Colorado’s AI Act: the Duty of Care Standard, Explained

Colorado’s AI Act does not tell you to test your model in a particular way. It imposes a duty of reasonable care, which is a standard rather than a rule, and standards are met by evidence of judgement rather than by ticking a box. That difference is the whole subject of this page.

The standard, phrase by phrase

Senate Bill 24-205, signed on 17 May 2024 and codified in Part 17 of Article 1 of Title 6 of the Colorado Revised Statutes (sections 6-1-1701 onwards), imposes essentially the same core duty twice: a developer of a high-risk artificial intelligence system, and a deployer of one, must each use reasonable care to protect consumers from any known or reasonably foreseeable risks of algorithmic discrimination. The bill and its history are on the Colorado General Assembly’s own bill page.

Four phrases carry the weight.

• “Reasonable care” is a comparative standard, not an absolute one. It asks what a reasonable developer or deployer in your position would have done, which means it moves as industry practice moves. A technique that was exotic in 2024 and routine in 2027 becomes part of what reasonable care requires without any amendment to the statute.
• “Known or reasonably foreseeable” is the phrase that decides how much investigation you owe. It covers risks you actually knew about — from testing, from complaints, from a developer’s disclosure — and risks you would have known about had you looked in the way a reasonable party looks. Not looking is not a defence; it is the classic way to fail a foreseeability test.
• “Protect consumers from” is a duty of mitigation, not a guarantee of outcome. The statute does not say a system must produce no disparate outcome. It says you must take reasonable steps against the risk of one.
• “Algorithmic discrimination” is defined in the statute, and the definition is narrower than the ordinary meaning — it is tied to unlawful differential treatment or impact on protected characteristics. That definition has its own page because the word “unlawful” inside it does more work than anything else in the Act.

Why it is shaped like negligence

Reasonable care is the language of tort. Colorado did not, however, create a negligence action: the Act is enforced as a deceptive trade practice under the Colorado Consumer Protection Act. What it borrows is the structure of a negligence standard — a duty defined by reasonableness, breach assessed against what a reasonable party would do, and foreseeability limiting the scope of what you must guard against.

That choice has a practical consequence which is easy to miss. A rules-based statute is satisfied by compliance: do the listed things and you are done. A standards-based statute is satisfied by justification: you must be able to explain why what you did was reasonable given what you knew. Contemporaneous documentation is therefore not administrative overhead under this Act; it is the substance of the defence. A mitigation you performed and did not record is very hard to prove two years later, and the reasonableness of a decision is judged on what you knew at the time, which only a contemporaneous record can establish.

The presumption shifts the work, not the duty

Both the developer section and the deployer section pair the duty with a rebuttable presumption: a party that does the specific things the section lists is presumed to have used reasonable care. That is a burden-allocation device, not a safe harbour. The presumption can be rebutted, and doing the listed things badly is unlikely to sustain it.

The deployer’s route to the presumption runs through a risk management policy and programme, which the statute says must be reasonable in view of guidance and standards including the NIST AI Risk Management Framework, ISO/IEC 42001, or another nationally or internationally recognised framework, or one designated by the Attorney General. That is an unusual piece of drafting: it makes a voluntary standard into the reference point for a statutory duty without mandating certification to it. See the NIST framework’s four functions and what ISO/IEC 42001 certification actually covers, and the presumption page for how it is pleaded.

Separately, the Act provides an affirmative defence where a party discovers and cures a violation through its own internal testing — including adversarial testing or red-teaming — or through feedback it acts on, and is otherwise in compliance with a recognised risk management framework. An affirmative defence is not the same thing as a presumption: it operates after a violation is established, and the party asserting it carries the burden. The two together create a deliberate incentive to look for problems, because finding one is better protected than not having looked.

Who can sue on it

The Act gives exclusive enforcement authority to the Colorado Attorney General and creates no private right of action. That is a significant limitation on the duty of care in practice: an individual who believes a hiring model discriminated against them cannot bring a claim under this Act. They may still have claims under Title VII of the Civil Rights Act of 1964, the Colorado Anti-Discrimination Act, or the Fair Credit Reporting Act, none of which the AI Act displaces. The Act also authorises the Attorney General to promulgate rules, so parts of the standard may be made more concrete by rulemaking rather than by amendment. This is covered further in the enforcement page.

What meeting it looks like in practice

Because the standard is about justified judgement, the artefacts that evidence it are the ones that record judgement being exercised over time. In rough order of how often they are the thing that is missing: a written record of what risks of algorithmic discrimination were identified for this specific use, and by whom; the testing that was done, including what was measured and against which subgroups, with the results kept whether or not they were good; the mitigations chosen and, importantly, the ones considered and rejected with the reason; the post-deployment monitoring arrangement and what it has actually surfaced; and the review cadence, since a risk that was not foreseeable at launch may be known by the second quarter.

Note what is not on that list. Certification to a standard is not required by the Act. A vendor’s assurance that its model is unbiased is not a mitigation you performed; at best it is an input to your own assessment, and a deployer that relies on it without more has a thin story about reasonable care. The disclosures a developer must hand a deployer are designed to be that input, which is why the division between the two roles matters so much — see the developer and deployer duties side by side.

Finally, the duty is continuing. It is not discharged at procurement and it does not end at launch. The Act pairs it with an obligation to notify the Attorney General within ninety days of discovering that a system has caused or is reasonably likely to have caused algorithmic discrimination, which is the point at which an internal finding becomes an external one.