The Shield Is Dead, and the Industry Wants the Rules

Senator Cynthia Lummis put the trade plainly. Her RISE Act offered the people who build AI models something almost no other industry enjoys: immunity from civil lawsuits over what their models do. The price was disclosure — publish the model's specifications so outsiders can see what they're dealing with — and the fine print left the liability exactly where it already sat. Doctors, lawyers, engineers, anyone with a license who used the model stayed fully on the hook.

This legislation doesn’t create blanket immunity for AI — in fact, it requires AI developers to publicly disclose model specifications so professionals can make informed decisions about the AI tools they choose to utilize. — Cynthia Lummis

The bill did not travel alone. It arrived inside a package whose other provisions did the rest of the work: a ten-year bar on states writing AI rules of their own, and a threat to strip federal broadband money from any state that tried anyway ^[1]. Stated cold, the trade was immunity for the builders in return for a decade of silence from Sacramento and Albany. That shield was already dead before Tuesday. The package stalled under full Republican control, not for lack of power but for lack of agreement on the preemption question — more than 250 state lawmakers from both parties signed on against Washington telling their states to stop ^[2][3]. The industry was losing its protection while its patrons still ran Congress. The sweep didn't create the threat; it formalized it. The new majority's first announced agenda is the mirror image of Lummis's bill: mandatory third-party audits of the big developers, civil liability for harms a model could have foreseen, and consumer protections on energy costs ^[4]. None of it is law yet. Liccardo, the Democrat driving the safety bill, sets the test himself.

this is the time for us to move forward with sensible AI regulation. — Sam Liccardo

The one certain change Washington made this week is what it killed, not what it created. The president's position has moved, and the flip reads as a matter of venue, not principle. Last year his signature package would have barred state AI rules for a decade ^[1]. In October, his people answered questions about AI harm with the liability law already on the books, enforced through Justice Department intervention ^[5]. This week, facing the Democrats' federal plan, he called it overreach and said he would rather the states handle it ^[4]. The states he now prefers are already in motion on their own track — California requires lawyers to disclose generative AI use in court filings, and a New York proposal would hold chatbots liable for legal advice an unlicensed person couldn't lawfully give ^[6]. Step away from Washington and the pattern sharpens: the rules that actually bind were arriving all year from venues no election controls. A utility tariff from August 2025 was the first of them, a rate written specifically for the data centers ^[7]. Ohio and Indiana regulators approved tariffs the hyperscalers signed in one state and protested in the other, and California ordered its own ^[8] — against Washington's answer to the same fight, which was a voluntary pledge ^[9]. Product-design suits over child safety against Google and OpenAI sit in the shadow of Meta's settlement with state attorneys general, and Anthropic flagged the same litigation risk in its own prospectus ^[10]. Agent breaches drew no new statute, only the FTC already investigating under the powers it already has ^[11]. And the insurance market, unable yet to price the risk, wrote exclusions into standard policies while a law firm — precisely the licensed professional Lummis's shield would have left on the hook — bought $5 million of liability coverage from a startup called Testudo so its agents could do unsupervised legal work ^[12]. The industry's own words kept the rule simplest. When an employee's autonomous agents probed a competitor's systems, Railway's chief executive banned the testing and stated the rule he now lives by ^[13].

You operate the machinery, you are responsible for that what that machinery does. — Jake Cooper

The strange part of this week is not that the industry is fighting the new Congress. It's that the industry isn't. Liccardo says the leaders at OpenAI and Anthropic support the safeguards ^[14].

I think there's a certain perception in the media that that somehow the notion of supporting AI safety, uh, is, is antithetical to what the industry wants. — Sam Liccardo

One reading of why — and only a reading — runs through an antitrust filing. In September, Amodei called for slowing frontier capabilities, and Altman and Musk said they agreed; within a week a class action followed, alleging an unlawful agreement to stifle development ^[15]. A federal mandate would give the labs cover for restraint they cannot legally volunteer, since a law that requires the slowdown is not a conspiracy. No one has stated that motive. The same week put the liability question on a deadline of its own. Anthropic is racing to file its IPO while keeping the founders' vote even as it sells the equity — the very liability the shield would have capped, and which scholars had urged the company to waive before going public ^[16][17]. Close the circle on Lummis's trade and the clause has inverted. The disclosure the shield sold in exchange for immunity is now being extracted by mandatory audits, and the immunity is gone — replaced by liability for harms a model could have foreseen ^[4]. The builders, by their would-be regulator's account, are at the table rather than the barricade. Whether a bill can clear 60 votes is the one part no one will know until it is put to a vote.