The Colorado Department of Law on November 19, 2026 adopted 76 pages of rules that will govern two new state laws at once: the Automated Decision-Making Technology in Consequential Decisions Act, enacted as SB 26-189, and the Conversational Artificial Intelligence Services Act, enacted as HB 26-1263 and called the "Chatbot Safety Act" in the rules. The text, numbered 4 CCR 904-6 and titled "Automated Decision-Making Technology & Conversational Artificial Intelligence Service Rules," sets out how companies that use algorithms in lending, hiring, housing and education must explain adverse decisions, and how operators of consumer chatbots must estimate users' ages, remind them they are talking to a machine, and report on suicide and self-harm safeguards. It is not yet in force. The Department has asked for a formal opinion from the Attorney General, after which the rules go to the Secretary of State and become final twenty days after publication.
In Short
Colorado has adopted detailed rules for two AI laws: one covering automated tools that help decide who gets a loan, job, apartment or school place, and one covering consumer chatbots. If a company uses such a tool and turns someone down, it will have 30 days to say so in plain language, and chatbot operators will have to check users' ages without relying only on government ID, remind them they are talking to AI, and file safety metrics with the state by July 1, 2027. The rules still need an Attorney General opinion and publication before they are final, and a federal court order in xAI's lawsuit already limits when the state can start enforcing the automated-decision law.
Where the rules stand
According to the Department's rulemaking page, informal comments were accepted through July 13, 2026, and the proposed rules were filed with the Secretary of State on August 11 under tracking number 2026-00360. An interim draft followed on October 6, a regulatory analysis on October 21, and a formal hearing on October 26 at 10:00 a.m., the same day written comments closed at 11:59 p.m. Mountain time. The Secretary of State's docket lists November 19 as the adoption date and November 20 as the date the Attorney General's opinion was requested. Its status shows the filing as pending.
The Department posted the adopted text as a double redline dated November 18. The rules themselves carry no date; their effective date clause points to whatever date the Secretary of State publishes in the Colorado Code of Regulations.
That clause is one of the visible changes from the earlier drafts. Both the August 11 proposal and the October 6 interim draft set a fixed effective date of January 1, 2027. The adopted Rule 1.3 instead reads: "These rules shall become effective by the effective date published by the Secretary of State in the Colorado Code of Regulations." The statutes themselves still take effect on January 1, 2027, and SB 26-189 requires the Attorney General to adopt rules on post-adverse outcome disclosures before that date.
A second change sits in the definitions. The earlier drafts described a continuous chatbot interaction as a sequence of exchanges that persists regardless of temporary pauses, without saying how long a pause could last. The adopted definition puts a number on it: an interaction begins when a user starts a conversation, thread, project or chatbot profile, and "Ends when sixty (60) consecutive minutes pass without the User interacting with the Conversational Artificial Intelligence Service." A pause of less than 60 minutes does not end it, and a continuous interaction can run across several threads or chatbot profiles. That definition drives every disclosure timer in the chatbot half of the rules.
The two statutes came out of a long fight. SB 24-205, the original Colorado AI Act signed on May 17, 2024, was repealed and replaced by SB 26-189, which the Colorado House passed 57-6 on May 9, 2026. According to the General Assembly's record, Governor Jared Polis signed SB 26-189 on May 14, 2026 and HB 26-1263 on May 29, 2026.
The deadlines in one place
The rules are dense with time limits. The table below collects them, with each figure taken from the text of 4 CCR 904-6.
