Bias audits for hiring AI

Prove your hiring AI treats people equally.

Isonomy audits AI screening and matching tools for adverse impact, finds where it comes from, and shows a fairer alternative that works as well. Every model release, not once a year.

Built for AI hiring vendors, employers and independent auditors. NYC Local Law 144-ready.

Impact ratio by groupSample audit · synthetic data
Resume screener · 12,000 applicants · dashed line = four-fifths rule
Single attributes
Women0.85
Black applicants0.66
Intersectional
White men1.00
White women0.81
Black men0.64
Black women0.56
Black women are selected at 56% of the top group's rate, a gap no single-attribute check shows. Main driver: a ZIP-region feature acting as a proxy for race.
Adverse impactMonitorReference
0.56

Lowest intersectional impact ratio found in the sample audit, well below the 0.80 four-fifths threshold.

7 → 0

Groups with statistically significant adverse impact, before and after the recommended alternative.

+6.4 pts

Higher precision for the fairer alternative: fewer unqualified people advanced, not more.

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

Product

More than a pass/fail on the four-fifths rule

A compliant table is the start. Isonomy also tells you who is affected, why, when it started, and what to change.

Impact ratios, done properly

Selection and scoring rates per category in the Local Law 144 format, with confidence intervals and significance tests so small-sample noise isn't reported as bias.

LL144 summary tables

Intersectional analysis

Every combination of attributes, such as race × sex. Finds the disparities that hide when each attribute is checked on its own.

Hidden-disparity alerts

Proxy detection

Shows which inputs stand in for protected attributes, like ZIP code for race or employment gaps for sex, and whether your features jointly encode them.

Feature-level evidence

Explanation fairness

Checks whether the model reasons differently for different groups, and which feature costs each group the most score points.

Procedural fairness

Drift monitoring

Tracks impact ratios over time and flags when a model release or a change in applicants makes fairness worse.

Audit every release

Less discriminatory alternatives

Retrains candidate models and finds an option with less adverse impact and the same or better job-related validity. No group-specific cut-offs, ever.

Mitigation lab

How it works

From a data export to audit evidence in a day

Export your decisions

One row per person assessed: the tool's decision or score, demographic categories, and optionally the model's inputs, a date and a job-related outcome. De-identified data is fine.

Isonomy runs the audit

Impact ratios, intersectional and proxy analysis, explanation fairness, drift and the alternative search run together. No integration needed to start.

Get the report, then keep watching

A report your auditor, counsel and customers can read, plus re-audits on every new batch of decisions or model release.

Mitigation lab

Don't just find bias. Show the fairer option.

Regulators and courts ask whether a less discriminatory alternative existed. Isonomy answers with evidence.

  • Tests dropping proxy features, reweighing and concept nullspace projection
  • Every candidate selects the same share of people as today
  • Ranked on significant adverse impact, worst impact ratio and validity
  • Never uses protected attributes or group-specific cut-offs at decision time
Current toolRecommended alternative
Worst impact ratio0.560.86
Groups with significant adverse impact70
Precision (selected people who succeed)79.2%85.6%
Change—Remove the employment-gap feature
Sample audit on synthetic data. Validity measured against a job-related outcome.

Regulations

One audit, many obligations

Rules on automated hiring decisions are multiplying. Isonomy produces the evidence they have in common: measured impact, documented testing and ongoing monitoring.

NYC Local Law 144

Annual independent bias audit and published impact ratios for automated employment decision tools.

Colorado AI Act

Reasonable care against algorithmic discrimination for high-risk AI, including employment.

Illinois HB 3773

Use of AI with a discriminatory effect in employment decisions falls under the Human Rights Act.

California ADS rules

Anti-discrimination regulations cover automated decision systems; anti-bias testing is relevant evidence.

EU AI Act

Recruitment AI is high-risk: risk management, data governance and bias monitoring obligations.

Title VII

Disparate-impact law asks whether a less discriminatory alternative was available.

Summaries are simplified for orientation. Isonomy provides decision-support software and analysis, not legal advice. Local Law 144 audits must be performed by an independent auditor; Isonomy can support that auditor.

Founder

"Most bias audits stop at one attribute and one number. The harm usually sits at the intersections and in how a model reasons. That's what my research measures, and what Isonomy is built to catch."

Gideon Popoola, Ph.D.Founder, Isonomy

Gideon's Ph.D. in Computer Science (Montana State University) focused on intersectional, procedural and explanation fairness in machine learning, with peer-reviewed work at ECML and IEEE venues. He has built production machine-learning systems in data science roles at Walmart, Riot Games and JPMorgan Chase.

Ph.D., Computer ScienceECML 2026IEEE CAI 2026WalmartRiot GamesJPMorgan Chase

Design partners

We're working closely with three design partners

If you build or use AI to screen candidates, we'll run Isonomy on your data and shape the product around your audit needs.

  • Full audit and monitoring at no cost during the partnership
  • Direct input on the roadmap
  • Founding-customer pricing afterwards
  • Mutual NDA before any data is shared

We use these details only to reply to you. See our privacy notice.