Comment on Colorado ADMT Proposed Rules: Training Data Certification

Box Commons recommends that the proposed rules implementing the Automated Decision-Making Technology Act (SB 26-189) and the Chatbot Safety Act (HB 26-1263) recognize independent, third-party data certification as a structured mechanism by which developers can satisfy their training data documentation obligations and deployers can satisfy their vendor oversight duties.

Specifically, we recommend that the final rules establish a rebuttable presumption that a developer's use of datasets certified by an independent standards body — one meeting defined governance criteria — constitutes adequate documentation of "the categories of data, including personal data, used to train the Covered ADMT" as required by the Act.

This comment draws on our experience operating a live audio data certification pipeline across a 22-station broadcast radio network, demonstrating that the training data transparency SB 26-189 contemplates is technically feasible, operationally affordable, and independently verifiable today.

SB 26-189 requires developers to provide documentation including "the categories of data, including personal data, used to train the Covered ADMT." However, neither the statute nor the proposed rules address how the deployer — or the Attorney General — can verify that the documentation is accurate.

The Verification Asymmetry: Developers know what data they used. Deployers and regulators do not. Disclosure without verification is attestation — the developer's word, unaudited. The insurance industry has already encountered this asymmetry. When ISO released CG 40 47 and CG 40 48 in January 2026, the exclusions exposed a consent verification gap that attestation alone could not resolve. Gallagher Re documented a 978.1% increase in generative AI-related lawsuits from 2021 to 2025. Emerging AI coverage providers — Armilla, Munich Re aiSure, Testudo — now require data provenance audits as a condition of coverage because self-attestation proved insufficient.

Independent data certification interposes a trusted intermediary between the data producer, the ADMT developer, and the regulator through a three-layer architecture:

Standards Body: Publishes technology-agnostic certification standards, accredits third-party certifiers, maintains a public registry of certified datasets.

Accredited Certifiers: Conduct independent audits of datasets against published standards, issue certifications with defined scope and validity.

Certified Data Producers: Submit datasets for audit; receive certification marks that travel with the data through the supply chain.

Under this architecture, the ADMT developer incorporates certified datasets and documents the certification scope. The deployer relies on the certification for vendor oversight. The Attorney General verifies compliance by checking a public registry — no inspection of proprietary models required.

BC-Certified covers audio data across a 22-station broadcast radio network operated by Audilous Media. The certification pipeline verifies consent chains (broadcaster retains original broadcast rights), content separation (human-created vs. AI-generated audio), metadata completeness (artist, title, timestamp, ISRC where available), and format integrity (uncompressed or lossless source files).

This is a live operational system, not a whitepaper. Audio data is flowing through the pipeline today. The 22-station proof demonstrates that training data transparency is technically feasible, not theoretical, and that independent certification can operate at broadcast scale without disrupting station operations.

Audio matters for ADMT because voice cloning, synthetic speech, and audio deepfakes are among the highest-impact AI harms currently facing consumers. Colorado's ADMT rules will govern AI systems that process audio data — systems whose training data provenance has immediate implications for consumer protection.

The HITRUST CSF certification provides the most direct precedent for what we propose. HITRUST resolved the same information asymmetry in healthcare: covered entities needed to verify the security practices of business associates who would not disclose proprietary security architectures. The solution was a trusted intermediary — the certifier — who evaluates the vendor under NDA and issues a credential the covered entity can rely upon.

Replace "cybersecurity" with "training data provenance," replace "business associate" with "AI developer," and the compliance pattern transfers intact to Colorado's ADMT framework.