| Obligation | Who | Time limit or threshold | Rule |
|---|---|---|---|
| Adverse outcome disclosure after a consequential decision | Deployer | Within 30 days after the decision | 7.4(A), 7.7 |
| Response to a request for more information about the AI system | Deployer | Within 15 days | 7.6(A)(1) |
| Instructions for accessing or correcting data, when requested by email, mail or phone | Deployer | Within 15 days | 8.2(E) |
| List of third parties holding the data, when the deployer holds none | Deployer | Within 15 days | 8.2(H) |
| Access to or correction of personal data | Deployer | Within 45 days, extendable once by 45 days | 8.3, 8.4, 8.6 |
| Acknowledging a request for meaningful human review | Deployer | Within 10 days | 8.7 (Timing) |
| Completing meaningful human review | Deployer | Within 45 days, one extension of up to 45 days | 8.7 (Timing) |
| Extension of human review | Deployer | Not available where the decision affects food, shelter or required medical care | 8.7 (Timing) |
| Records of human review | Deployer | Three-year retention | 8.7 (Documentation) |
| Printed disclosures sent by post | Deployer | Not less than 12-point font | 7.2(A)(2) |
| End of a continuous chatbot interaction | Operator | 60 consecutive minutes without user activity | 2.2 |
| AI disclosure to minors | Operator | At the start of each interaction and at least once every three hours | 11.2(A)(3) |
| AI disclosure to adults | Operator | First interaction of each calendar day, then persistent or at three-hour intervals | 11.2(B) |
| First annual report to the Attorney General | Operator | On or before July 1, 2027, covering the preceding calendar year | 14.5(A) |
| Producing documents to verify the annual report | Operator | Within 30 days of a Department request | 14.3(C) |
Automated decisions: when an algorithm counts
The first eight rules deal with automated decision-making technology, abbreviated throughout as ADMT, in what the statute calls consequential decisions: those affecting access to education, employment, housing, financial or lending services and other covered domains. The central question is when a tool's output "Materially Influences" a decision, because only then do the automated decision-making duties attach.
Rule 6.2 answers with a presumption. An output is presumed to materially influence a consequential decision if it "constrains an option set, sets a threshold, or produces a rank, score, classification, recommendation, prediction, or other inference that pertains to the individual about whom the decision is being made," and it is reviewed by the deployer during the decision, used to screen the data the deployer sees, or used to decide without human involvement. "Other inference" is defined as "any opinion, characterization, or non-objective conclusion reached by the ADMT."
The presumption can be rebutted by showing the output was a "De Minimis Factor." Evidence can include that the deployer never consulted the output; that the decision rested on a hard cut-off that was documented beforehand, is not derived from the ADMT and is applied consistently; or that a person conducted an independent review with enough time, authority and subject matter understanding. Rule 6.2 also states that "the absence or presence of any one of the examples above, standing alone, does not preclude a finding that the ADMT output was a de minimis factor."
Some tools fall outside the definition entirely. Under the statute, tools used "solely to summarize, organize, translate, draft, route, or present information for human review of administrative processing" are not ADMT. An AI system that sorts applications from highest to lowest grade point average "without any subjective input" is excluded because it only organizes data. A tenant-placement tool that sorts applicants by unit size and date is excluded too, but becomes covered ADMT once it labels applicant data "positive" or "negative," because, according to the rule, it "opined on whether qualities of the tenant applications were 'positive' or 'negative' rather than objectively summarizing each applicant's data."
Other examples are pointed. A private college that uses an AI system to "predict how that student would respond to different amounts of institutional aid" and sets aid accordingly is making a consequential decision, because the aid amount is "a differentiated price that is reasonably likely to materially limit or fundamentally alter a consumer's access to education." In hiring, an algorithm that ranks candidates from "strongest" to "weakest" without showing a numeric score still triggers the presumption, because it "produced a non-objective conclusion about the applicants."
Developers and supply chains
Rules 4 and 5 deal with the companies that build these systems. A developer that integrates another developer's covered ADMT into its own product must either pass the upstream documentation to downstream customers or fold it into its own disclosures. Downstream parties "may reasonably rely on information made available by preceding Developers absent actual knowledge that the information is materially inaccurate or incomplete." A company whose general AI product is integrated into covered ADMT "without the party's knowledge" and without having designed, marketed or contracted for that use is not a developer.
Developer documentation must name the decisions for which a system's use "is known to be inappropriate or harmful" and say whether training data includes sensitive or biometric data. Trade secrets may be withheld, but the developer must cite its legal authority and still provide "a summary, categorical description, or other generalized information."
When a third party runs the system on a deployer's behalf, the deployer remains responsible. In the rules' example, an employer that has a staffing agency shortlist applicants is the deployer, though it may contract with the agency for help.
What a denial letter must now say
Rule 7 turns the statute's 30-day disclosure duty into a format: a plain-language description of the decision and the system's role, instructions for requesting more information, and an explanation of the rights to access and correct data and to ask for human review.
