Anthropic has circulated a draft federal bill that would let a court block a new AI model if it could cause mass casualties or major damage. TechCrunch reports the trigger is a foreseeable, imminent risk of more than 50 deaths or more than $1 billion in property damage. The draft is confidential and not enacted law, and it has no reported introduction date. For founders, it matters less as a rule today than as a sign of how frontier AI oversight may be designed.
What Would Trigger a Block
The risk has to come from one of three sources. According to the reporting, the draft covers a biological weapon, a model evading its developer's controls, or a cyberattack. The harm has to be both foreseeable and imminent.
The thresholds are reported differently. TechCrunch describes "more than 50 deaths" and "more than $1 billion." TechSpot describes "at least" $1 billion or "50 or more" people. The exact wording matters at the margin, so wait for the bill text before relying on either version.
The thresholds describe a single event. The proposal is aimed at catastrophic incidents, not routine harms. That is why critics say the trigger is narrow.
Why a Judge Decides
Evaluators come first. Under the draft, independent evaluators registered with the Commerce Department would assess a lab's frontier models. If an evaluator finds an imminent catastrophic risk, it has 72 hours to report to Commerce.
Commerce asks a court to act. Commerce would present the evidence to a federal judge in Washington, D.C., who would decide whether release should stop. According to TechCrunch, no federal agency could halt distribution on its own.
This shifts the decision. Routing the choice through a judge gives companies a formal hearing rather than an administrative order. It also means the timing and outcome depend on a court, not on an agency's discretion.
The Evaluator Conflict Question
Labs would pay the evaluators they use. TechCrunch says labs could pay evaluators "provided payment is not contingent on the evaluations' results." TechSpot describes the same rule as applying only if evaluators' independence "remains unaffected." Neither report explains how payment arrangements would be checked.
This is the sharpest criticism. A lab that pays an evaluator has an interest in the result, even if no bonus depends on it. Without rules on fees, rotation or disclosure, the independence safeguard is hard to verify. Founders who sell into regulated markets already know this problem from audits.
How It Compares to California's SB 53
California already has a frontier AI law. According to Orrick's summary, SB 53 uses the same 50-death and $1 billion thresholds, applies to frontier models trained with more than 10^26 operations, and covers developers with more than $500 million in prior-year revenue. It took effect January 1, 2026.
SB 53 focuses on disclosure. Under that summary, large developers must publish safety frameworks, release transparency reports before deploying new frontier models, and report critical incidents to the state. Penalties can reach $1 million. The federal draft goes further by allowing a court to halt release.
Read the statute, not the summaries. I checked these details against secondary summaries, not the law's text.
Where the Industry Stands
Positions differ widely. TechCrunch reports that DeepMind chairman Demis Hassabis has suggested a self-governing AI body modeled on FINRA. Nvidia CEO Jensen Huang has argued the industry does not need new laws. TechCrunch says Anthropic's proposal goes further than either.
Labs have resisted mandatory oversight. TechCrunch notes that AI labs have generally favored voluntary evaluations over binding rules. The draft's supporters will say a voluntary system has not been enough. Its critics will say a judge-led system creates a new kind of uncertainty.
Read the critiques carefully. TechSpot's author argues the thresholds seem to allow $999 million in damage or 49 deaths without intervention, which the author calls "eyebrow-raising." That is one reading of the draft, not a settled judgment.
What It Means for a Small AI Company
Start by checking whether you're in scope. The draft's full thresholds haven't been published in the reporting I could read. SB 53's compute and revenue thresholds are a reference point, but they don't tell you whether the federal draft would apply to you.
Document your evaluations now. Keep records of how you tested each model for dangerous capabilities, what you found, and what you changed. These records would matter under any evaluator-based regime.
Plan an incident process. SB 53 requires critical incident reporting within 15 days, and a 24-hour notice when risk is imminent. Even without a federal mandate, a written process for escalation helps you respond quickly.
Recommendation Blocks
Watch closely if: - You train or fine-tune models above the compute levels SB 53 uses - You sell AI products to regulated industries such as finance or healthcare Focus on documentation if: - You build on top of other labs' frontier models - You want to show customers and investors you test for misuse Wait for the text if: - Your models are small, and the draft's thresholds may not reach them - You're making product or hiring decisions that depend on how the law reads
A Decision Framework
Step 1: Map your exposure. Identify whether your models or company size could fall under SB 53 today and under the federal draft later.
Step 2: Document testing. Keep a dated record of evaluations and their results.
Step 3: Write an escalation plan. Name who reports incidents, to whom, and within what time.
Step 4: Track the bill. Follow the text once it's introduced, and compare thresholds against your own numbers.
Bottom Line
The proposal is a draft, but it shows where frontier AI oversight may go. A judge-led block would be unusual and could be decisive, and its effectiveness depends on how independent evaluators are paid and checked. For founders, the practical steps are the same whether or not this bill passes: know your exposure, keep records, and plan how you would report an incident.
This isn't legal advice. Talk to counsel before relying on any draft or summary.
Explore More on Automation Opportunities
Visit Bitroot's homepage for founder guides, tools, and automation resources built for early-stage SaaS teams.
