Klobuchar, Thune and Cruz would turn AI safety pledges into an enforceable duty of care

A duty of care has two moving parts: harm, and a failure to take reasonable precautions ^[5]. The first is decided after the fact by someone else. The second is decided against documents the developer wrote before release, and Anthropic has published a threat report raising doubts about whether its newer models could assist in developing biological weapons ^[10]. Under the standard the senators describe, that report is evidence of what the company knew.

So a lab that publishes candidly builds part of the record a plaintiff will use, and a lab that publishes less has less to explain. The description of the bill does not say whether completing the mandated tests would be a defence. That distinction sets the size of the number. With a safe harbour tied to the testing protocol, compliance is a fee for audits and reports ^[6]. Without one, it is an open tail that insurers will price slowly and expensively.

The more immediate cost is scheduling. Crypto Briefing reports that the bill would give the government authority to intervene and prevent the release of models deemed hazardous ^[7]. A launch date then becomes contingent on a review the developer does not run. Mandatory pre-release testing and independent third-party audits ^[6] also need auditors with capacity, and there is no queue today because there is no requirement today.

OpenAI has publicly advocated for mandatory national safety standards and testing protocols, a position it takes as a company that would be directly subject to the rules ^[8]. A lab already running internal evaluations pays less at the margin for a rule requiring them than a new entrant does. It also gets a federal answer to the question of what counts as enough. The other reading is that the position is sincere and costs OpenAI money.

The figure underneath all of this is a probability. Evan Hubinger, another Anthropic researcher, has publicly estimated a greater than 10% chance of an AI-driven extinction event within the next decade ^[9]; spread evenly across ten years, that is roughly a 1% chance a year ^[1]. Congress is legislating against an estimate produced inside a company the bill would regulate ^[4].

Passage is the weak link. Lawmakers have flagged the difficulty of writing technically sound rules for a technology that evolves faster than Congress can schedule hearings ^[11]. Some are reluctant to vote on a complex regulatory framework while running for re-election in November ^[12]. I would expect the pre-release review power to bind operators before the liability does, because a hold on a launch moves a revenue date this year while a duty of care is litigated over years. If the introduced text pairs the duty with a safe harbour for labs that complete the mandated tests, that expectation is wrong, and the whole thing prices as an audit fee.