The rules supply sample sentences: "Bank ABC closed your checking account," "Leasing Company ABC was unable to approve your application to rent 123 Main Street," and "Credit Union ABC has denied your loan application." Vagueness is not allowed. According to Rule 7.4, "a statement that the role of the Covered ADMT was to automate a review of the Consumer's Personal Data is insufficient," while a statement that the system reviewed "credit report information and employment verification information" is specific enough. Where a factor was an automatic denial, such as criminal history, credit score or years of work experience, the disclosure must say so. Where the decision followed incomplete information, it must say that too.
The sector examples come with figures. A college awarding a 50% tuition scholarship might write: "You are only eligible for a 50% tuition scholarship, covering $16,850 in your first year." A landlord whose system sets a $2,500 deposit, where other tenants of the same three-bedroom unit paid between $1,500 and $2,500, must explain that the automated system "was the sole basis used in making a decision regarding the amount of your security deposit." A closing-cost grant program might tell an applicant they were "automatically denied because your income is above the program limit of $75,000." An employer that uses ADMT to pick 10 interview candidates from 87 applicants and sends federal Fair Credit Reporting Act adverse action notices to the other 77 needs no separate notice, provided the existing notice carries the information on further requests and consumer rights.
Under Rule 7.7, notices already required by the Equal Credit Opportunity Act or the Fair Credit Reporting Act satisfy the Colorado duty if they carry the extra information, in the notice or a simultaneous supplement within 30 days.
Naming the sources
For the marketing and data industry, the most consequential sentence may be in Rule 7.6. When a consumer asks for more information, the deployer must give the name of the system, its version number, its developer and the types, categories and sources of the personal data it used, and "must identify each source by name when that information is known to the Deployer."
The rule spells out the presumption with a data broker example: "if a Deployer receives Personal Data from Data Broker A and that Personal Data was considered by the Covered ADMT, it must at a minimum disclose Data Broker A as a source of Personal Data." It continues: "specific data brokers, databases, courts, social media companies, schools, employers, and any other source of Personal Data for the Covered ADMT, must be identified by name when that information is known to the Deployer." Where data passed through an aggregator, the deployer must name "the original source of the information and any intermediary or vendor source(s)" when known.
Inferences count as data. "If the Covered ADMT generated any profiles or inferences, the types and categories of those profiles or inferences constitute types and categories of Personal Data that must be disclosed." The rules give "Inferences about your socioeconomic status" and "profile about your educational attainment and future earning potential" as examples of adequate descriptions.
How the consumer rights process runs
According to Rules 7 and 8, the process after an adverse decision has eight stages.
- Disclosure. Within 30 days, the deployer sends a written or electronic disclosure through every channel it normally uses with the consumer. One way to request more information must be an online link, usable "at any time of day."
- Request for more information. The deployer answers within 15 days and need answer only one such request per adverse outcome. No identity check may be required.
- Request for instructions. The consumer asks how to obtain or correct the personal data used. A deployer that deals with consumers only online and has a direct relationship with them may offer just an email address; others must offer two or more methods, including a webform where they have a digital presence. A consumer cannot be forced to create a new account.
- Authentication. For access and correction, the deployer uses a "commercially reasonable" identity check. It may not charge a fee, and may not demand a notarized affidavit "unless the Deployer compensates the Consumer for the cost of notarization." Data collected for the check may not be reused and must be deleted as soon as practical.
- Access and correction. Within 45 days, extendable once by 45, the deployer provides "all the specific pieces of Personal Data used," including the final score or inference and all ADMT inputs. Opinions, predictions and scores need not be corrected, and a correction may be refused if the data is "more likely than not" accurate.
- Human review request. The consumer can ask for meaningful human review at the same time as access or correction, explain why, and attach evidence. The deployer confirms receipt within 10 days and says whether it considers review commercially reasonable.
- Review and reconsideration. The review is completed within 45 days. One 45-day extension is allowed only if the deployer met every other deadline, the delay causes no material harm and the adverse outcome is stayed, and never where the decision affects "a Consumer's access to food, shelter, or required medical care."
- Response and records. The deployer states whether the reviewer confirmed or overrode the decision, the type of review, the factors considered and the reasons, which must be "specific to the evidence provided, and not a recitation of the Covered ADMT's general logic." It keeps records of the reviewer's identity, training, timestamps and written justification for three years.
What makes a review meaningful
Rule 8.7 sets standards for the reviewer, who may be internal or external but must have "actual authority to approve, modify, or override a consequential decision," access to the evidence and relevant training. The reviewer "must not be subject to steering by the upper management that would influence the reviewer's decision, and they must be shielded from potential retaliation."
