The US AI Rulebook Is Being Written in State Capitals

In the absence of a comprehensive federal AI statute, US AI policy is increasingly being set state by state, and that fact alone reorders how the rulebook gets written. When Washington does not occupy the field, the first movers are state legislatures and governors, whose statutes become the de facto national reference points simply because nothing larger exists to override them. Factrail's model tracks two living architects of that shift: California state Senator Scott Wiener and Colorado Governor Jared Polis, whose laws now anchor the Digital-Rights and Platform-Power Regulation driver in the US context. Reading their work as anchors rather than experiments is the model's central interpretive move: in a vacuum, the earliest binding rules set the gravity that later proposals must answer to.

Colorado moves first on the substance

Colorado moved first on the substance of automated decisions, and the substance is what makes the move consequential. The fact Colorado Enacts the First US Comprehensive AI Anti-Discrimination Law records Polis signing SB 24-205 in May 2024, imposing notice, disclosure and impact-assessment duties on "high-risk" AI used in hiring, lending and housing. These three domains are not arbitrary; they are precisely the gatekeeping decisions where an opaque automated system can quietly reproduce or amplify discrimination at scale, which is why a law reaching them counts as regulating outcomes people feel rather than abstractions.

Notably, Polis signed it "with reservations," warning that a state-by-state patchwork could burden developers, and the original act was repeatedly delayed and later rewritten. The model treats that detail as load-bearing rather than incidental. A governor who signs his own state's law while publicly doubting the wisdom of fifty separate regimes is, in effect, registering the central tension of the whole state-led approach from inside it. The subsequent delays and rewrite are recorded as evidence that a first-mover statute is a moving target, not a finished standard.

California's more contested path

California's path was more contested, and the contest is instructive. The fact Newsom Vetoes California's SB 1047 Frontier-AI Safety Bill shows Governor Gavin Newsom rejecting Wiener's first, broader safety bill in September 2024, arguing its thresholds were not empirically grounded. The veto is analytically interesting because it was not a rejection of regulation as such but a dispute over calibration: the objection was to where the lines were drawn, not to the idea of drawing them, which left the door open to a narrower successor rather than slamming it shut.

A year later the fact California Enacts SB 53, the First US Frontier-AI Transparency Law records that narrower successor becoming law, focused on public safety frameworks, incident reporting and whistleblower protection rather than prescriptive testing mandates. The shift from SB 1047 to SB 53 is a study in how state-level lawmaking iterates under constraint: a broad mandate that drew a veto was reworked into a transparency-and-disclosure regime that could survive one. That progression from rejection to enacted compromise is exactly the kind of sequence the model is built to follow.

Whether the state-level approach converges into a coherent national standard, or fragments into conflicting regimes, remains an open question the model flags rather than resolves.

How the model weighs it

Factrail treats these as net-strengthening moves for protective digital-rights regulation, but with honest caveats that keep the assessment from overreaching. The veto, for instance, is recorded as a temporary pause rather than a harm. That distinction matters: a bill that fails on calibration and returns in amended form a year later is a setback in a process, not a defeat of the underlying direction, and conflating the two would distort the trajectory the model is trying to trace. Treating SB 1047's veto as a harm would have mismeasured a moment that ultimately produced an enacted law.

The contested patchwork concern that Polis raised is real, and the model carries it as a genuine uncertainty rather than dismissing it. Fifty state regimes can converge toward a shared template, which would make the early laws genuine national anchors, or they can diverge into conflicting compliance burdens that a federal statute would eventually have to reconcile. The model does not pretend to know which outcome will hold. It records the direction of travel, marks the open question, and leaves the resolution to events.

Why state capitals matter here

The deeper significance is institutional. With no federal frontier-AI law on the books, the rulebook is being drafted in Sacramento and Denver, and the choices made there about disclosure, impact assessment and whistleblower protection are setting the terms other states will adopt, adapt or resist. That is why Factrail tracks individual governors and legislators as architects of a driver rather than as bystanders to a technology story. The state-led model is not a placeholder for federal action so much as a live experiment in whether a coherent national standard can emerge from the bottom up. For the contrasting top-down approach, the related analysis of how the EU writes its rules for the machine follows the same questions through a single supranational statute rather than a patchwork of state capitals.