AI Bias Audit for a Hiring Technology Platform

Problem: An AI-powered hiring and candidate-scoring platform needed an independent, annual bias audit to comply with NYC Local Law 144, which requires employers and employment agencies using Automated Employment Decision Tools to undergo an independent bias audit and publicly summarize the results.

Goal: Determine, with statistical rigor and no intervention in the live system, whether the platform’s scoring outputs showed adverse impact across sex, race and ethnicity, the White versus Non-White binary, and every relevant intersectional subgroup, and deliver a report the client could publish to meet its legal obligation.

Geography: United States, New York City jurisdiction.

Stage 1: Data intake and preparation

The client provided anonymized historical candidate-scoring records as CSV data, alongside system configuration and output data as JSON. No synthetic or simulated data was used, every record reflected a real, historical scoring decision the system had already made. BI Company did not access or modify the live system at any point.

Stage 2: Demographic classification

Because the platform does not directly collect race, ethnicity, or gender, BI Company inferred these attributes using established, disclosed methods: a name-to-gender inference approach for sex, and Bayesian Improved Surname Geocoding, which combines surname and geography, for race and ethnicity.

Stage 3: Statistical testing

BI Company ran four separate analyses in SPSS against the full dataset of 81,165 historical candidate-scoring records, covering a one-week audit window in the summer of 2025:

• Sex bias: a Chi-square test and Adverse Impact Ratio comparing male and female selection rates
• Race and ethnicity bias: selection rate and Impact Ratio across five groups, Arabic, Asian, Black, Hispanic or Latino, and White
• White versus Non-White bias: a Chi-square test and Adverse Impact Ratio on the binary grouping
• Intersectional bias: selection rate and Impact Ratio for every combination of sex and race or ethnicity

Stage 4: Grading against a fixed standard

Every result was measured against the EEOC’s four-fifths rule (an Impact Ratio below 0.80 signals potential adverse impact) and graded on a three-tier scale: Clear (0.80 or above), Consider (0.60 to 0.79), or Concern (below 0.60). The platform’s own pass threshold for an individual candidate score was a composite score of 60 points or higher, a separate, product-level threshold from the audit’s own statistical grading scale.

If your organization needs an independent AI bias audit for NYC Local Law 144, or wants an audited trust signal for an AEDT you sell to enterprise customers, talk to our research team. See our AI bias audit services page for the full methodology.