Colorado's Chatbot Safety Act: One Duty Is Live Now

The date most compliance calendars miss

Most compliance calendars have a single Colorado date on them: January 1, 2027. That is when the Automated Decision-Making Technology Act takes effect, when most of the Chatbot Safety Act's operator duties begin, and when the Attorney General's proposed rules would become effective.

If your product has a conversational surface, it is the wrong date to plan around.

House Bill 26-1263 — enacted as Conversational Artificial Intelligence Service Operator Requirements, though nearly everyone calls it the Chatbot Safety Act — was signed on May 29, 2026. Its effective-date section does not say January 1, 2027. It says the act takes effect "at 12:01 a.m. on the day following the expiration of the ninety-day period after final adjournment of the general assembly," and then supplies the date in a parenthetical: August 12, 2026. That was subject to a referendum petition, which was not filed.

So the Act has been law since mid-August. The definitions it added to C.R.S. § 6-1-1701 — "conversational artificial intelligence service," "operator," "minor," "self-harm" — have been operative since then.

That alone would be a footnote. What makes it matter is how the operative section is dated.

Read the dates inside § 6-1-1708

Section 6-1-1708 carries the operator duties, and four of them have an express start date:

• Consumer AI-nature disclosure — "on and after January 1, 2027" (§ 6-1-1708(3))
• Suicide and self-harm response protocol — "on and after January 1, 2027" (§ 6-1-1708(4))
• The professional-equivalence prohibition — "on and after January 1, 2027" (§ 6-1-1708(5))
• The annual report to the Attorney General — "on and after July 1, 2027" (§ 6-1-1708(6)(a))

Subsection (2), which governs minors, is dated as well — but only in part. Its fourth sentence begins "on and after January 1, 2027, if an operator knows that an account holder or user is a minor," and the eight lettered duties that follow all hang off that clause.

The first two sentences carry no date at all:

> An operator shall use commercially reasonable methods or generally accepted methods to estimate the age of account holders or users. An operator shall not willfully disregard clear and convincing information that an account holder or user is a minor.

Where a legislature dates four adjacent provisions, dates the back half of a fifth, and leaves two sentences undated inside that same subsection, the sound reading is that the undated sentences operate on the act's own effective date. That date was August 12, 2026.

There is a contrary reading, and it is not frivolous: that the whole of subsection (2) is prefatory to the January 1 duties, existing only to define when an operator "knows" a user is a minor for purposes of (a) through (h). We do not think it is the better reading — a sentence phrased "an operator shall use" reads as a duty, not a definition — and we would not advise a client to rely on it.

Why this is not academic

Two features of the age-estimation duty make it worth attention now rather than in December.

The Office has ruled out the common approach

The Attorney General's proposed rules say what the Office expects. Proposed Rule 9.2(B) would require that a compliant method "not solely rely on user self-declarations of age; general contractual restrictions or terms regarding the use of the internet website, online service, online application, or mobile application; or online payment methods that are available to minors."

That is a precise description of how most consumer applications handle age today: a birthdate field at signup, an 18-and-over representation buried in the terms of use, and a credit card. Under the proposed rule, all three together are not enough.

Rule 9.2(E) lists methods the Department would treat as generally accepted — an age-assurance system conforming to ISO/IEC 27566, a zero-knowledge cryptographic proof, government-ID matching against a live photo, or an assessment of the user's digital footprint from a verified email address. Rule 9.2(F) makes commercial reasonableness turn in part on "the size and financial resources" of the operator, which helps a small firm. It does not excuse having no method at all. And Rule 9.2(G) would provide that an inconclusive result cannot be the basis for concluding a user is not a minor.

The definition of "operator" is broader than most people assume

Section 6-1-1701(15.5)(a) reaches a person that "develops and makes publicly available" a conversational AI service or that "offers a conversational artificial intelligence service to a consumer." The second prong means licensing the model from a vendor does not move the duty off your books. And proposed Rule 2.2 defines "accessible to the general public" to include services "provided through a free or purchased subscription, or services or products behind a paywall" — so a subscription gate is not a defense either.

The limits, stated plainly

Two things soften the near-term exposure, and neither eliminates it.

The rules are not in force. Proposed Rule 1.3 sets their effective date at January 1, 2027, so the detail in Rule 9.2 does not bind anyone today. It tells you how the Office reads the statute, which is worth a great deal, but it is not yet law and it will change — a revised draft is due September 23.

And the enforcement hook is incomplete. SB 26-189 adds C.R.S. § 6-1-105(1)(uuuu), making a violation of part 17 a deceptive trade practice, effective January 1, 2027. Until then there is no express deceptive-trade-practice designation attached to a § 6-1-1708 violation, which makes direct enforcement under the Colorado Consumer Protection Act awkward, though not obviously unavailable. From January 1, penalties run to $20,000 per violation, enforced exclusively by the Attorney General, with no private right of action.

A live statutory duty with a soft enforcement path is still a live statutory duty. It is also the kind of thing that reads badly later: in an examination, in diligence, or in a demand letter, the sequence "the duty existed, the firm had notice, nothing was documented" is not one you want to explain.