The rules limit machine help in the review itself. "Except as described below, ADMT may not assist in the Meaningful Human Review." The exception covers deterministic tools that apply "predefined, human-understandable, and human-authored rules" and do not use "machine learning, foundation models, or large language models." Software that merely checks whether materials arrived on time is allowed.
An override is "one factor that may indicate" a meaningful review, but "non-reversal does not, on its own, mean that human review was not meaningful." When a consumer corrects data, reconsideration is presumed commercially reasonable if it means "re-running the exact same ADMT system process that was used to make the initial Consequential Decision with the corrected Personal Data."
Deployers can argue that a full review is not commercially reasonable, weighing nine factors including the magnitude of harm, reversibility and their own size. The burden is theirs, backed by "specific evidence," and for mass challenges to the same system that evidence may be a review of "a randomly-selected sample of those requests, of sufficiently significant size."
Chatbots: who is in and who is out
Rules 9 through 14 implement HB 26-1263. The statute applies to conversational AI services made accessible to the general public, which the rules define broadly: free services "with or without a gated sign-up," subscription services and those "behind a paywall." Internal workforce tools running in authenticated, non-consumer-facing environments are excluded.
The statute also carves out a list of services, and Rule 9.2 explains how the Department will read it. It will "look beyond an Operator's stated design intent to objective evidence" of a product's design and function. Commerce bots are excluded if their functions are limited to tasks such as purchase-status updates, billing questions, shopping, appointment scheduling, "product or service recommendations," orders, payments, deliveries, returns and customer service. Menu-based and rule-based bots qualify for the narrow-topic exclusion only "as long as these chatbots cannot generate outputs related to sexually explicit conduct or depictions" or "maintain dialogue related to suicidal ideation or self-harm."
Business tools are judged on availability to consumers, marketing and design. Voice assistants are assessed partly on whether the service "is designed to simulate emotional companionship or encourage emotionally dependent interaction." Video game dialogue is excluded if it stays on the game and conversation is not the game's main component.
The companionship test matters most, because several exclusions are lost if a product simulates companionship. The Department will weigh whether a product lets users assign a chatbot "a name, gender, avatar, or fictional backstory"; whether training and deployment "prioritized, promoted, or failed to address sycophantic behavior or outputs that simulate empathy and offer Users nonjudgmental responses and continual validation"; and whether it can recall past conversations "to create or reinforce a sense of an ongoing personal relationship." No single factor decides the question. The traits overlap with features that other states' companion chatbot laws have targeted, including New York's.
Age estimation without ID-only checks
HB 26-1263 requires operators to use "commercially reasonable methods or generally accepted methods" to estimate the age of users. Rule 10.2 lists what the Department will consider, and it starts with privacy. Data collected for age estimation may be used only for that purpose and "should be deleted after the minimum time required for compliance." The rule adds: "Obtaining more precise personal data than required is not evidence of a more reasonable determination."
Then comes the provision in the headline. Methods should provide "User choice in how the User's age can be estimated," and "An Operator must not use as the sole method of age estimation a method that requires the User to provide government-issued identification." Self-declared ages, contractual terms and online payment methods available to minors cannot be relied on alone either.
A method should be "reasonably effective at identifying Users under age 18" and tested, with results showing "the rate at which Minor Users are incorrectly determined to be over age 18, reported by relevant Age Category." The three age categories are under 13, 13 to 15, and 16 to 17. Operators should consider operating system, app store and device signals where available and lawful, and re-assess estimates when conversational content, usage patterns or references to "age or grade level" contradict them.
Rule 10.2(D) names five generally accepted methods:
- A system that conforms to ISO/IEC 27566-1:2025, the international standard for age assurance systems, and where applicable IEEE 2089.1.
- A cryptographic technique "such as a zero-knowledge proof" that lets a user show whether they are a minor without revealing anything else.
- Matching a scan of a government ID with a live photo or video using facial recognition.
- An assessment of the user's digital footprint "originating from a verified email address."
- For school-provisioned accounts, an age or grade supplied by the school through a roster, single sign-on or similar integration.
The statute bars operators from willfully disregarding "clear and convincing information" that a user is a minor. Evidence of that includes designing a service "to avoid receiving or acting on information that a User is a Minor," ignoring parental complaints or system flags, and "age reporting tools that automatically fill in a birth date over age 18."
