Blog
Latest
“We've been audited” is not an answer. It's the start of a question you haven't finished asking yet.
Ask almost any AI hiring vendor whether their tool has been audited for bias, and the answer is yes. Ask a second question — audited by whom, using what standard, checking for what specifically — and the confidence can thin out fast. Here's what a real audit actually checks, so you know what you're asking for.
Independent, Not Self-Certified
The first and most basic requirement: the audit has to be conducted by a party independent of the vendor.
A vendor's internal QA team testing its own product is not an audit in any meaningful sense — it's quality control, and quality control has an obvious incentive problem when the product in question is the auditor's own paycheck.
Ask specifically who performed the audit and what their relationship to the vendor is.
The Four-Fifths Rule, in Plain Terms
The most common technical standard for detecting bias is the “four-fifths rule,” established by the EEOC in 1978.
It works like this: take the selection rate for the group selected most often — say, 60% of one group advances past screening.
Any other group selected at less than 80% of that rate (in this case, below 48%) is flagged as evidence of adverse impact requiring further review. It's not a perfect test, but it's the baseline test, and it predates AI hiring tools by nearly fifty years. A vendor unfamiliar with it hasn't done basic homework.
Aggregate Scores Hide More Than They Reveal
A vendor reporting “no bias detected” as a single number is reporting the least useful version of an audit.
A tool can look clean in aggregate while still producing a real gap for one specific subgroup, or at the intersection of two — race and gender together, for instance, rather than either alone.
A real audit reports selection rates broken out by group, not collapsed into one reassuring headline figure.
An aggregate “no bias detected” claim can hide a real gap for a specific group. Ask for the breakdown, not the summary.
What to Specifically Ask For
Four concrete requests get you most of the way there: the name of the independent auditor, the specific groups tested (race, sex, ethnicity, and intersectional categories — not just one dimension), the date of the most recent audit, and whether results are published somewhere you can verify without asking permission. If a vendor treats any of these as an unusual request, that's useful information in itself.
—
Part of the Transparent AI in Hiring Series. For the full vendor evaluation framework, download the Transparent AI Vendor Evaluation Scorecard at getclara.io.
—
Sources
U.S. Equal Employment Opportunity Commission. Uniform Guidelines on Employee Selection Procedures (1978).
New York City Department of Consumer and Worker Protection. Automated Employment Decision Tools (Local Law 144).
About CLARA
CLARA is a skills-based hiring platform built for mid-market companies. We measure critical thinking, learning agility, and Distance Traveled—the validated competencies that predict performance, not pedigree. Filter great talent in, not out. Learn more at getclara.io.