Two AI “moratorium” fights are running in Washington and the states right now, and they are not the same fight. One is about whether AI development itself should pause. The other — the one this piece tracks — is about whether federal action can stop states from enforcing their own AI safety laws, like California’s SB 53 or Colorado’s AI Act. As of September 2026, that second fight is still unresolved, and the past four months have moved it in several directions at once.
How it got here, briefly
The federal-preemption push has already gone through one legislative defeat and one executive-branch pivot. In May 2025, the House-passed reconciliation bill included a 10-year moratorium blocking state enforcement of AI regulation; the Senate stripped it by a 99-1 vote before the bill’s final passage that July. Five months later, the administration tried again through executive action: Executive Order 14365, “Ensuring a National Policy Framework for Artificial Intelligence,” signed December 11, 2025, directed a Justice Department litigation task force to challenge state AI laws, told the FCC and FTC to develop federal-preemption theories of their own, tied some federal funding to states dropping “onerous” AI rules, and — in its Section 8 — instructed the administration to prepare a proposed federal preemption bill for Congress to take up. CASRAI’s guide to both executive orders covers that order’s mechanics in full; this piece picks up where it leaves off, with what has happened since.
The 2026 attempt: a discussion draft, not yet a bill
That congressional bill Section 8 called for has a real, public shape now, though it has not been introduced as formal legislation. On June 4, 2026, a bipartisan group of House members — Reps. Jay Obernolte (R-CA), Lori Trahan (D-MA), Suhas Subramanyam (D-VA), Scott Franklin (R-FL), Scott Peters (D-CA), and Erin Houchin (R-IN) — released a discussion draft known as the Great American AI Act. Its preemption mechanism is narrower than the 2025 reconciliation-bill version in one respect and broader in another: it would preempt state laws that specifically regulate how AI models are developed, for three years, while leaving laws that govern how AI is deployed or used alone. Rather than simply wiping state frontier-safety and transparency rules off the books, the draft would federalize versions of some of them — while naming specific California statutes it would preempt outright, including AB 2013 (training-data transparency) and parts of SB 942 (AI content watermarking).
The draft is still a draft: it was released for stakeholder feedback, not filed as a bill, and it has stayed there through the summer. It also drew real, organized opposition rather than passing quietly — over 200 state lawmakers and labor groups, including the Association of Flight Attendants, publicly urged Congress to reject the preemption provision in July 2026. Nothing about its trajectory since suggests that opposition has been resolved either way.
What the states did while Congress deliberated
The preemption fight has not frozen state AI legislating in the meantime — if anything, the clearest sign of where authority currently sits is that states kept amending their own laws on their own timelines. Colorado is the sharpest example: on May 14, 2026, its legislature passed SB 26-189, which repeals and reenacts the Colorado AI Act rather than simply delaying it again. The rewrite drops the original law’s risk-management-program, impact-assessment, and algorithmic-discrimination-prevention duties and replaces them with four narrower obligations — user notification, adverse-outcome disclosure, a data-correction right, and meaningful human review — with a new effective date of January 1, 2027, the same deadline by which the state attorney general must issue implementing rules. Colorado’s own stated rationale for keeping a state law on the books at all, per counsel tracking the bill, was to preserve “the policy goal of filling the AI oversight vacuum given the lack of a comprehensive federal law” — a state legislature explaining, in real time, why it isn’t waiting on Washington.
California’s SB 53 has stayed in force the same way, and has already produced at least its second public compliance dispute of the year: Fortune reported on September 14, 2026, that an AI watchdog group had alleged OpenAI’s latest model releases might not meet SB 53’s frontier-safety-framework and transparency-reporting requirements — echoing a similar dispute over an earlier model release in February 2026 that OpenAI publicly contested. Whatever the merits of the specific allegation, its existence is itself informative: a year into the federal government’s stated effort to preempt state AI-safety disclosure law, the law is active enough to be the subject of a live enforcement argument.
Congress’s own uncertainty about acting at all
The clearest recent signal about where the Great American AI Act’s odds actually stand didn’t come from AI policy staff — it came from House Speaker Mike Johnson, in a string of public remarks in mid-September 2026. Johnson said Congress shouldn’t “panic” into “emergency” AI regulation and argued that AI companies, not the federal government, should be primarily responsible for AI safety. Separately, he rejected calls for an AI “moratorium,” warning it would hand China a competitive advantage.
That second comment is worth pausing on precisely because it’s easy to misread against this piece’s topic: Johnson was responding to calls for a pause on frontier AI development — the kind of proposal behind the Statement on Superintelligence, which CASRAI covers separately — not to the state-preemption fight this piece is about. But his broader position, that Congress shouldn’t be the one leading on AI safety rules, cuts directly against the odds of the Great American AI Act, or anything like it, becoming a statute soon. That matters specifically because a statute is the one thing that could preempt state AI law with real legal certainty. An executive order cannot, by itself — it has to work through litigation, agency rulemaking, or funding conditions, each of which a court can uphold, narrow, or reject. As long as Congress stays where Johnson described it in September, EO 14365’s litigation-and-agency-theory route remains the only federal preemption mechanism actually in motion, and every one of the state laws it targets remains in effect exactly as written until a court, Congress, or the state itself says otherwise.
Why the jurisdiction question is a vocabulary problem, not just a legal one
Underneath the litigation and the draft bill is a plainer governance question: when a frontier AI model is deployed, whose safety-framework rules actually apply to it, and how would a compliance record show that? That’s the specific gap CASRAI’s own NIKOLAI project — an independent, unendorsed reference dictionary of frontier-AI-safety elements, not a standard adopted by any lab, regulator, or evaluator — was built to name. Its N1 track, Actors, models and scope, catalogs eight elements a governance record needs to identify an AI artifact and the rules that govern it, and two of them describe exactly the axis this fight turns on. NIKOLAI’s safety framework version element is defined as the identifier, effective date, and change log of the governing safety framework under which an artifact was produced — recorded so an assessment or safeguard can be tied to the exact framework version in force at the time. Its coverage scope threshold element records which population, deployment surface, or model class a given framework actually reaches.
Neither element resolves the preemption question; NIKOLAI has no position on whether SB 53, Colorado’s SB 26-189, or a future Great American AI Act statute governs a given deployment, and nothing here should be read as any of those bodies having adopted or endorsed NIKOLAI’s vocabulary. What the current fight makes concrete is why a governance record needs fields like these at all: in a landscape where the governing framework for the same model could change by court ruling, executive order, or statute within the same calendar year, “which framework version, in force over which scope, applied to this model on this date” is a question an organization needs to be able to answer precisely — and right now, in the U.S., the honest answer keeps changing.
What to watch next
- Whether the Great American AI Act moves from discussion draft to introduced bill — and whether its three-year, development-only preemption scope survives contact with the opposition it drew in July 2026.
- Colorado’s January 1, 2027 effective date for SB 26-189’s narrower notification, disclosure, correction, and human-review duties, and the attorney general’s implementing rules due the same day.
- How the SB 53 compliance dispute reported September 14, 2026 is resolved — the first real test of whether California’s frontier-safety reporting duties hold up in practice, not just on paper.
- Whether the DOJ AI Litigation Task Force that EO 14365 created actually files and wins a case against a state AI law, which would be the first time any part of the federal preemption effort produced a binding legal result rather than a policy directive.
