Colorado's ADMT Law and Synthetic or Test Data: What the Text Says and Doesn't
Nothing in SB 26-189, the Attorney General's interim draft rules, or the regulatory analysis mentions synthetic data, and nothing in them exempts anyone for using it. The law's reach turns on whether a system processes personal data and materially influences a consequential decision, not on where its training or test data came from.
Last reviewed: November 13, 2026 · Independent coverage by Synthetic Data News · Editorial standards
Not legal advice. This article is general information from an independent publication, based on the public sources listed below. It is not legal advice and does not create an attorney-client relationship. Consult qualified counsel about your situation.
What SDN checked
SDN searched the enacted text of SB 26-189, the AG's October 6, 2026 interim draft and the October 21, 2026 regulatory analysis for the words "synthetic," "test data," "training data" and "de-identified." The term "synthetic" does not appear in any of the three. The statute does refer to data used to train a system, in the developer documentation provision discussed below. The absence of a term is not a finding that synthetic data is treated any particular way; it means the law is silent, and readers should not infer an exemption.
Where the definition of ADMT points
The statute defines automated decision-making technology as technology that "processes personal data and uses computation to generate output," such as predictions, recommendations, classifications, rankings or scores, used to make, guide or assist a decision about an individual (C.R.S. 6-1-1701(2)(a)). "Personal data" has the meaning in section 6-1-1303(17) (C.R.S. 6-1-1701(16)). The draft rules define "process" by reference to 6-1-1303(18) as collection, use, sale, storage, disclosure, analysis, deletion or modification of personal data (draft Rule 2.2).
On a plain reading, the ADMT definition describes what a technology does when it runs. A model that scores real applicants or employees in a covered domain processes their personal data at that point, whatever data was used to build it. That is SDN's reading of the text, not AG guidance, and it has not been tested in enforcement or court.
Developer documentation and training-data categories
A developer of a covered ADMT must make available to each deployer, on and after January 1, 2027, a description of "the categories of data, including personal data, used to train the covered ADMT, to the extent known" (C.R.S. 6-1-1702(1)(b)). The same subsection requires a statement of intended and known harmful uses (1)(a), known limitations and risks (1)(c), instructions for use, monitoring and meaningful human review (1)(d), and information the deployer reasonably needs for its own notices (1)(e).
The draft rules would add that the description of training-data categories use a level of detail that gives deployers a meaningful understanding of the data, including whether it would be considered sensitive data or biometric data under the Colorado Privacy Act (draft Rule 5.2(D)). Developers may withhold trade secrets or legally protected information but must notify the deployer (C.R.S. 6-1-1702(1)(e)), and the draft says they must cite the legal authority and still provide as much information as possible (draft Rule 5.3).
What the law does not say: whether a developer that trained on synthetic data must describe it as synthetic, name the source population it was derived from, or characterize its privacy properties. Those are open questions. A developer's description would still need to meet the "to the extent known" and meaningful-understanding standards, and counsel is the right source on how to apply them.
Test and evaluation data
The enacted text does not require testing, bias audits or impact assessments (see our explainer on what changed from SB 24-205). It does require developers and deployers to keep records "reasonably necessary to demonstrate compliance" for at least three years: for developers, three years after a record is created (C.R.S. 6-1-1702(4)); for deployers, three years after the date of a consequential decision (C.R.S. 6-1-1703). Records may include system version identifiers, changelogs and documentation of material mitigation changes (6-1-1703).
Organizations that evaluate a system before release, using real, synthetic or hybrid data, may find evaluation results useful as a source for the limitation and risk statements in 6-1-1702(1)(c) and for change documentation. The statute does not say test data or test results must be kept, and does not set any standard for what a test must show.
Who is outside the developer definition
A person that develops and uses an ADMT solely for research, where it is not used in a consequential decision in the research, or for internal purposes without making it available to another person for use in a consequential decision, is not a "developer" (C.R.S. 6-1-1701(8)(b)(I)). Building a model purely on synthetic data does not by itself decide which side of that line a team is on; what the system is used for and who it is made available to does.
Liability is not displaced
The act states that use of an ADMT in a consequential decision does not excuse or provide a defense to obligations or liability under state or federal law, including discrimination and consumer protection law (C.R.S. 6-1-1707(9)), and that compliance with part 17 is not a defense to other law (6-1-1709(2)). Nothing in the text suggests that the origin of training data changes that.
Frequently asked questions
Does Colorado's AI law exempt systems trained on synthetic data?
No exemption appears in the enacted text of SB 26-189. The law does not mention synthetic data. Coverage depends on the ADMT, consequential decision and covered domain definitions (C.R.S. 6-1-1701).
Does synthetic data mean a system doesn't process personal data?
The statute does not address this. The ADMT definition refers to technology that processes personal data (6-1-1701(2)(a)); a system scoring real individuals in a covered domain would process their data when deployed. That is a reading of the text, not AG guidance.
What must a developer say about training data?
A description of the categories of data, including personal data, used to train the covered ADMT, to the extent known (6-1-1702(1)(b)). The AG's draft Rule 5.2(D) would require enough detail for a meaningful understanding.
Does the law require testing the system on test data?
The enacted text contains no testing, audit or impact-assessment requirement. It requires records reasonably necessary to show compliance (6-1-1702(4), 6-1-1703).
Where can I find out what the AG thinks?
The AG's rulemaking page (coag.gov/ai) and eDocket 2026-00360. As of November 13, 2026 no final rule is shown and the draft does not mention synthetic data.
Sources
- Enacted SB 26-189, Session Laws of Colorado 2026, ch. 131 (C.R.S. 6-1-1701 to 6-1-1709), approved May 14, 2026
- Colorado Department of Law, ADMT & Chatbot rules, interim draft redline, October 6, 2026 (proposed; subject to change)
- Colorado Department of Law, Regulatory Analysis, October 21, 2026
- Colorado Attorney General, AI rulemaking page
Facts in this article were checked against these sources on November 13, 2026. Draft rule content is proposed and subject to change.
Other resources
- Colorado Attorney General: AI rulemaking — Official; rulemaking documents and status.
- Colorado General Assembly: SB 26-189 session law — Official enacted text.
- CertifiedData.io: Colorado SB 26-189 guide (vendor resource) — A vendor resource from a platform operated by the same organization as SDN (see our editorial standards). SDN has not evaluated its claims; it is one of several approaches alongside internal logging, general model-risk and governance tooling, and open-source tooling, and legal counsel.
Get the SDN weekly digest
Independent AI governance and synthetic data coverage, once a week.
Related coverage
Not legal advice. This article is general information from an independent publication, based on the public sources listed below. It is not legal advice and does not create an attorney-client relationship. Consult qualified counsel about your situation.