the probability of successful federal court intervention against an AI lab, not decrease it ^[2]. Current law has no clear cause of action for AI catastrophic risk; a plaintiff today would have to improvise. Once Congress writes a specific mechanism into statute, courts have a defined path to use it, even if the triggering bar is high.

The same analysis complicates the preemption story too. Section 14(b)(1) appears to preempt state disclosure, evaluation, and reporting rules, but the state's authority to suspend a model outright may survive - and critically, the preemption of that suspension power is itself conditioned on the federal blocking mechanism in section 12 actually being enacted: 'this paragraph shall have no effect unless section 12 is enacted' ^[2]. That drafting detail means the simple 'Anthropic wants to gut state AI law' framing, however resonant online, understates how contingent and legally tangled the preemption actually is.

In June, Anthropic published its 'Advanced AI Framework,' arguing that absent a strong federal regime, states should keep their authority to legislate on AI ^[5]. Four months later, the bill it is pitching lawmakers would preempt exactly the kind of state rules that framework seemed to defend - including California's AB 2013, which requires model developers to publicly post high-level summaries of their training data ^[6]. Whether that counts as a reversal or simply the fulfillment of a conditional promise depends on how the federal blocking mechanism itself fares in Congress, but the optics are difficult: a company that spent the summer telling states they could keep regulating is now asking Congress to take some of that power away.

Anthropic's own spokespeople have tried to close that gap by insisting there is no inconsistency. One spokesperson described the draft as representing 'the proposals outlined in Anthropic's Advanced AI Framework, which we have been advocating for with federal and state lawmakers since June' ^[5], and another called it simply 'sensible policies we believe Congress should act on' ^[7]. Both statements frame the bill as continuity rather than retreat - but neither addresses why a framework that once protected state authority is now paired with language built to override it.

Anthropic isn't drafting into a vacuum. The bipartisan FRONTIER Act, from Rep. Lori Trahan and Rep. Jay Obernolte, already gives government stronger authority to stop deployment of risky models, including rulemaking power and emergency shutdown authority that Anthropic's draft notably omits ^[4]. Senate leaders including John Thune, Amy Klobuchar, and Ted Cruz are negotiating a parallel track, with Sen. Bernie Sanders pushing for stricter measures than either proposal contains ^[7].

Pitching its own text to lawmakers now, before any of those efforts converge, is one of the most concrete attempts yet by a frontier AI lab to shape binding federal law in its own preferred image rather than leaving the outcome to Congress. The money behind that effort has been rising sharply. Anthropic nearly tripled its first-half 2026 federal lobbying spend to $3.53 million, up from $3.1 million for all of 2025, including $1.97 million in the second quarter alone - more than Nvidia spent in that same quarter ^[8]. That followed what was reportedly Anthropic's biggest-ever lobbying quarter in April 2026, when it outspent OpenAI in Washington ^[9]. Read together, the spending and the bill look like the same bet: that getting a seat at the table early, with a draft already written, is worth more than waiting to react to whatever the FRONTIER Act or the Senate talks eventually produce.