Bias audit
An assessment of whether a hiring tool produces different selection rates across protected groups, performed by a party independent of the tool’s vendor.
Last reviewed
In practice
Independence is the load-bearing word. A vendor testing its own tool and publishing a favourable result has produced marketing, not an audit, and regulations that require audits generally say so explicitly.
A standard audit computes selection rates by sex and by race or ethnicity, and the intersection of the two, then reports the impact ratios. It describes outcomes; it does not certify that a tool is fair, because a tool can produce equal selection rates while measuring the wrong thing entirely.
The most common practical obstacle is data. An employer that never collected voluntary demographic data has nothing to audit, and assembling it retrospectively under time pressure is how audits end up superficial.
Not to be confused with
- A fairness guarantee
- An audit reports what happened to selection rates in a period. It says nothing about whether the tool measures something job-related, and a passing audit is not a defence against a claim that it does not.
Related
Adverse impact
A substantially different rate of selection for a protected group that results from a neutral-seeming employment practice, regardless of whether anyone intended it.
Automated employment decision tool
A computational process that substantially assists or replaces discretionary decision-making in hiring or promotion, typically by scoring, ranking or filtering candidates.
Selection rate
The proportion of candidates from a given group who pass a particular stage of a hiring process.
How CalHire handles this: AI governance, bias auditing and the audit trail