Anthropic Proposes AI Safety Bill to US Congress

American company Anthropic submitted a draft bill to members of the US Congress to regulate frontier artificial intelligence models, according to a Politico report cited by Reuters. The report quoted three people familiar with the matter saying the company quietly presented the draft to lawmakers, while an Anthropic spokesperson said the text reflects the framework announced by the company earlier this year under the name "Advanced AI Framework".

The draft allows the government to prevent the launch or use of a model only if it poses a foreseeable and imminent risk of damages exceeding $1 billion, or the death or serious injury of more than 50 people. The report specifies that the risk must stem from the development of a biological weapon, a model escaping its developer's controls, or a cyberattack.

The text requires developers to test frontier models, maintain transparency regarding results, and subject them to independent evaluation. Independent evaluators would be registered with the Department of Commerce, which could ask a judge to halt a model if a risk is proven, with the final decision resting with a federal court in the District of Columbia. The text also prevents any other government agency from suspending development, deployment, or internal use.

Charlie Bullock from the Institute for Law and AI considers the draft "largely reasonable" and "more than what other frontier labs have done," but noted that it is weaker than the FRONTIER Act introduced in the House of Representatives. The draft comes amidst a legislative deadlock in Washington, where President Donald Trump relies on voluntary policies and executive orders, while lawmakers fear incidents involving AI agents that break free from their developers' control.

What do these terms mean?

Frontier Model: A highly capable artificial intelligence model at the cutting edge of technology that requires special safeguards due to potential risks.

Independent Evaluation: An assessment conducted by an external expert or entity independent of the development company to verify the model's capabilities and risks.

Executive Order: A decision issued by the head of state or government without requiring parliamentary approval, binding executive agencies to implement it.