What Anthropic's draft AI bill would let Washington block?

Fifty deaths or serious injuries, or $1 billion in damage. That's how bad a frontier AI model's risk would have to be before the government could block it under a draft bill Anthropic has pitched to lawmakers.

The Claude maker confirmed on Friday that it's circulating the draft, which Politico reported on first.

Under this draft, the release or usage of a model could be prohibited only if there is an imminent threat of over one billion dollars in damages or of death or severe injury to over 50 individuals. The threat should be due to a biological weapon, the escape of the model from the control of its creator or a cyberattack. This clause is similar to California's SB 53.

Independent evaluators from the Commerce Department registry could evaluate AI laboratories. Laboratories could pay them as long as the payments were not based on results. In case the evaluator identified an imminent threat, they should inform the Commerce Department within 72 hours.

Who can pull the plug?

Not many people. Commerce could ask a federal judge, and a judge for the US District Court for the District of Columbia could order a suspension. The bill explicitly bars any other department, agency or officer from suspending a model's development, deployment or internal use.

The bill would also override some state AI laws on risk disclosure, outside evaluation and incident reporting. Democrats have resisted such proposals, Politico noted.

The House FRONTIER Act, introduced by Representatives Lori Trahan and Jay Obernolte, gives the government the authority to halt the deployment of dangerous models. 

Meanwhile, in the Senate, Majority Leader John Thune, Senators Amy Klobuchar and Ted Cruz have drafted their own bill regarding catastrophic risks from AI.

Reporters noted that the requirements for $1 billion and 50 persons in the draft are high standards. The authors of the draft itself are also companies that will be regulated by it. An Anthropic spokesperson said that the bill is aligned with its Advanced AI Framework and called the policies "sensible".

Anthropic said that access to its two models was halted on June 12 because of the restrictions under US export control. Those were lifted on June 30, and access was re-enabled on July 1.