Colorado documentation for deployers

Colorado's Automated Decision-Making Technology Act (SB 26-189, signed 14 May 2026, effective 1 January 2027; it repealed and re-enacted SB 24-205) gives developers and deployers of systems used in employment decisions specific duties. This page is the developer documentation we owe you and the checklist for what you owe candidates.

Last updated 9 September 2026.

Facts on this page were verified against the Colorado Attorney General's AI page and the enacted bill on 5 September 2026. The Attorney General's rulemaking is open (proposed rules filed 11 August 2026, comments to 26 October 2026); we update this page when the rules are final.

Who is who

InterviewStack is the developer of the screening interview system. The employer that configures a screening and uses its output is the deployer. The candidate is the consumer.

Developer documentation

Intended use

A first-round screening interview for knowledge-work and technical roles: a 30-minute AI-conducted conversation from an employer-configured plan, producing a transcript, a recording where enabled, a written assessment with four reasoned dimension scores and an overall 0-100 signal, and session context. It is intended to give a human recruiter evidence for their own decision. It is not intended to make or recommend the decision, to compute a ranking of candidates, to verify identity, or to assess anything other than the content of the answers. See how the assessment works.

Training data categories

InterviewStack does not train or fine-tune models. The interviewer and the scorer are large language models provided by OpenAI, used through prompts we author (the plan, the rubric and the scoring instructions). We hold no information about those models' training data beyond OpenAI's published documentation. No candidate data is used to train models by us or by our providers (see subprocessors).

Known limitations and risks

Language-model variability: the same transcript can score a few points differently on two runs.
Evidence sensitivity: short answers score low for lack of evidence; the scorer cannot distinguish a candidate who knows but said little from one who does not know.
For interviews in a language other than English, language and fluency are excluded from scoring by instruction; for interviews in English there is no separate fluency rule. An instruction excluding fluency in English was tested on 9 September 2026 and not shipped: it raised borderline scores by about four points regardless of fluency. We test name and pronoun effects on synthetic transcripts; we cannot rule out effects that come through dialect or phrasing, which the synthetic test does not vary.
Transcription: spoken answers pass through speech recognition. An answer is marked as a transcription, answer by answer, when our own speech-to-text produced most of its text or the employer required spoken answers; the marking never comes from the candidate's browser. It proves that our transcription produced the text, not that a person spoke it: a machine-read answer would pass, and where recording is enabled the recording is the employer's check. Accents can still reduce transcription accuracy.
No identity verification and no detection of outside help beyond the session context shown to the employer.

Performance evaluation and bias testing

Internal test on synthetic transcripts with name and pronoun variants (first run 9 September 2026: 17 transcripts, 8 variants each, largest group gap 1.1 points inside run-to-run noise, four-fifths ratio 1.00 at a threshold of 60 and 0.88 at 70 with an interval from 0.71 to 1.00; method and limitations in how the assessment works). No independent audit has been performed yet. We will support one when a customer's jurisdiction requires it, including supplying synthetic test data, which NYC Local Law 144 permits.

Human review instructions

The review page shows the transcript, the recording, the four dimensions with reasoning, coverage of the plan, the session context and the AI signal with its caveat. The recruiter records Advance or Reject; nothing advances or rejects on its own. A recruiter can record their own score with a written reason beside the AI signal, and can change that score or their decision at any time; the AI signal is produced once and never regenerated. Treat session flags as prompts for a conversation with the candidate, not conclusions.

Data governance

Candidate data handling is described in the candidate privacy notice and governed by the data processing agreement. Retention is set per workspace and fixed for each candidate at the moment they consent, so a later change never lengthens it for them. Deletion on request is available to the employer at any time; it removes name, email, transcript, recordings and the written assessment, and the numeric score stays with the employer's record of the decision.

Material updates and records

We notify workspace owners by email before deploying a change to the dimensions, their weights, what is excluded from scoring, or what the deployer receives, and we record it in the changes list below with its date. Records: we keep this documentation, its dated change history and the anti-bias log results for at least three years. Interview data (plan, transcript, assessment, session context) is separate: held for the workspace's retention period; export what you need before it ends.

Deployer checklist

Notice at the point of interaction. Every candidate sees the candidate privacy notice and gives consent before the interview starts. Add your own notice where your process requires it.
Plain-language description after an adverse decision. Point the candidate to how the assessment works and, from the review page, the dimension reasoning that applied to them. The Act sets a 30-day window after an adverse decision.
Access and correction. The review page and the CSV export cover what you would disclose to a candidate: name, email, transcript, recordings, assessment, session events and your decision. To correct a name or email, or for a full data inventory, email support@interviewstack.io.
Meaningful human review and reconsideration. Built in: no decision is made by the system, and any recorded decision can be changed.
Records. Keep your screening configuration and decisions for the period the Act and its rules require; our retention setting controls how long we hold the underlying data.

A violation of the Act is enforced by the Attorney General as a deceptive trade practice. This page is documentation, not legal advice; your counsel should confirm how the final rules apply to your process.

Changes

9 September 2026: spoken answers are marked as transcriptions answer by answer from our own transcription record, also when speaking is optional. Per-turn times removed from the transcript the scorer reads.
9 September 2026: first version.