AI hiring bias audit: what it measures and how to run one

An AI hiring bias audit is a structured check of whether an automated hiring tool, such as a resume screener or an interview scoring model, produces different outcomes for candidates based on sex, race, ethnicity, age, disability, or other protected traits. In practice, the auditor compares how often each group gets advanced by the tool and flags gaps large enough to need an explanation. As the hiring manager, you probably won't run the statistics yourself. You do own the decisions the tool feeds into, so you need to know what a sound audit looks like and what to ask for.

What is an AI hiring bias audit?

It is a fairness test for a hiring step that software either makes or heavily influences. The auditor takes real outcomes, or carefully built test data, and asks a narrow question. Did this tool advance one group of candidates at a meaningfully lower rate than another?

A useful audit has four parts:

1. A clear definition of the tool and the exact decision it affects, such as "ranks applicants and the top 30 percent get a phone screen."
2. Outcome data broken out by demographic group, usually from voluntary self-identification.
3. A comparison of selection rates or average scores between groups.
4. A written summary of the results, including the audit's limits and what changed afterward.

Vendors often run their own audits. Your company can also run one, or hire an independent auditor. Some local laws now require an independent audit before an employer uses certain automated tools, along with notice to candidates. The details vary a lot by place, so check with your legal team about the rules where you hire.

Why should a hiring manager care?

In most places, the employer is responsible for discriminatory outcomes in its hiring process, even when a vendor built the tool. Pointing at the software is unlikely to get you off the hook. If a screener filters out older applicants or people with gaps in their work history, that lands on you.

There is also a plain business reason. A biased filter throws away qualified people before anyone looks at them. You end up with a smaller, less varied shortlist, and you never see what you missed. An audit is one of the few ways to find out who the tool rejected.

What does a bias audit actually measure?

Most audits start with selection rates. A selection rate is the share of candidates in a group who pass a step. If you want a refresher on rate math, our recruiting metrics glossary with worked formulas covers pass-through rates the same way.

The next number is the impact ratio. Take each group's selection rate and divide it by the rate of the group that was selected most often. In the US, a common rule of thumb is the four-fifths rule. If a group's ratio falls below 0.8, the gap may indicate adverse impact and is worth investigating. A ratio under 0.8 does not prove discrimination on its own, and a ratio above it does not prove the tool is fair.

Here is a worked example with made-up numbers for a resume screener:

Group B's ratio is 30 divided by 50, or 0.60. That is well under 0.8 and needs a closer look. Group C at 0.90 is less concerning, though a small group can swing a lot from one month to the next.

Tools that produce a score instead of pass or fail, such as interview scoring models, are often audited by comparing the share of each group scoring above the median, or by comparing average scores. Better audits also check intersectional groups, such as women of a particular race, because a gap can hide inside two averages that each look fine.

How is an AI hiring bias audit done, step by step?

You may never touch the spreadsheet. You should still understand each step well enough to ask good questions.

1. List every automated step

Write down every place software sorts, scores, ranks, or rejects candidates. Include knockout questions in your applicant tracking system, resume parsers, assessment scoring, interview analysis, and scheduling bots that drop people who do not respond. The list is usually longer than people expect.

2. Define the decision and the cutoff

An audit only means something for a specific decision. "Top 25 percent get an interview" and "score above 70 gets an interview" can produce different results from the same model.

3. Gather outcome data by group

Use voluntary self-identification data where you have it, and keep it separate from the hiring decision. Never ask hiring managers to guess a candidate's race or age. If data is missing for many candidates, the audit should say so.

4. Calculate rates and ratios

Compute selection rates and impact ratios for each group, and for intersections where sample sizes allow. Note the time period and the number of candidates.

5. Investigate any gap

A low ratio tells you where to look, not why the gap exists. Check which inputs drive the score. Common culprits include zip codes, school names, employment gaps, and speech or video features that track accent or disability.

6. Fix, retest, and record

Remove or reweight problem inputs, change the cutoff, or stop using the tool for that step. Then rerun the numbers and keep a written record of what you changed and why.

What should you ask your vendor about bias audits?

Vague claims like "our AI is unbiased" are a warning sign. Ask for specifics, in writing:

• Who ran the most recent audit, and were they independent of your company?
• Did it use real customer outcomes, or test data you built?
• Which decision and cutoff did it cover, and does that match how we will use the tool?
• Which groups and intersections were measured, and what were the impact ratios?
• What did you change after the last audit?
• Can we turn off automated scoring and use the tool only to organize information?
• Can we export our own outcome data to audit our use of the tool?
• How do candidates request an accommodation or an alternative process?

If you are running a formal buying process, add these to your video interview software RFP template so every vendor answers the same questions.

What can a bias audit not tell you?

An audit describes one period of time. It does not prove a tool will stay fair. Vendors retrain models, and the people applying to your jobs change over time. If someone moves a cutoff next quarter, that is a new decision and it needs its own check.

Audits also struggle with small numbers. If only a dozen candidates from a group applied, the ratio can look terrible or perfect by chance. A good report says when a sample is too small to support a conclusion.

Most audits measure outcomes, not whether the tool measures the job. A tool can pass a ratio test and still score candidates on things unrelated to the work. Ask what job-related evidence supports each scored factor.

How can you reduce AI hiring bias in your own process?

You control more than the vendor's model. These habits are within reach of any hiring team.

Score against a written rubric

Before anyone reviews candidates, define what a strong, adequate, and weak answer looks like for each question. Our guide to building an hourly candidate scorecard has a template you can adapt for most roles.

Treat AI scores as advice

Use AI output as one input that a trained person checks, not as the decision. In DeepHire, for example, AI summaries and scores run on the recruiter's own rubric and are advice only, so a person still decides who moves forward.

Ask everyone the same questions

Structured interviews with identical questions and time limits are easier to compare and easier to audit.

Write down why a reviewer overrides a score

When a reviewer advances someone the tool scored low, or rejects someone it scored high, record why. Those notes let you audit human decisions as well as the software's.

Offer accommodations up front

Tell candidates how to request extra time, a different format, or a human review. Our ADA hiring basics for recruiters explains what that looks like in practice.

A simple audit record you can copy

Keep one row per automated step and update it at least once a year, or whenever the tool or how you use it changes. The entries below are a made-up example.

This record is not legal advice, and on its own it will not satisfy every local requirement. Hand it to your legal team or an outside auditor as the starting point for the next review.