4 hours ago
22
Anthropic, the company behind the Claude family of AI models, has handed US lawmakers a draft federal bill on frontier AI safety. It is the company’s own blueprint for how Washington should police companies like Anthropic.
Critics are not impressed. One opinion piece calls the proposal a fairly transparent attempt to calm public fears that Anthropic’s own people helped stoke, and argues that governments, not AI firms, should be writing the rules.
What the draft actually says
The confidential proposal was drafted in mid-October 2026. It has not been formally introduced as legislation, so for now it is a pitch, not a law.
The core of the bill is a set of thresholds. The federal government could block the release or use of an advanced AI model only if it posed a foreseeable, imminent risk of more than $1 billion in damages.
The other trigger is human harm. Regulators could step in if a model threatened to kill or seriously injure more than 50 people.
Not every kind of danger counts, either. The risks covered are limited to three scenarios: a model helping create or release a biological weapon, a model slipping out of its developer’s control, or a cyberattack.
Who holds the off switch
The draft names the Department of Commerce as the only agency that could go to court to pause a specific model’s deployment. Other agencies would not have that power under the proposal.
The bill also calls for repeated independent evaluations of frontier models. If an evaluator spots an imminent risk, it would have 72 hours to report it to Commerce.
Beyond the shutdown mechanism, the proposal includes a structure for whistleblower protections and incident reporting.
The California template
The draft borrows from California’s SB 53, which was signed into law in 2025. The proposal reflects that law’s thresholds while aiming for a higher bar before federal intervention kicks in.
The draft also ties into Anthropic’s Advanced AI Framework, which the company has been promoting to lawmakers since June 2026. In other words, this is not a sudden pivot; it is the latest step in a months-long lobbying effort.
Why critics are pushing back
The sharpest criticism centers on who is writing the rules. The opinion piece argues the bill exists largely to address concerns raised by Anthropic’s own current and former staff, and by its executive leadership, that the systems it is building could wipe out humanity.
AI safety advocates have also argued that the intervention thresholds are too lenient. They point to other proposals, including the bipartisan FRONTIER Act, as offering stronger oversight.
The $1 billion and 50-person thresholds are the obvious targets. Requiring that a risk be both foreseeable and imminent narrows the window further, since harms that are plausible but not yet looming would fall outside the bill’s reach.
Anthropic has acknowledged that parts of the language need work. The company committed to refining certain wording by October 18, 2026.
The bigger fight over who regulates AI
The debate here is less about the specific numbers and more about a familiar question: should an industry draft its own guardrails? The opinion piece comes down firmly on one side, arguing that governments must regulate AI firms rather than let them set their own terms.
Industry-drafted legislation is not new in Washington. What makes this case unusual is that the company pitching lighter-touch triggers is also one of the loudest voices warning about catastrophic AI risk.
What this means
For lawmakers, the draft is one more option on a crowded table. Anthropic’s version will now compete with proposals like the FRONTIER Act, and the contrast between them will likely shape how strict any eventual federal standard turns out to be.
The single-agency design deserves close attention. Concentrating shutdown authority in Commerce could speed decisions, but it also creates a single point of failure if that agency is slow, under-resourced or reluctant to act.
The revised language Anthropic has promised is the next thing to watch. Whether the company lowers the thresholds, broadens the covered risks or simply tidies the wording will signal how seriously it is taking the backlash.
Then comes the harder test: whether any member of Congress formally introduces the bill. Until that happens, the draft remains a statement of what one AI company would like the rules to be, which is precisely what its critics are worried about.
Disclosure: This article was edited by Diego Almada Lopez. For more information on how we create and review content, see our Editorial Policy.
English (US) ·