On Capitol Hill, the folders that circulate in the office buildings along Independence Avenue are usually written by committee staff. A set of them that moved through meetings over the past several weeks was written with the help of a company that builds the technology it would regulate.
Anthropic’s representatives have spent weeks walking members and staff through a draft bill on frontier artificial intelligence, according to a report from Politico on October 16. The draft sets the bar for a government-ordered shutdown of a model at a level that is deliberately high: a foreseeable and imminent risk of more than 1 billion dollars in losses, or of fifty or more deaths and serious injuries.
The risk has to come from one of three sources. The first is a biological weapon, meaning a model that materially helps someone produce or deploy one. The second is a model that escapes the control of its developer. The third is a cyberattack of a severity the draft defines.
The machinery around those triggers is where the detail sits. Evaluators would register with the Commerce Department. A developer that finds a problem would have 72 hours to report it. Only a federal judge in the District of Columbia could order a model paused, a provision that routes every emergency dispute to a single courthouse and away from state attorneys general and agency hearing officers.
The draft would also preempt some state laws covering risk disclosure and incident reporting, though not, according to the description, state rules on other subjects. That clause is the one that will draw the most opposition, because several states have already passed statutes in this area and their legislators did not write them to be replaced by a federal text drafted in part by an interested party.
A company writing its own regulator’s rulebook is not new. Airlines helped design the certification regime they live under, drug makers negotiated the user-fee system that pays for their reviewers, and banks have commented on capital rules for as long as there have been capital rules. Industry involvement brings technical knowledge legislators lack; it also produces thresholds calibrated to what the industry can bear.
The calibration is the interesting part here. A billion dollars is a large number for a harm threshold, larger than the penalties that follow most product failures, and it means a model could cause substantial damage without meeting the test. Fifty deaths is a mass-casualty line, the kind of figure that appears in chemical-plant rules and in the definitions used for catastrophic events.
Supporters of a high bar argue that a low one would produce frequent false alarms, that a regulator empowered to pause a model on thin evidence would chill development, and that a system which fires constantly gets ignored. Critics argue that a threshold written in dollars and bodies is a threshold that is met only after the harm has happened, which makes the power to pause a model a power to respond rather than to prevent.
The three categories are narrower than the general language about safety that usually appears in these debates. They cover weapons of mass destruction, loss of control and cyber, and exclude the automated production of persuasive content, the displacement of work and the use of a model in surveillance. Those exclusions are choices, and civil-society groups will contest them in hearings.
The 72-hour reporting window is tighter than most incident-reporting regimes in financial services, where the standard is often four business days. A three-day clock assumes the developer can tell, within three days, whether an event meets the statutory definition, which requires both monitoring and a legal judgment made under time pressure.
Routing the pause power to a single district court is the provision with the clearest institutional logic and the clearest political cost. On the logic side, emergency orders on a national technology raise questions that one court can accumulate expertise in. On the cost side, state attorneys general who have filed their own complaints against technology companies would lose a forum, and members of Congress from both parties have objected to provisions that displace state authority.
The politics of passing anything are poor. Congress has held hearings on artificial intelligence for years without producing a floor vote on a comprehensive bill. A draft that arrives with a company’s fingerprints on it can move faster through the gap between hearings and legislation, or it can become the exhibit that opponents use to argue that the rules were written by the regulated.
Anthropic’s commercial interest in the outcome is not hidden. Rules that require third-party evaluation, incident reporting and registered auditors impose fixed costs that a large lab can absorb and a small one cannot, and they raise the price of entering the frontier. The same is true of the preemption clause, which replaces a patchwork of state requirements with one national standard that a company with a Washington office is better equipped to navigate.
What the draft does not address is the part that researchers have asked about most often: how a model’s capabilities are measured before release, who pays for the evaluations, and what happens when an evaluator and a developer disagree about a result. Those questions sit underneath every threshold in the text.
For the members and staff who received the folders, the decision is whether to treat the document as a draft to be edited or as a proposal from a stakeholder with an interest. Both readings are available. The bill’s thresholds are high enough that the industry can live with them, and low enough that its authors can say they wrote a real constraint. Which of those descriptions survives the committee process will depend on who is in the room when the language is marked up.
