A year of deleting AI rules, then moving the liability away from the labs

On May 20, hours before signing, the president canceled the voluntary federal vetting order that would have touched a model before it reached the public. The order would have asked the leading labs to share their newest, most powerful models with the Treasury Department and the National Security Agency up to 90 days before release. The sharing was voluntary, with no penalty for refusing. Elon Musk, Mark Zuckerberg and David Sacks had warned the White House that even a voluntary review would harden into a mandate ^[1]. The president gave his reason days later.

I really thought that could have been a blocker. — Donald Trump

The year runs in one direction: strip away every rule that can constrain a model before it does harm, keep liability formally on the books, and put the government's weight on the labs' side of whatever is left. On October 10, the administration repealed the binding risk order that had governed frontier AI and replaced it with a voluntary pact signed by OpenAI, Anthropic, Meta, Google, SpaceXAI — the merged SpaceX and xAI — and Nvidia ^[2]. The pact encourages internal monitoring teams but sets no detailed safety measures and mandates no transparency ^[3]. The president offered his own description of what it accomplished.

I'm seeing tremendous self-policing. — Donald Trump

Two weeks before signing, Dario Amodei had named the standard any real constraint would have to meet.

The most effective method of pacing is via regulation that targets all U.S. frontier AI companies, as that covers even those who are unwilling to cooperate voluntarily. — Dario Amodei

The pact he signed reaches only the willing, on their own terms. The states were handled faster and harder. A House budget bill freezing state AI rules for a decade passed by a single vote in June 2025, and the Senate rejected the moratorium 99 to 1 ^[4]. By November the administration had drafted, then shelved, an executive order creating a Justice Department task force to sue states, with $42.45 billion in broadband funds withheld from any state that refused to repeal its AI laws ^[5]. The president put the goal in his own words.

one Federal Standard instead of a patchwork of 50 State Regulatory Regimes. — Donald Trump

Where states would not yield, the industry wrote the exceptions itself. Google supplied model-bill language to at least ten state legislatures, language that exempted its own chatbot products from liability ^[6]. Colorado's Chatbot Safety Act, as signed, carves out twelve categories of chatbot. Families of suicide victims have a name for those exemptions.

I will not protect big technology companies and AI chat bots more than children. — Katie Hobbs

Arizona's governor vetoed her state's version rather than sign it ^[6].

You can't legislate tech companies unless you have their input. — Sean Camacho

That is the removal. The question is where the weight landed. The administration's stated alternative is a liability-based model, in which companies self-regulate but remain legally responsible for the harms that result ^[7]. Liability stays alive on paper. The year then shows which way the weight tilts. In September the Justice Department filed its first intervention in the training-data litigation, backing OpenAI against the New York Times ^[8].

This Administration will never let our Nation be at a disadvantage relative to our foreign adversaries based on a plainly incorrect understanding of copyright law. — Stanley Woodward Jr.

In April the department had joined xAI's suit to strike down Colorado's AI law ^[9].

The Justice Department will not stand on the sidelines while states such as Colorado coerce our nation’s technological innovators into producing harmful products that advance a radical, far-left worldview at odds with the Constitution. — Harmeet Dhillon

And in October a federal appeals court added a rule of its own, holding that AI agents are legal tools of their human users rather than independent actors, which keeps liability off the developer ^[10]. In both fights where the government chose a side, it chose the labs'. None of this makes the labs immune. The Federal Trade Commission has open inquiries into consumer harms, and product-liability suits allege chatbots groomed minors and coached teenage suicides; OpenAI faces one over a Florida State University shooting ^[11][12]. Forty-two state attorneys general have subpoenaed OpenAI ^[13]. The accountability arrives afterward, in a courtroom. The endpoint is already named. xAI, suing Colorado, said plainly what it wants.

Government regulation that is applied at the state level in a patchwork across the country can have the effect to hamper innovation and deter competition in an open market. — xAI

In June the Supreme Court ruled 7 to 2 in the Bayer Roundup case that federal preemption bars state failure-to-warn suits — the claims arguing a company should have added a warning the federal regulator decided not to require — because a state tort suit would impose a requirement in addition to or different from the federal label ^[14]. Compliance with a federal standard defeats the state claim. No one has made that connection from Roundup to AI. But the administration supported the ruling, and its stated goal is one federal standard. If the labs write the federal benchmarks themselves, their own compliance becomes the defense. What makes that shape more than a thought experiment is the document on the other side of the year. This is the largest IPO window in history. SpaceX executed the biggest debut on record on June 12, raising $75 billion at roughly $1.8 trillion after acquiring xAI ^[2]. Anthropic filed its prospectus, the risk and finance disclosure every company must publish before listing, on September 29, targeting a $2 trillion valuation and up to $100 billion raised. A third of that document is risk factors, warning in its own words of "catastrophic or existential risks to humanity." Anthropic is committing $518 billion in mostly non-cancelable infrastructure spending ^[15]. OpenAI filed confidentially on June 8, then ruled out a 2026 debut and pushed toward 2027 ^[13]. A prospectus must disclose, in a company's own words, every material risk a reasonable investor would want: the copyright suits, the product-liability claims, the state laws it is fighting. That is where the year's two halves meet. The risks are disclosed in the labs' own words, under securities law. And the defense assembled against them is built from the same self-authored words: the compliance the labs write for themselves, pointed at a liability that now lands on users, deployers, and states that can no longer legislate. The labs are not immune. Nothing that arrived this year constrains a model before it can do harm. What did arrive turns the labs' own paperwork into their answer.