No US State's Mandatory AI Impact-Assessment Anti-Discrimination Law Survives in Force Through Jan 1, 2028

The Prediction

No US state's comprehensive, deployer-side AI anti-discrimination law carrying mandatory pre-deployment impact-assessment or risk-management duties takes effect and remains in force on January 1, 2028. The Colorado retreat — where the original SB 24-205 was stayed and effectively replaced by the much narrower SB 26-189 — marks the high-water mark of the impact-assessment model in the United States, not the beginning of a wave.

Confidence Level: 68%

To resolve inaccurate, at least one US state must have, as of January 1, 2028, a statute that is all of the following at once: (1) comprehensive in scope, covering high-risk or consequential automated decision systems across multiple domains such as employment, lending, housing, insurance, or essential services rather than a single sector; (2) deployer-side, placing duties on the organizations that use AI to make consequential decisions, not merely on developers; (3) carrying a mandatory pre-deployment obligation, an impact assessment, algorithmic risk assessment, or risk-management program required before or as a condition of deployment; and (4) actually in effect and enforceable on that date, not merely enacted with a future effective date, and not stayed, repealed, or gutted down to a notice-only regime.

Why This Matters

For two years the Colorado AI Act, SB 24-205, was the template every other state borrowed from. It was the first comprehensive, deployer-facing algorithmic discrimination law in the country, and it leaned hard on the impact-assessment model: deployers of high-risk systems would have to perform and document risk assessments, maintain risk-management programs, and demonstrate reasonable care to avoid algorithmic discrimination. That structure was widely treated as the likely national baseline, the thing other legislatures would copy the way they copied California's privacy law.

Then Colorado blinked. As detailed in the Colorado AI Act retreat, the original effective date was pushed back, the law was stayed, and the legislature replaced the ambitious impact-assessment regime with a far narrower successor that strips out the heavy pre-deployment duties. The first state to adopt the model became the first to abandon it before it ever took hold. That sequence is the single most important data point for anyone forecasting where state AI regulation goes next, and it points away from the impact-assessment model rather than toward it.

The Reasoning

Three forces push in the same direction.

The template that was being copied just collapsed. Legislatures imitate what appears to be working and durable. Colorado's law was that reference design, and its retreat sends an unambiguous signal that the compliance burden was judged unworkable, premature, or economically risky by the very state that pioneered it. States that were drafting Colorado-style bills now have a worked example of what not to ship, and the most natural reaction is to wait, narrow, or shelve.

The federal environment is actively hostile to the model. The prevailing federal posture favors deregulation and AI acceleration, and there has been sustained pressure toward preemption or discouragement of aggressive state AI rules. Even where an outright federal preemption does not pass, the threat of it chills state action: legislators are reluctant to spend political capital on a comprehensive regime that Washington might override or that could brand their state as anti-business in an AI investment race.

The compliance model itself is genuinely hard. Mandatory pre-deployment impact assessments across all consequential decision systems impose real cost and legal uncertainty on every bank, insurer, employer, and landlord operating in the state. Industry opposition is well-funded and effective, and the assessment requirement is precisely the provision lobbyists target first, because it is the expensive one. The Colorado successor shows the typical compromise: keep some transparency or notice language, drop the binding pre-deployment duty.

By contrast, the alternatives that are advancing tend to be narrower: sector-specific rules (employment automated-decision tools, insurance underwriting), developer-side transparency obligations, or notice-and-disclosure regimes. Those can pass and survive precisely because they do not carry the comprehensive, mandatory, deployer-side assessment duty that defines the model this prediction is about.

What Would Falsify It

This prediction is falsified if any of the following happens before January 1, 2028:

• A state with an already-enacted comprehensive law — Colorado being the obvious candidate if its delay reverses, or another state that quietly enacted one — has that law in force on the target date with the mandatory pre-deployment assessment intact.
• A large state with both the institutional capacity and the political will (California, Illinois, New York, New Jersey, Connecticut, or Texas) enacts a Colorado-style comprehensive bill with a near-term effective date that lands before 2028 and survives industry challenge without being narrowed below the threshold.
• A coalition of states adopts a model act — for instance through a multistate drafting effort — that includes a binding impact-assessment duty and at least one member state brings it into force in time.

The prediction does not require that no AI legislation passes. Plenty will. It requires specifically that the comprehensive, deployer-side, mandatory-assessment model fails to take and hold in any single state by the date. A flurry of narrower bills, even strong ones, leaves the prediction intact.

Signposts to Watch

• Colorado's revised effective date. Whether the narrowed SB 26-189 holds, or whether reformers restore assessment duties in the next session, is the leading indicator. A genuine restoration would be the fastest path to falsification.
• California's trajectory. California's automated-decision and frontier-model rulemaking is the most likely place a comprehensive deployer duty could re-emerge at scale. Watch whether its rules cross from disclosure into mandatory pre-deployment assessment with real teeth.
• Federal preemption activity. Any serious federal moratorium or preemption proposal raises confidence in this prediction by freezing state action; its failure lowers confidence by clearing the lane for states.
• Industry posture. If major employers and insurers stop fighting the assessment model and start treating it as inevitable compliance overhead, that would be an early sign the political economy is shifting against this prediction.

Confidence Calibration

Sixty-eight percent reflects genuine uncertainty. The case for the prediction is strong: the flagship law retreated, the federal wind is at the deregulators' backs, and the assessment duty is the most lobbied-against provision in the whole model. The case against it is that the United States has fifty laboratories, the target window is eighteen-plus months, and it only takes one determined state with a near-term effective date to falsify the claim. California in particular has both the appetite and the machinery to surprise on the regulatory side. I am deliberately not above 70 percent, because forecasting that zero of fifty states does a specific thing over eighteen months should carry humility, even when the trend line points the right way.