The White House released a four-page national AI policy framework on March 20. It is not long. It does not need to be. The document, developed by OSTP Director Michael Kratsios and White House AI/Crypto adviser David Sacks, lays out what the administration wants Congress to do about artificial intelligence. The headline ask is preempting state AI laws. The real ask is buried in paragraph three.
“Congress should avoid open-ended liability that could give rise to excessive litigation,” the framework states. That single sentence, if enacted into legislation, would make it substantially harder to sue AI developers when their models cause harm. It would reshape the legal architecture around AI accountability in the United States. And it landed in a four-page document released on a Thursday afternoon.
Who Wrote It, and Who Benefits
This matters more than it’s getting credit for. Kratsios ran Thiel Capital before joining the White House Office of Science and Technology Policy. Sacks is the founder and general partner of Craft Ventures, a venture capital firm with investments across the AI sector. Before this role, he was PayPal’s COO. He is now the White House’s top adviser on AI policy. He co-authored a framework that proposes shielding AI developers from state-level regulation and from liability for third-party misuse of their products. Those are the facts.
Nobody is calling this corruption. But the revolving door between the AI industry and AI policymaking is operating in public, in real time, and the framework reads exactly the way you’d expect a document written by AI investors to read. It protects builders. It limits legal exposure. It consolidates regulatory power at the federal level, where industry lobbying is most effective.
The framework follows Trump’s December 2025 executive order directing a review of state AI laws. That order specifically targeted regulations the administration considered obstacles to AI development. The framework is the output. It asks Congress to “preempt state AI laws that impose undue burdens” and replace them with “a minimally burdensome national standard, not fifty discordant ones.”
That language sounds reasonable in isolation. But the framework also says Congress should not “penalize AI developers for a third party’s unlawful conduct involving their models.” If someone uses an AI model to commit fraud, produce illegal content, or cause measurable harm, the developer who built and deployed the model should bear no legal consequence. That is not deregulation. It is immunization.
The Liability Shield
The preemption language is getting the coverage. State attorneys general are furious. But the liability provisions are where the framework does its real work.
There are currently no federal AI liability standards. States have been filling the gap. California, Colorado, Illinois, and others have passed or proposed laws that hold AI developers accountable when their products cause demonstrable harm. Some of these laws are clumsy. Some are well-crafted. All of them give citizens a path to court when AI systems damage their lives.
The framework proposes to erase that patchwork without replacing it with anything equivalent at the federal level. No new regulatory body. No federal enforcement mechanism. Seven categories for Congress to consider, including children’s safety, copyright, and “community effects,” but no specific accountability structure for when AI systems produce harmful outputs.
On child safety, the framework affirms that COPPA applies to AI and calls for mandatory parental controls. TechCrunch reported that critics say this shifts the safety burden from developers to parents. That tracks with the framework’s broader logic: companies build, users manage the consequences.
The Anthropic Contradiction
The administration’s AI policy has an internal coherence problem, and nobody seems to be putting the two pieces together. Laterstack has been tracking the Anthropic situation since February. The same administration that blacklisted Anthropic for refusing to remove safety restrictions from its military AI is now proposing to shield AI developers from liability for third-party misuse of their models.
In one case, the government punishes a developer for limiting how its product can be used. In the other, it proposes protecting developers from consequences when their products are misused. Those two positions are contradictory on their face. Unless you read them as a single policy: comply with government requests for unrestricted access, and in exchange, you are shielded from accountability. Refuse, and you are designated a supply chain risk.
The framework does not mention Anthropic. It does not need to. The message is structural.
The Opposition, and the Question Nobody Is Asking
Thirty-six state attorneys general, from both parties, issued formal objections to the preemption provisions. More than 280 state lawmakers from both parties signed a letter urging Congress to reject federal preemption of state AI regulations. The Alliance for Secure AI said the framework “provides no path to accountability for AI developers.” The Center for Data Innovation, a pro-industry think tank, praised it as a step toward regulatory clarity. That split tells you roughly who benefits.
But here is the question that should be getting more attention: does this framework have any realistic chance of becoming law? This is not legislation. It is a wish list. Congress would need to draft and pass actual bills to implement any of it. And 36 attorneys general from both parties plus 280+ state legislators from both parties is an extraordinary level of bipartisan opposition. That kind of unified state-level resistance kills federal preemption proposals. It has happened before, on everything from data privacy to gun regulations.
So what is the framework actually for? If it cannot pass Congress, it still serves a purpose. It sets the terms of debate. It tells the AI industry where this administration stands. And read alongside the Anthropic blacklisting, it delivers a message to every AI developer in the country: build what the government wants, remove restrictions when asked, and you will be protected. The framework may never become law. The signal it sends already has.