Reminders every three hours
Rule 11 sets the cadence of the statute's "this is AI" disclosures. Minors must be told they are "interacting with artificial intelligence that is artificially generated and not human." On screens, a persistent disclaimer must stay visible "throughout the conversation and not require a Minor User to scroll, click, hover, or take other affirmative action to view the disclaimer," and must stand out by size, contrast and location. Audio-only products repeat a spoken disclaimer, including whenever the user asks whether the service is human.
Where the statute requires periodic disclosures to minors, the rules set them at the beginning of each interaction and "at least once every three hours during a Continuous Conversational Artificial Intelligence Service Interaction." A new interaction includes starting a conversation with a different character or profile. The rule's own example: a minor who begins at 9:00 a.m. must get a disclosure then and another by 12:00 p.m.; if the minor leaves at 10:00 a.m. and returns at 10:30 a.m., the clock keeps running; if the minor returns at 11:30 a.m., the 60-minute gap has ended the interaction, and a fresh disclosure is due.
Adults get a lighter version. The disclosure must appear at the start of the first interaction of each calendar day, and then either persistently or at three-hour intervals during a continuous interaction. It must also be given "in response to User prompts regarding whether the Conversational Artificial Intelligence Service is artificially generated and not human, including in role-playing settings."
The approach tracks laws in two larger states. California's SB 243, approved on October 13, 2025, requires companion chatbot operators to notify minors at least every three hours, and New York's existing AI companion law has required notices at least every three hours since November 5, 2025. A bill the Connecticut Senate passed on April 21, 2026 proposed disclosures at least hourly.
Engagement design and emotional dependence
Rule 12 addresses the statute's protections for known minors. Operators may not offer minors "points or similar rewards at unpredictable intervals with the intent to encourage increased engagement." The Department may consider whether a reward "increases, sustains, or extends a Minor's session length, return frequency, or platform use time" through variable reinforcement. Covered features may include "leaderboards, points, badges, or login streaks," features requiring users to "play a game, answer trivia questions, spin a wheel," and access to extra features "based on session length, return frequency, or platform time."
The statute requires "technically feasible measures" against sexual content involving minors. The Department will weigh whether comparable safeguards exist in the industry and whether alternatives exist, including "substantially altering the Conversational Artificial Intelligence Service or choosing not to deploy" it. One factor stands out for anyone who builds on large language models: "Whether the safeguard depends on a model following natural-language instructions." Operators are expected to test safeguards before deployment and periodically after, "using realistic and adversarial scenarios involving Minors or through adult human evaluators simulating interaction as Minor Users."
A separate list covers responses that simulate emotional dependence or isolation from real-world supports. Operators should test whether their safeguards stop a chatbot from suggesting it is human, "Claiming to be in a Minor-adult romantic relationship," encouraging a minor to withdraw from family, friends or professionals, encouraging secrecy, portraying itself as the minor's "primary, exclusive, or preferred relationship," using "expressions of loneliness, abandonment, guilt, jealousy or emotional distress to encourage continued engagement," conditioning affection on engagement, pressuring a minor to stay, and "Implying sentience or suffering." Safeguards are expected to hold up "during prolonged conversations."
Minors and their parents or guardians must get tools to control whether a chatbot "retains information from prior interactions or sessions" for personalization and whether the minor's personal data is used for training. Neither set of tools may use a dark pattern, and both must follow the principles in the Colorado Privacy Act rules. The rules do not require operators to retrain or alter "any underlying Artificial Intelligence model, model weights, or training corpus."
Rule 13 deals with the statute's ban on claiming professional credentials. The Department will consider whether an operator has controls "reasonably designed to prevent the creation of Conversational Artificial Intelligence Service profiles with names that include terms such as 'therapy,' 'therapist,' 'psychologist,' 'doctor,' 'lawyer,' 'dietitian,' or similar words," as well as controls on outputs and the way the service is advertised.
