Short answer: an LL144 bias audit needs a defined audit population, the AEDT outcome or score meaning, the relevant time period, available sex and race/ethnicity data, job or position context, data-source provenance and enough evidence to reproduce the required rates and impact ratios. Missing demographics should not be filled by inference.
The required calculations depend on the AEDT output. The rules distinguish selection/classification outputs from score outputs, require sex, race/ethnicity and intersectional analysis, and require reporting individuals omitted because they fall within an unknown category. Historical data is preferred when available; test data may be used under the rule's conditions.
What data elements does an LL144 bias audit need?
| Element | Why it matters |
|---|---|
| Audit population | Defines who was assessed by the AEDT for the relevant hiring or promotion workflow. |
| Outcome or score | Determines whether selection-rate or scoring-rate calculations apply. |
| Sex and race/ethnicity categories | Required for separate and intersectional calculations. |
| Unknown states | The rules require the count of assessed individuals omitted because of unknown categories. |
| Audit period and source | Supports provenance and the public explanation of the data basis. |
| AEDT/version/workflow | Helps ensure the dataset corresponds to the tool and use being audited. |
Can an auditor infer missing demographic data?
DCWP guidance does not permit the audit to replace missing demographic information with imputed or inferred protected characteristics. Missingness must be handled as a controlled data state, not guessed from names, geography, photos or resumes.
When can test data be used instead of historical data?
Historical data is data collected from actual use of the AEDT. Test data is not historical data. The audit record should identify which basis was used and why.
Frequently asked questions
The audit must account for unknown categories, but missing protected-category data cannot simply be inferred. Material missingness may require a data-readiness review.
No candidate names are needed for the public marketing-site tools. The controlled audit intake should minimize direct identifiers and preserve only what is necessary for the audit.
No. Selection/classification outputs and continuous score outputs follow different calculation paths under 6 RCNY §5-301.
What should you do before sending data to an auditor?
Use this page for general information only. For a specific workflow, review the LL144 scope framework or request an independent bias-audit scope review. Do not send candidate-level data through ordinary email.