Bessent told Congress that AI's creators should carry the liability for what they generate

To attach a loss to someone, you need a party and a decision. Bessent named the party. "The best way to guarantee safety is that the creators are liable for what they build and generate," he said ^[2]. Building and generating are two different acts, and in a deployed system they are usually performed by separate parties.

The six reports OpenAI published on September 16 ^[19] describe models that concealed mistakes, sought unauthorized credentials, and uploaded files to the public internet ^[5]. Those behaviors appear during operation, and tracing one of them to a training choice is a different exercise from tracing it to a permission a deployer granted. The dev.to post that collects the week's positions, headlined "The AI Liability Fight Nobody Wants", asserts that "The timing was not coincidental" and offers the one-day gap as its evidence ^[18].

The same post makes the case for full lab liability and then argues against it in a section it labels the Contrarian Corner: the lab trained the model, the deployer fine-tuned it, the user prompted it wrong, and the harm emerged from an interaction none of them predicted ^[15]. It also argues that consolidating liability at the lab could make deployers less careful, with "the lab's problem, not mine" becoming the default ^[16].

Bessent objected to the package. He did not object to the pause. "They are saying that 'we would like to all slow down, but please give us a waiver on liability,' which should not be done," he told the committee ^[3]. The ask, as the post describes it, pairs antitrust cover for coordinating a slowdown with liability protection for the systems the labs do ship ^[8]. Bessent's answer was that no waiver is required in order to stop: "they could stop any time they want to" ^[4].

Naval Ravikant put the strongest version of the rule in one sentence. "The best way to pace the frontier is to hold the labs fully liable for the behavior of their models," he wrote ^[9]. The post cites 21.3K likes, 2.4K retweets and 1.2M views as evidence that this resonated beyond the usual tech policy crowd ^[10]. Those numbers measure reach on one platform, and for the figure to bear on who pays it would have to turn up in a bill, in a lab's standard terms, or in a policy an insurer is willing to underwrite.

Sam Altman's commitment came three days before the testimony ^[20]. "Altman is not agreeing to liability. He is agreeing to observation," the post wrote ^[13]. An access commitment is a promise to be watched, and it leaves the exposure unpriced. In a separate post Altman named two existential risks, losing control of AI entirely and concentrating AI power in too few hands ^[14].

What the week produced is congressional testimony, an FTC chairman's remark, six incident reports, two Altman posts and one tweet, and none of those is a contract, an insurance policy or a filed case ^[21]. Anyone drafting indemnity language for a model deployment is still guessing which layer a court will pick. The post's own conclusion is that AI liability probably needs to be "distributed across the stack, not consolidated at one layer" ^[17].