In December 2025, President Trump signed Executive Order 14365 with a specific goal: prevent states from enforcing AI laws that the administration considered “inconsistent” with federal policy. The EO established an AI Litigation Task Force whose job was to challenge state laws in court and make states with “onerous” AI rules ineligible for non-deployment federal funds. It was an aggressive move — arguably the most consequential executive action on AI since Biden’s order was revoked.

It hasn’t worked as intended.

Congress weighed in with a 99-1 Senate vote to strip a 10-year federal moratorium on state AI laws from the “One Big Beautiful Bill Act.” That wasn’t a close call. The political calculation was straightforward: most senators represent states that are themselves passing AI legislation, and the idea of a federal government telling states they can’t regulate the algorithms affecting their citizens’ healthcare, hiring, and criminal justice systems was a vote-loser in both parties. The preemption strategy, as currently constructed, has no legislative path.

And so state legislatures have kept going. As of March 2026, lawmakers in 45 states have introduced 1,561 AI-related bills — surpassing the total volume from all of 2024. Colorado and California have passed comprehensive AI laws that are now enforceable or approaching their effective dates. These aren’t vanity resolutions. They create real compliance obligations for companies operating in those jurisdictions.

The irony is that the federal government has, through enforcement actions and NIST partnerships, quietly acknowledged that some oversight is necessary. The announcement that CAISI — the renamed US AI Safety Institute — has signed evaluation agreements with Google, Microsoft, and xAI is notable precisely because of who it comes from. The Trump administration came into office explicitly opposing AI oversight. It is now a party to formal evaluation partnerships with frontier AI developers. This isn’t because it changed its philosophy. It’s because Anthropic’s Mythos model — with its advanced cybersecurity capabilities — apparently prompted a national security reconsideration at the senior levels. The government’s own intelligence assessment, reportedly, concluded that certain frontier models posed risks that couldn’t be left to market incentives.

The US is left in an unusual equilibrium: a federal government that publicly opposes AI regulation but quietly participates in it, states that are legislating with increasing sophistication and speed, and companies that are being pulled in both directions simultaneously. Planning for this environment requires assuming that both tracks continue — federal posturing and state enforcement — and that the real compliance burden lives in the state layer that Washington can’t seem to neutralize.

Whether this is a stable arrangement is unclear. What is clear is that for anyone building or deploying AI systems in the US, the state laws are not theoretical. They’re already here.