Colorado scales back AI rules
- Colorado lawmakers moved to replace the state’s 2024 AI law with SB 26-189, a narrower disclosure-and-appeals regime that passed the House 57-6 on May 9. - The rewrite drops the old “reasonable care” anti-discrimination duties and shifts the start date to January 1, 2027, with 60 days to cure violations. - Colorado is backing away from the country’s toughest state AI law just weeks before its June 30, 2026 launch.
Colorado’s AI fight just turned into something much smaller. The state passed a landmark law in 2024 that tried to police discriminatory AI systems before they harmed people. But days before that law was supposed to bite, lawmakers moved to replace it with a narrower bill built around disclosure, consumer access, and post-decision appeals. Basically, Colorado is not abandoning AI regulation. It is swapping a broad duty-to-prevent model for a lighter tell-people-and-let-them-challenge-it model. ### What changed this week? The vehicle is Senate Bill 26-189. It repeals and reenacts the core of SB 24-205, the 2024 law that made Colorado the first state to impose broad obligations on both AI developers and deployers. The new bill cleared the Colorado House on May 9 by a 57-6 vote after moving through committee earlier in the week, so the rewrite is no longer theoretical — it is the live legislative path. (leg.colorado.gov) ### What was the old law trying to do? The 2024 law targeted “high-risk” AI systems used in consequential areas like jobs, housing, lending, insurance, health care, and government services. It told developers and deployers to use “reasonable care” to protect consumers from “algorithmic discrimination,” do impact assessments, keep risk-management programs, and give consumers appeal rights when AI played a substantial role in an adverse decision. (leg.colorado.gov) That was a pretty muscular framework by U.S. standards. ### So what does the new bill do instead? SB 26-189 changes both the vocabulary and the burden. It replaces “high-risk AI systems” with “automated decision-making technology,” or ADMT, and focuses on systems that materially influence consequential decisions. Developers still have to give deployers technical documentation — intended uses, training-data categories, limits, and human-review instructions. But the center of gravity moves to notice, recordkeeping, and consumer rights after an adverse outcome. (leg.colorado.gov) ### What got scaled back? The biggest cut is the affirmative duty. The old law said companies had to use reasonable care to prevent foreseeable algorithmic discrimination. The new bill drops that structure and instead leans on disclosure, access to data, correction rights, and meaningful human review after a bad result. It also appears to narrow what companies must explain up front, replacing the earlier detailed disclosure approach with simpler notice plus more detail on request. (leg.colorado.gov) ### When would the new rules start? Not immediately. The old law had already been delayed once to June 30, 2026. The rewrite would push the new framework to January 1, 2027. The bill also gives the attorney general until January 1, 2027 to clarify post-adverse-outcome disclosure rules, and it gives targets of enforcement a 60-day chance to cure violations when a cure is possible. (leg.colorado.gov) ### Why did lawmakers retreat? Because the original law ran into a wall of opposition. Business groups and tech companies argued that the old definitions were vague, the compliance burden was too high, and the law risked chilling ordinary software tools along with more serious AI systems. The compromise that emerged tracks a policy-group draft backed by Gov. Jared Polis’s process and was explicitly framed as a narrower replacement, not just a tweak. (leg.colorado.gov) ### What still matters for consumers? People would still get some real rights. If a covered system helps drive an adverse decision, consumers can ask for a plain-language description of the system’s role, request the personal data involved, correct factual errors, and seek meaningful human review and reconsideration. That is weaker than a duty to prevent harm up front, but it is not nothing. Think seat belts instead of speed limits — more recourse after risk, less constraint before it. (coloradonewsline.com) ### Bottom line? Colorado is still trying to regulate AI. But the state is no longer trying to be the toughest cop in the country. The new bet is that narrower rules have a better chance of surviving politics, surviving implementation, and actually taking effect. (leg.colorado.gov)