The annual report
The statute requires operators to report annually to the Attorney General. Rule 14 defines the content, starting with size. Each operator must state its tier:
| Tier | Monthly Active Users |
|---|---|
| Tier 1 | Greater than 10,000,000 |
| Tier 2 | Between 5,000,000 and 10,000,000 |
| Tier 3 | Between 1,000,000 and 5,000,000 |
| Tier 4 | Between 500,000 and 1,000,000 |
| Tier 5 | Fewer than 500,000 |
The statute already requires a count of crisis service provider referrals. The rules require context to read it: the total number of conversations in the period and how a conversation is counted; the share of referrals directed at outputs "determined, using an Evidence-Based Method, to involve suicidal ideation or self-harm risk"; and, "Where such measures can be responsibly estimated," precision, recall, false-positive and false-negative rates. The rules name the Columbia-Suicide Severity Rating Scale, the Patient Health Questionnaire item 9 combined with the Ask Suicide-Screening Questions, and the SAFE-T protocol as examples of evidence-based methods. Time from detection to referral is reported where measured.
Operators must describe their suicide and self-harm protocols element by element, including whether each works "through model behavior, system-level instructions, automated classifiers or filters, human review." The rules list 12 possible elements, among them when law enforcement is contacted and "Any circumstances in which risk is detected but no crisis resource is provided and no human is involved, and the reason."
Testing data must separate "failures to detect relevant risks from failures to respond appropriately to risks that were detected," and flag material differences across languages and model versions.
Rule 14.4 adds age metrics: the distribution of users across the three minor age categories, a description of age-estimation methods and testing results, the number of outside reports that a user was a minor and how they were resolved, the average and median time to a determination, and "The number and percentage of Users whom the Operator initially estimated to be adults but subsequently determined to be Minors." For four categories of sexual content involving minors, operators must report outside reports, how many proved true, how many the operator found itself, and resolution times.
The first report is due "on or before July 1, 2027, covering the preceding calendar year," through a form on the Attorney General's website. For periods before a duty applied, an operator may state that the requirement "was not applicable." A report prepared for another jurisdiction counts if it is "reasonably similar in scope and effect," with an appendix for anything missing. California's SB 243 also starts annual reporting to that state's Office of Suicide Prevention on July 1, 2027, according to PPC Land's coverage of the law. The Department may demand underlying documentation or "a demonstration sufficient to verify any element of the Annual Report," which operators must provide within 30 days.
Accessibility, language and standards
Rule 3 requires every notice under both laws to avoid "technical or legal jargon," work on mobile screens and appear in every language the company uses with consumers. Online material should follow the Web Content Accessibility Guidelines 2.2, which Rule 15 incorporates by reference.
The lawsuit that sets the clock
The rules arrive under a federal court order. xAI sued Attorney General Philip J. Weiser on April 9, 2026 in the U.S. District Court for the District of Colorado, claiming the original AI Act violated the First Amendment, and the U.S. Department of Justice intervened on April 24, 2026 under the Equal Protection Clause. On April 27, Magistrate Judge Cyrus Y. Chung granted a joint motion to suspend case deadlines and temporarily stay enforcement.
According to the court docket in X.AI LLC v. Weiser, No. 1:26-cv-01515, the minute order bars the Attorney General from initiating enforcement, including investigations, for violations of SB 24-205 or any replacing or amending legislation enacted during the 2026 session that occur on or before 14 days after the court rules on a forthcoming xAI motion for a preliminary injunction. xAI must file that motion, and any amended complaint, within 28 days after final adoption of rules implementing SB 24-205 or its replacement.
The order does not define final adoption. If the November 19 adoption counts, the 28 days would run out on December 17, 2026, by PPC Land's calculation; if it means the rules becoming final after publication, the deadline falls later. The order also refers only to SB 24-205 and its replacement, and does not name HB 26-1263.
Federal pressure extends beyond the courtroom. On July 1, 2026, the Federal Trade Commission released a proposed policy statement that named Colorado's law as amended by SB 26-189 as one that could push companies toward undisclosed output steering, and argued for implied preemption. The statement remains a proposal.
Why this matters for marketers and platforms
The ADMT rules reach well beyond banks and landlords. Any company that uses scoring or classification in employment decisions, or in pricing that limits access to education, housing, credit or other covered services, now has a detailed template for what it must disclose when the answer is no. The source-naming requirement in Rule 7.6 pulls data suppliers into view: a consumer who asks will learn which broker, aggregator or social platform supplied the data behind a denial. PPC Land has tracked similar pressure in California, where SB 947 bars employers from relying solely on automated systems for discipline or firing from July 2027.
