Colorado AI Act

On this page 7

The Colorado AI Act is the Colorado statute, enacted as Senate Bill 24-205 in May 2024, that imposes a duty of reasonable care on developers and deployers of high-risk artificial intelligence systems to protect consumers from algorithmic discrimination. In AI regulation, it is widely described as the first comprehensive state AI law in the United States, because it governs AI used across many sectors rather than a single use case such as hiring or facial recognition.

The law borrows ideas familiar from the EU AI Act, including risk tiers, provider and deployer roles and impact assessments, but it is narrower in scope and rooted in consumer protection and anti-discrimination law. Its original effective date of February 1, 2026 was postponed to June 30, 2026 by a 2025 special session, and further amendments have been debated, so practitioners should confirm the current text before relying on any detail.

What the Colorado AI Act regulates

The Act regulates high-risk AI systems, defined as systems that make, or are a substantial factor in making, a consequential decision about a consumer. A consumer is any Colorado resident, so the law reaches out-of-state companies whose systems affect people in Colorado.

A consequential decision is one with a material legal or similarly significant effect on the provision, denial, cost or terms of an important opportunity or service. The statute lists eight domains, shown in the table below with typical examples of covered systems.

DomainTypical covered system
EducationAdmissions scoring or scholarship ranking
EmploymentResume screening or candidate ranking
Financial or lending servicesCredit underwriting models
Essential government servicesBenefit eligibility triage
Health care servicesPrior authorization or care allocation tools
HousingTenant screening scores
InsurancePricing or claims decision models
Legal servicesAutomated case intake prioritization
Domains of consequential decisions under the Act

Algorithmic discrimination means any condition in which the use of an AI system results in unlawful differential treatment or impact that disfavors individuals or groups on the basis of protected characteristics such as age, color, disability, ethnicity, national origin, race, religion, sex or veteran status. The definition captures disparate impact, not only intentional bias, which is why testing and monitoring sit at the center of compliance.

Several categories are excluded from the high-risk definition, including systems that perform narrow procedural tasks, detect decision-making patterns without replacing human assessment, and common technologies such as spam filters, firewalls, calculators and spreadsheets, provided they do not make consequential decisions.

Obligations of developers and deployers

The Act splits duties between the developer, who builds or intentionally and substantially modifies an AI system, and the deployer, who uses a high-risk system in Colorado. This mirrors the AI provider and AI deployer roles in European law, although the specific obligations differ.

ObligationDeveloperDeployer
Duty of reasonable careYes, against known or foreseeable discrimination risksYes, in use of the system
DocumentationProvide data summaries, limitations and evaluation resultsRetain records of assessments
Risk managementNot required as a formal programRisk management policy and program
Impact assessmentSupply information that enables itAnnually and after substantial modification
Public statementSummary of high-risk systems offeredSummary of systems deployed
Attorney General noticeKnown discrimination risks within 90 daysDiscovered discrimination within 90 days
Core duties by role

Deployers carry the consumer-facing obligations. Before a consequential decision, they must notify the consumer that a high-risk system is used. After an adverse decision, they must explain the principal reasons, describe the data used, offer a chance to correct inaccurate personal data and provide an appeal with human review where technically feasible.

A separate general disclosure duty applies to any AI system intended to interact with consumers, high-risk or not: the business must tell people they are interacting with AI unless that would be obvious to a reasonable person. This resembles the AI transparency obligations for chatbots in the EU.

Impact assessments and risk management

The impact assessment is the deployer’s central evidence of reasonable care. It must be completed before deployment, repeated at least annually and refreshed within 90 days of any intentional and substantial modification of the system.

ElementWhat it must describe
Purpose and contextIntended uses, deployment setting and benefits
Discrimination risk analysisKnown or foreseeable risks and mitigation steps
DataCategories of input data and outputs produced
CustomizationData used to tailor the system, if any
Performance metricsMetrics used and known limitations
Transparency measuresHow consumers are informed
Post-deployment monitoringOversight, user safeguards and monitoring process
Required content of a deployer impact assessment

The risk management program must be reasonable in light of the deployer’s size, the nature of the system and the sensitivity of data. The statute points to the NIST AI RMF and ISO/IEC 42001 as recognized frameworks, and alignment with one of them supports both a rebuttable presumption of reasonable care and an affirmative defense.

Enforcement, defenses and exemptions

Enforcement is exclusive to the Colorado Attorney General. Violations are treated as unfair or deceptive trade practices under the Colorado Consumer Protection Act, which allows civil penalties assessed per violation. There is no private right of action, so consumers cannot sue directly under the Act.

  • Rebuttable presumption: a developer or deployer that complies with the statutory duties is presumed to have used reasonable care.
  • Affirmative defense: available when the business discovers and cures a violation through feedback, adversarial testing or internal review and follows a recognized framework.
  • Small deployer exemption: deployers with fewer than 50 full-time employees that do not train the system with their own data are relieved of several duties, such as the impact assessment.
  • Rulemaking authority: the Attorney General may adopt rules on documentation, notices, impact assessments and the presumption.

Businesses already regulated under equivalent federal standards, such as certain banks and insurers subject to comparable oversight, may qualify for sector-specific carve-outs. The scope of these exemptions is one of the most contested points in pending amendment proposals.

How it compares with the EU AI Act

The Colorado AI Act is narrower and more consumer-focused than the EU AI Act. It targets one harm, algorithmic discrimination, whereas the European regulation addresses safety, fundamental rights, prohibited practices and GPAI models.

FeatureColorado AI ActEU AI Act
Primary harm addressedAlgorithmic discriminationSafety and fundamental rights
Risk tiersHigh-risk onlyProhibited, high-risk, limited, minimal
Foundation modelsNot specifically regulatedGPAI model obligations
EnforcerState Attorney GeneralNational authorities and EU AI Office
Conformity assessmentNoneRequired for high-risk systems
Maximum penaltiesConsumer protection penaltiesUp to 7% of global turnover
Colorado AI Act and EU AI Act compared

A common misconception is that compliance with one regime satisfies the other. Overlap is partial: an EU fundamental rights impact assessment covers much of Colorado’s content, but notice, appeal and Attorney General reporting duties must be mapped separately.

In plain words

The Colorado AI Act is a rule for companies that let software help decide important things about people. Examples include who gets a job, a loan, an apartment or insurance. These companies must take reasonable steps to prevent unfair treatment based on race, sex, age, disability or similar traits.

Think of it like a building inspection for decision software. The company checks the system before use and again every year. It tells people when the software is involved. It also lets them fix wrong information and ask a human to look again. If something goes wrong, the state Attorney General can take action. Individual citizens cannot sue under this law.

A real-world example

A regional property management company with about 300 employees uses a third-party tenant screening model to rank rental applicants in Denver. Because housing is a consequential domain, the company is a deployer of a high-risk AI system. Its compliance team obtains the developer’s documentation, which includes training data summaries and known limitations, and builds a risk management program aligned with the NIST AI RMF.

The first impact assessment reveals that applicants who rely on housing vouchers receive systematically lower scores, a pattern that could disfavor protected groups. The company removes the variable that drives the gap, retests the model over a quarter of applications and documents the result. It adds a pre-decision notice to the application form and an appeal path with human review for every denial.

When the developer later releases a substantially modified version, the company completes a new assessment within 90 days. The outcome is a defensible compliance record that supports the presumption of reasonable care if the Attorney General ever investigates.