Skip to content

Stockertim Dictionary · Human resources · Technology

Bias audit

Also called fairness audit, algorithmic audit, bias assessment

In short

An internal or independent test of whether a tool's outcomes differ across groups, usually by calculating selection rates and impact ratios.

  • Advanced
  • Process
  • Checked

In detail

A bias audit asks a concrete question: does this tool, in this use, treat groups differently? The core work is quantitative. The auditor gathers the tool's real or test outcomes, links them to group data such as sex, ethnicity, age or disability, and calculates selection rates and impact ratios at each stage. Good audits also test intersections, such as older women, check statistical significance and look at why gaps arise.

An audit is more than a number. It reviews what the tool measures and whether that is job-related, how the training data and target outcome were chosen, whether reasonable adjustments are possible for disabled candidates, how humans use the scores and what happens when a gap is found. NIST's guidance on bias stresses that bias comes from data, from the way a system is designed and from how people use it, so an audit looks at all three.

Some local laws in the United States require independent bias audits of automated employment decision tools before use, with a public summary. Elsewhere audits are good practice and support duties under equality and data protection law and, for high-risk systems in the European Union, the AI Act's data governance and monitoring duties.

Worked example

An auditor tests a made-up screening tool on 1,000 past applications: 500 from men, 500 from women.

  1. Men passed: 250 / 500 = 0.5
  2. Women passed: 200 / 500 = 0.4
  3. Impact ratio: 0.4 / 0.5 = 0.8

The ratio sits exactly at the four-fifths line, so the auditor tests subgroups and significance before concluding.

In practice

Buyers ask vendors for recent audit reports, run their own checks on their own applicant data after launch, and repeat audits after model updates or a change in how the tool is used.

Watch out

  • An audit on the vendor's data says little about outcomes in your applicant pool.
  • A one-off audit ages: models and applicant pools drift.

Learn it properly

Official sources

Facts last checked on 3 November 2026.

Education, not advice: the Stockertim Dictionary explains what terms mean. It never recommends buying or selling anything.

English