The conventional wisdom about American AI regulation was so widely held it stopped sounding like a prediction and started sounding like weather. Washington would stall, the story went, so the states would act — California, Colorado, Texas, New York — and companies would face a fifty-jurisdiction patchwork. To survive it, you'd build to the strictest rule and ship that everywhere. The patchwork was the threat, and the strictest state was the de facto regulator of the whole country.
Six months into 2026, that map is being redrawn so fast it's worth stopping to notice the direction. The patchwork isn't thickening. It's unraveling — and from both ends at once.
Colorado deleted its own flagship
Start with the law that was supposed to be the model. Colorado's AI Act, passed in 2024, was the first comprehensive, risk-based state AI statute in the country — a near-cousin of the EU's framework. Its core was a genuine obligation: developers and deployers of "high-risk" systems owed a duty of reasonable care to protect consumers from algorithmic discrimination, backed by impact assessments and disclosures to the attorney general. That duty was the one substantive thing any American AI law actually demanded. Everything else was paperwork; this was a standard you could be sued under.
It was scheduled to take effect February 1, 2026. Then, in a special session, the legislature pushed the date to June 30, 2026 (Alston & Bird). Then, on May 14, 2026, Governor Polis signed SB 189, which pushed it again — to January 1, 2027 — and, in the same stroke, gutted it. The amended law strips out the duty of care, the algorithmic-discrimination framework, the risk-management programs, and the impact-assessment regime, leaving a narrower scaffold of disclosure and transparency around automated decisions (Hunton).
Read that sequence twice. The most substantive obligation in any U.S. AI law was repealed by the same state that invented it, before its effective date, before one company ever had to comply. The duty of care didn't lose a court fight. It never got to the fight.
The flagship statute wasn't struck down. It was quietly deleted by its own authors, three weeks before anyone had to obey it.
Texas shipped narrow on purpose
The other big state to act didn't build a Colorado either. The Texas Responsible AI Governance Act (TRAIGA, HB 149) took effect January 1, 2026 — and it is deliberately, almost defensively narrow (Norton Rose Fulbright). The original draft had Colorado-style ambitions; the version that passed was pared back to intent-based liability for specific harms — using AI to incite self-harm, to facilitate crime, to produce certain illegal content — plus restrictions aimed mostly at government use, and a regulatory "sandbox" for testing. There is no general duty of care. The drafters looked at the comprehensive model and chose, on purpose, to ship something far smaller.
So the two largest red-and-purple-state experiments converged on the same answer: not the EU-style risk regime everyone expected, but thin, harm-specific, transparency-first rules. The strictest-state-wins logic assumed the states were racing to be strict. In 2026 they're racing each other to be modest.
And the federal government is suing the rest
The second end of the unraveling is in Washington — not as legislation, but as litigation. Congress already rejected the blunt instrument: a proposed ten-year moratorium on state AI enforcement, attached to last year's reconciliation bill, was stripped out by a Senate vote that was nearly unanimous. Preemption-by-statute failed.
So the administration switched weapons. Executive Order 14365, signed December 11, 2025, directs the Attorney General to stand up an AI Litigation Task Force to challenge state AI laws the administration deems hostile to U.S. "AI dominance" — arguing, case by case, that they unconstitutionally burden interstate commerce or are preempted by federal policy, while threatening to condition federal funding on state restraint (White & Case; Paul Hastings).
Here's the part the headlines underplay: an executive order can't actually preempt a state law. Preemption flows from Congress, and Congress declined. What the order really does is convert the federal government from a missing regulator into an active anti-regulator — a litigant whose job is to make passing AI rules legally expensive enough that statehouses think twice. The deterrent isn't a rule. It's the lawsuit you'll lose money fighting.
The Brussels effect meets a vacuum
Now set the American picture against Europe's, because they are moving in opposite directions on the exact same calendar. While Colorado was deleting its duty of care, the EU was centralizing — reinforcing the powers of its AI Office and consolidating oversight of general-purpose models — and counting down to August 2, 2026, when penalties for GPAI providers become enforceable (European Commission). The "Brussels effect" — the theory that the strictest large market sets the global default — was supposed to drag the U.S. along by sheer gravity.
It still might, for any company that ships into Europe. But domestically, it's now pulling against a vacuum. The genuinely non-obvious lesson of the first half of 2026 isn't which level of American government wins the AI-rulemaking fight. It's that the substantive standard everyone organized around — care, applied to discrimination, enforceable by lawsuit — turned out to be the first thing every American actor abandoned the moment it got close to real. Colorado abandoned it to avoid enforcing it. Texas declined to write it. Washington is suing to keep anyone else from trying.
The patchwork was never the danger. The danger, for anyone who wanted AI held to an actual standard of care in the United States, is that the patchwork is being erased faster than it was ever sewn — and the blank space it leaves behind is not neutral. It's the policy.
