Anthropic drafts its own AI safety law for US Congress
WASHINGTON: Anthropic, the company behind the AI assistant Claude, has confirmed it is circulating a draft bill among American lawmakers to regulate artificial intelligence, stepping into a debate Washington has failed to settle for years.
The proposal was first reported by Politico on Friday, citing three people with knowledge of the effort. The Anthropic AI safety bill would let the government block a model’s release or use only in narrow, catastrophic cases, and only with a judge’s approval.
What the Anthropic AI safety bill proposes
Under the draft, Washington could halt a model if there is a foreseeable, imminent risk of killing or seriously injuring more than 50 people, or of causing more than $1 billion in damages.
The risk would have to stem from one of three causes: the creation or release of a biological weapon, the model evading its developer’s controls, or a cyberattack, Politico reported.
The bill would require frontier developers to test their most powerful models, publish the findings, submit to independent evaluation and maintain a robust security programme. Independent evaluators would register with the Commerce Department and assess labs periodically. Companies could pay these evaluators, as long as payment was not tied to the outcome.
If an evaluator found an “imminent enumerated risk”, it would have to report to the Commerce Department within 72 hours. Commerce could then ask a federal judge in Washington DC to suspend the model, and no other federal agency would be allowed to do so.
Why Anthropic wrote its own law
A company spokesperson told Reuters the bill reflected proposals in Anthropic’s “Advanced AI Framework”, released earlier this year. The proposal, the spokesperson said, grants the government “the ability to block or deter dangerous deployments of the most powerful models” and represents “sensible policies we believe Congress should act on”.
It is the first time a leading AI lab has drafted legislation for Congress itself, rather than simply lobbying against someone else’s. The move follows revelations of security incidents caused by AI agents that escaped their developers’ control. Feature Pakistan recently reported on Tencent’s admission that its AI agents showed “rogue” browsing behaviour, and on calls for safety laws at a Dublin summit.
Anthropic has had a rough year with Washington. The Trump administration banned the company from the Pentagon in February, and in June forced it to suspend its most advanced models for nearly three weeks.
Chief executive Dario Amodei last month urged the industry to slow down, warning AI agents could take control of “the entire internet” within a year. Trump dismissed such calls as coming from “negative forces”.
Long odds in Congress
The bill faces a steep climb. Congress is out of session until after next month’s midterm elections, and the United States still has no federal AI law. Policy under Trump has relied on voluntary industry pledges and executive orders meant to keep an AI boom going.
The draft would also override some state laws, including California’s, which could draw fire from states that have moved faster than Washington. Charlie Bullock of the Institute for Law and AI called the proposal “pretty reasonable” but “weaker” than the rival FRONTIER Act introduced in the House this summer.
For Pakistan, the debate matters. The country is pushing AI ambitions while wrestling with basic infrastructure, as our look at Pakistan’s AI push showed. Whatever rules America sets will shape the tools Pakistani developers and businesses end up using.
What happens next
With Congress away until after the midterms, the draft is best read as an opening bid for next year’s legislative session. If it advances, it would give the US its first real lever over the most powerful AI systems, while deliberately tying the government’s own hands everywhere else. The original report was published by Dawn, with further detail from Reuters.
FAQ
What does the Anthropic AI safety bill propose?
It would let the US government, through a judge, block an AI model’s release or use over an imminent risk of 50+ deaths or serious injuries, or $1bn+ in damages, from bioweapons, loss of control or cyberattacks.
Who would enforce the proposed AI law?
The Commerce Department would register independent evaluators and could ask a federal judge in Washington DC to suspend a model. No other federal agency would be allowed to halt AI development.
Has the US passed any federal AI law?
No. AI policy under President Trump has relied on voluntary industry pledges and executive orders, and Congress is out of session until after the midterm elections.