For brands running conversational assistants, the commerce exclusion offers room, but only for bots whose functions are limited to shopping, orders, support and similar tasks. A brand character that remembers a shopper's past conversations, invites them to name it, or offers constant validation edges toward the companionship factors in Rule 9.2. Engagement mechanics familiar from loyalty programs, such as streaks and spin-to-win rewards, are listed as features that may count as prohibited rewards when aimed at minors.
The age rules add work. Operators must test their methods, report error rates by age band and stop relying on self-declared birthdates, yet cannot make government ID the only route. Research PPC Land covered in August showed how loose current gating is: 2,589 of 4,346 AI companion apps carried store ratings accessible to minors. Federal scrutiny runs alongside: the FTC ordered seven companies, including Alphabet, Character Technologies, Meta, OpenAI, Snap and X.AI, to report on chatbot child-safety measures in September 2025, and California enacted SB 1119, a companion chatbot risk-assessment law, on September 10, 2026.
The reporting regime may matter most in practice. Precision and recall figures for crisis detection, failure data and age-misclassification counts will reach the Colorado Attorney General from every covered operator, sorted by size tier, starting in July 2027.
What the rules leave open
The rules do not define Monthly Active Users, and the tier bands share their boundary figures, so an operator with exactly 5,000,000 users fits two tiers. Several cross-references do not match the text's own lettering: Rule 8.7 refers to subsections (F) and (G) for timing and documentation that appear under (E) and (F), and Rule 9.2(E) refers to factors "9.1(H) and (I)" that sit in Rule 9.2. The Department's page gives July 1, 2026 as HB 26-1263's signing date, while the General Assembly's record shows May 29, 2026. The rules' effective date depends on an Attorney General opinion with no stated deadline.
Key facts
- Official title: Automated Decision-Making Technology & Conversational Artificial Intelligence Service Rules
- Citation: 4 CCR 904-6, Colorado Department of Law, Consumer Protection Section
- Length: 76 pages, Rules 1 to 15
- Statutes implemented: Automated Decision-Making Technology in Consequential Decisions Act (SB 26-189) and Conversational Artificial Intelligence Services Act, "Chatbot Safety Act" (HB 26-1263), sections 6-1-1701 to 6-1-1710, C.R.S.
- Rulemaking authority cited: sections 6-1-108(1), 6-1-1707, 6-1-1704(4), 6-1-1705(3) and 6-1-1706(5), C.R.S.
- SB 26-189 signed: May 14, 2026; obligations from January 1, 2027
- HB 26-1263 signed: May 29, 2026 (General Assembly record); effective January 1, 2027
- Secretary of State tracking number: 2026-00360
- Proposed rules filed: August 11, 2026; interim draft October 6, 2026; regulatory analysis October 21, 2026
- Hearing and comment deadline: October 26, 2026
- Adopted: November 19, 2026; adopted text posted as a double redline dated November 18
- Attorney General opinion requested: November 20, 2026
- Effective date: the date published by the Secretary of State; the Department says the rules become final twenty days after publication
- First chatbot annual report due: July 1, 2027
- Standards incorporated: WCAG 2.2 (December 12, 2024), ISO/IEC 27566-1:2025 (December 2025), IEEE 2089.1 (May 24, 2024)
- Related case: X.AI LLC v. Weiser, No. 1:26-cv-01515-DDD-CYC (D. Colo.), minute order of April 27, 2026 (ECF No. 24)
- Source: https://coag.gov/resources/ai/
Timeline
- May 17, 2024 - Colorado AI Act, SB 24-205, signed
- August 2025 - Special session moves SB 24-205's effective date from February 1 to June 30, 2026
- September 10, 2025 - FTC issues 6(b) orders to seven AI chatbot companies on child safety
- October 13, 2025 - California approves SB 243, with three-hour reminders for minors and reports from July 1, 2027
- November 5, 2025 - New York's AI companion law, requiring notices at least every three hours, takes effect
- April 9, 2026 - xAI sues Colorado Attorney General Philip J. Weiser
- April 21, 2026 - Connecticut Senate passes SB 5 with hourly companion chatbot disclosures
- April 24, 2026 - U.S. Department of Justice intervenes in the xAI case
- April 27, 2026 - Court stays the xAI case and bars enforcement until 14 days after a ruling on a preliminary injunction motion due within 28 days of final rule adoption
- May 9, 2026 - Colorado House passes SB 26-189, 57-6
- May 14, 2026 - Governor signs SB 26-189
