Colorado AI Act (SB24-205) — CRS §6-1-1703
CRS §6-1-1701 — Categories of consequential decisions
Training data is predominantly from large financial institutions; may not generalize well to credit unions or community banks with different hiring patterns.
System performance degrades for roles with fewer than 50 historical hiring records in the training data.
Resume parsing accuracy drops to 78% for non-standard resume formats (e.g., creative layouts, non-English sections).
Historical hiring data reflects industry-wide underrepresentation of women in senior financial analyst roles (32% vs 50% population baseline). Mitigation applied via reweighting.
Name-based features were removed after detecting proxy discrimination patterns in v2.1 audit. Current version uses anonymized candidate IDs during scoring.
Demographic reweighting applied to training data to correct for historical underrepresentation in senior roles.
All personally identifiable information (name, address, age indicators) removed before model scoring. Candidates are evaluated on skills, experience, and assessment scores only.
Quarterly disparate impact testing using 4/5ths rule across race, gender, age, and disability status. Results published internally and available upon request.
Human review required for all candidates ranked in bottom 20% before rejection — ensures AI recommendations are not sole basis for adverse decisions.
| Metric | Value | Benchmark |
|---|---|---|
| Selection Rate Parity (Gender) | 0.87 | ≥ 0.80 (4/5ths rule) |
| Selection Rate Parity (Race) | 0.83 | ≥ 0.80 (4/5ths rule) |
| Selection Rate Parity (Age 40+) | 0.91 | ≥ 0.80 (4/5ths rule) |
| Quality of Hire Score | 4.2/5.0 | 3.8/5.0 (industry avg) |
| Time-to-Hire Reduction | 60% | 30% (industry avg) |
| False Negative Rate | 8.3% | ≤ 15% |
Qualified candidates may be ranked lower than warranted if their experience doesn't match patterns in training data (e.g., career changers, non-traditional backgrounds).
Candidates from underrepresented groups may face compounded disadvantage if historical bias in training data is not fully mitigated.
Over-reliance on AI rankings by hiring managers could reduce consideration of qualitative factors not captured by the model.
I certify that this impact assessment has been conducted in good faith and accurately represents the current state of the AI system described above, in accordance with CRS §6-1-1703.