§Bias audits for hiring AI

Prove your hiring AI treats people equally.

Isonomy audits AI screening tools the way a regulator or a plaintiff’s expert would: across every intersection of race and sex, down to the input causing the gap, and against a fairer model that works as well.

NYC LL144 tablesIntersectional analysisProxy detectionDrift monitoringFairer alternatives

Exhibit AIntersectional audit
Impact ratio by race and sex
Selection rate relative to the most-selected group ยท 12,000 applicants
Impact ratio by race and sex
WomenMenAll
White0.811.001.00
Asian0.880.890.97
Hispanic / Latino0.640.660.71
Black0.560.640.66
Two or more racesn<2%n<2%0.78
All races0.851.00
Adverse impactMonitorPassReference
Read across, then down. The margins are what a single-attribute audit reports. The cells show what it misses: Black women are selected at 0.56× the rate of white men. Synthetic demo data.

§ 01The blind spot

Most bias audits check one attribute at a time. Harm rarely does.

In the sample audit behind Exhibit A, women pass with an impact ratio of 0.85. Black applicants fail at 0.66. Black women, the group that sits in both, are selected at 0.56× the rate of white men. A table that reports race and sex separately never shows that number.

Isonomy reports every combination, tests each one for statistical significance so small groups aren’t mislabelled, and then traces the gap to the inputs and the release that caused it.

How intersectional analysis works

§ 02One audit file

Six analyses, from “is there a gap?” to “what do we change?”

  1. Impact ratios

    Selection and scoring rates in the format NYC Local Law 144 requires, with confidence intervals and Holm-adjusted significance tests. Details

  2. Intersectional analysis

    Every combination of protected attributes, with an alert when a group fails although every single attribute passes. Details

  3. Proxy detection

    Which inputs stand in for race or sex, and whether your features jointly encode them even after the protected columns are removed. Details

  4. Explanation fairness

    Whether the model reasons differently by group, and how many score points each feature costs each group. Details

  5. Drift monitoring

    Impact ratios month by month, tied to model releases, so a regression is caught in weeks, not at next year’s audit. Details

  6. Fairer alternatives

    Retrains candidate models and tests whether one reduces adverse impact without losing job-related validity. Details

§ 03The question that decides cases

“Was there a less discriminatory alternative that worked as well?”

Under U.S. disparate-impact law, that question can decide the outcome. Isonomy’s mitigation lab answers it with evidence, without group-specific cut-offs and without using protected attributes when a decision is made.

0.56 → 0.86

Worst impact ratio, current tool vs. recommended alternative

7 → 0

Groups with statistically significant adverse impact

+6.4 pts

Precision against a job-related outcome. Fairer, and more accurate.

Figures from Isonomy’s demo dataset of 12,000 synthetic applicants with planted bias. Your results will differ.

Exhibit BMitigation lab
Is there a fairer model that works as well?
8 retrained alternatives, each selecting the same share of applicants
0.50.60.70.80.976%80%84%88%Precision vs job-related outcomeWorst impact ratioCurrent toolRecommended
Up and to the right is better. Every alternative here is both fairer and more accurate than the deployed tool: the evidence a less-discriminatory-alternative analysis looks for. Synthetic demo data.

§ 04Who it’s for

Built for the people who answer for the model.

  1. AI hiring vendors

    Answer customers’ audit requests in days, re-test every release, and show the fix, not just the finding. For vendors

  2. Employers

    Check the tools you buy before the law, a candidate or a regulator does, with evidence your counsel can read. For employers

  3. Independent auditors

    Run deeper LL144 audits in less time, on methods you can inspect and defend. For auditors

§ 05Regulation

One evidence base for a set of rules that keeps moving.

RuleStatus, Nov 2026What it asks for
NYC Local Law 144In force since July 2023Annual independent bias audit; published impact ratios; candidate notice
Illinois HB 3773In force since Jan 2026No AI with a discriminatory effect in employment; no ZIP codes as a proxy
California ADS regulationsIn force since Oct 2025Anti-bias testing counts as evidence; AI records kept four years
Colorado SB 26-189Applies from Jan 2027Notice and human review for automated consequential decisions
EU AI ActEmployment rules from Dec 2027Hiring AI is high-risk: risk management, data governance, monitoring

Each rule, with sources

Gideon Popoola, founder of Isonomy

§ 06Founder

An audit that stops at one attribute and one number misses where the harm usually sits: at the intersections, and in how the model reasons.

About Isonomy

§ 07Design partners

Bring one data export. Leave with the evidence.

We’re working closely with three design partners: a full audit and monitoring at no cost while we build the product around your needs.

  1. Day 1

    30-minute call

    Walk through a live audit on synthetic data and your questions.

  2. Week 1

    Mutual NDA, then one export

    Decisions or scores plus demographic categories. De-identified is fine.

  3. Week 2

    Your audit file

    Report, findings walkthrough, and the alternatives worth testing.