Colorado’s Chatbot Safety Act: What It Requires and Who It Covers

Starting January 1, 2027, a business that offers a general-purpose AI chatbot or a companion app to Colorado residents has new obligations under HB 26-1263, the Chatbot Safety Act, codified at C.R.S. § 6-1-1710. It has to tell users they are talking to AI, run a protocol for users at risk of self-harm, protect minors from sexual content and emotional dependence, and report to the Attorney General every year.

This is a separate law from SB 26-189. The ADMT law governs decisions software makes about people. The Chatbot Safety Act governs conversations software has with them. Most customer service and single-purpose bots are exempt. The businesses most likely to be caught off guard are the ones that added a branded assistant willing to talk about anything.

Who it covers

The Act applies to operators of a conversational artificial intelligence service: an AI system available to the general public that primarily simulates human conversation through text, images, or voice. There are two ways to be an operator:

• You build it and make it publicly available. The Attorney General’s draft rules count licensing it to another business as making it available.
• You offer it to consumers, even if someone else built it.

Both can apply to the same chatbot. A Colorado business that puts a licensed assistant in front of customers is an operator, and in most cases so is the vendor that built it. Neither the statute nor the draft rules divide the duties between them, so the vendor contract has to.

The protections run to Colorado residents. There is no size threshold, and the Attorney General has said the law applies to operators regardless of where they are located. Companion apps are the core target, and an app store download does not change that. The app store is not the operator; the developer is.

Who is exempt

The statute lists twelve exemptions. In practice they fall into four groups:

• Commerce and customer support. Shopping, ordering, payments, returns, and service bots.
• Narrow tasks. Order status, billing, scheduling, and menu-driven bots, as long as they cannot produce sexual content or carry on a conversation about self-harm.
• Business and internal tools. Developer and research tools, business-to-business tools, internal-only use, and device voice assistants.
• Specific settings. Video games, theme parks, HIPAA covered entities and their business associates, school tools, and features embedded in other software.

One caveat does most of the work. The school and embedded feature exemptions do not apply to anything designed to simulate emotional companionship or encourage emotional dependence. Draft Rule 9 lists what the Attorney General will look for: names, avatars, and backstories; responses built on constant validation; and memory used to create a sense of an ongoing relationship.

What it requires

The statute sets the duties. The draft rules, 4 CCR 904-6, say how the Attorney General will judge compliance. What follows reflects the interim draft released October 6, 2026; final rules are due by January 1.

• Disclose that it is AI. To every user at the start of each day’s first conversation, at least every three hours or through a persistent notice, and whenever a user asks. Minors get a persistent disclaimer that stays on screen without scrolling and stands out from the surrounding content.
• Run a self-harm protocol. A response that refers the user to a crisis line or similar service, never to law enforcement, with escalation for repeated or severe warning signs.
• Do not imply professional credentials. Advertising, the interface, and outputs cannot suggest the service comes from a licensed doctor, lawyer, mental health professional, or dietitian. Under draft Rule 13, that includes controls on bot profiles named “therapist” or “doctor.”
• Estimate users’ ages. By commercially reasonable or generally accepted methods, without ignoring clear signs that a user is a minor. Draft Rule 10 says government ID cannot be the only method, accuracy has to be tested by age band, and device and app store age signals count. So does what a user says in conversation, such as repeated references to high school. The ISO/IEC 27566-1 age assurance standard is the benchmark.
• Protect minors. No streaks, badges, or other unpredictable rewards meant to drive engagement. Technically feasible measures against sexual content, and reasonable measures against responses that simulate emotional dependence or isolation from family and friends. Draft Rule 12 treats a safeguard that depends only on the model following instructions as weak evidence, and expects testing before and after launch, with records.
• Give minors and parents control. Tools to manage privacy settings, including whether the service remembers past conversations and whether the minor’s data is used for training.
• Report every year. Beginning July 1, 2027, operators report crisis referrals, their self-harm protocols, and the metrics the Attorney General requires. Draft Rule 14 adds accuracy measures for referrals, the age distribution of users, and counts of reports involving minors. A report prepared for another state can be reused with an appendix.

The October 6 draft eased several August proposals, including how large the minors’ disclaimer must be and a requirement that minors’ settings default to the most protective option. The rulemaking is covered in more detail in Colorado’s AI law after the October 26 hearing.

How it is enforced

The Attorney General enforces the Act through the Colorado Consumer Protection Act, and a violation is a deceptive trade practice carrying civil penalties of up to $20,000 per violation. The statute creates no new private right of action. The 60-day opportunity to cure in SB 26-189 is written for developers and deployers under the ADMT law, so I would not count on it for a chatbot operator.

Where it stands

• The statute took effect August 12, 2026. The operator duties begin January 1, 2027, and the first annual report is due July 1, 2027.
• The rules are not final. The hearing was October 26, and the statute requires adoption by January 1.
• The xAI court order may or may not reach this law. There is a credible argument either way, which I walk through in the October 30 update. Until a court decides, I would treat the duties as live.

Two examples

A Longmont credit union runs a chat assistant that answers balance, payment, and loan questions. That is customer service, and it is exempt. If the credit union adds a “financial wellness coach” built on a licensed model that will talk about anything a member raises, it is likely covered, and its vendor probably is too.

A Colorado developer publishes a companion app on the Apple and Google app stores, with named characters that remember past conversations. That is squarely covered. Because some of its users will be minors, every minor protection applies, along with the annual report.

What to do now

• Sort every chatbot you run against the exemptions, and write down why each exempt one qualifies.
• For a covered chatbot, choose an age estimation method, test your safeguards and keep the results, and write the crisis protocol.
• Settle responsibility with your vendor for disclosures, protocols, and the annual report.
• Start collecting report data on January 1.

If you are not sure which side of the line your chatbot falls on, that is usually a short conversation, and I would welcome it. You can also see how I approach AI governance and compliance work.