- May 29, 2026 - Governor signs HB 26-1263, according to the General Assembly's record
- July 1, 2026 - FTC proposed policy statement names Colorado's amended AI law
- July 13, 2026 - Informal pre-rulemaking comments close
- August 3, 2026 - Deepsee research finds 2,589 of 4,346 AI companion apps rated accessible to minors
- August 11, 2026 - Department of Law files proposed rules with the Secretary of State
- September 10, 2026 - California enacts SB 1119 on companion chatbots
- September 30, 2026 - California approves SB 947 on AI-only discipline and firing decisions
- October 6, 2026 - Interim draft of the rules released
- October 21, 2026 - Regulatory analysis released
- October 26, 2026 - Formal hearing; written comments close
- October 28, 2026 - New York's governor signs the law barring human-posing AI companion features for minors
- November 19, 2026 - Department adopts 4 CCR 904-6
- November 20, 2026 - Department requests the Attorney General's formal opinion
- January 1, 2027 - SB 26-189 and HB 26-1263 take effect
- July 1, 2027 - First chatbot annual reports due to the Colorado Attorney General
Related PPC Land coverage
- Colorado passes SB 26-189: the AI decision law that rewrites the rules - The May 2026 vote that repealed and replaced the Colorado AI Act and set the January 1, 2027 rulemaking deadline.
- xAI sues Colorado to block AI bias law, claiming First Amendment violations - The April 9, 2026 complaint against Attorney General Philip J. Weiser.
- U.S. joins xAI's fight against Colorado's AI bias law - The Justice Department's Equal Protection complaint in intervention.
- FTC move could force Colorado to rewrite new AI bias law - The FTC's July 2026 proposed policy statement on AI output accuracy and preemption.
- California law requires AI to tell you it's AI - SB 243's disclosure, suicide protocol and July 2027 reporting rules for companion chatbots.
- Hochul signs law barring human-posing AI chatbots for under-18s - New York's list of 12 unsafe AI companion features and its age assurance gate.
- Connecticut's AI bill targets companion bots, hiring tools and frontier models - The Connecticut Senate's SB 5, with hourly chatbot disclosures.
- FTC orders seven AI chatbot companies to detail child safety measures - The September 2025 6(b) orders covering age gating, engagement metrics and outputs to minors.
- Pinterest rules out companion AI as California enacts Adam's Law - California's SB 1119 risk-assessment requirements for companion chatbot operators.
- Deepsee: buyers lose age ratings on 60% of 4,346 AI companion apps - Research on app store age ratings and how easily age gates were bypassed.
- California bans AI-only firing decisions from July 2027 - SB 947's limits on automated decision systems in employment discipline.
- Treat every Character.ai chat as a child's, groups ask bankruptcy judge - Child-safety groups' requests in the Character.ai bankruptcy sale.
Summary
Who: The Colorado Department of Law, under Attorney General Philip J. Weiser, and the companies covered by two state laws: developers and deployers of automated decision-making technology used in consequential decisions, and operators of consumer-facing conversational AI services. xAI and the U.S. Department of Justice are challenging the underlying AI law in federal court.
What: The Department adopted 4 CCR 904-6, 76 pages of rules implementing SB 26-189 and HB 26-1263. They define when an algorithm materially influences a decision, prescribe 30-day adverse outcome disclosures that name data sources including data brokers, set 10-, 15- and 45-day deadlines for consumer rights and human review, bar ID-only age checks for chatbots, require AI disclosures at least every three hours, and specify the content of annual chatbot safety reports.
When: Adopted on November 19, 2026, with an Attorney General opinion requested on November 20. The rules become final twenty days after publication by the Secretary of State. The statutes take effect on January 1, 2027, and the first chatbot reports are due on July 1, 2027.
Where: Colorado, covering consumers and, for automated decisions, employees and job applicants evaluated by companies doing business in the state.
Why: SB 26-189 requires rules on post-adverse outcome disclosures and consumer rights before January 1, 2027, and the Department chose to issue rules under HB 26-1263 as well. The result gives companies operating in Colorado detailed obligations on disclosures, data sourcing, age estimation and safety reporting, while the xAI litigation determines when enforcement can begin.
Discussion