AI Bias Audit

An AI bias audit is an independent statistical evaluation of an automated employment decision tool to determine whether it produces disproportionately different outcomes across race, ethnicity, or sex categories. It's the compliance mechanism behind laws like NYC Local Law 144, and it's rapidly becoming a baseline expectation wherever employment AI is regulated.

What Does a Bias Audit Actually Measure?

Most bias audits calculate a selection rate for each demographic group the tool screens, then compare those rates using a version of the adverse impact ratio, often anchored to the EEOC's long-standing four-fifths rule. If one group is selected at a rate below 80% of the highest-selected group's rate, the tool shows adverse impact for that category, regardless of whether the outcome was intended.

The audit also has to be conducted by an independent party — the vendor auditing its own tool doesn't satisfy most legal definitions of independence.

Who Is Required to Get One?

NYC Local Law 144 requires a bias audit for any AEDT used on NYC-based candidates or employees. Illinois's amended Human Rights Act and several other pending state frameworks apply similar audit-adjacent obligations to AI used in employment decisions. Employers operating across states increasingly run one audit standard that satisfies the strictest applicable jurisdiction, rather than maintaining separate audits per state.

How Often Should an Audit Happen?

Annually at minimum, and immediately after any material change to the tool's model, training data, or scoring logic. A tool audited once at launch but never revisited can drift into bias as real-world hiring data feeds back into it — a phenomenon distinct from, but related to, general AI model drift.

What Happens if a Tool Fails Its Audit?

A failed audit doesn't automatically mean the tool is illegal, but it does mean the employer has documented notice of a problem. Continuing to use the tool without adjustment — reweighting inputs, adding human review, or discontinuing use — significantly increases legal exposure under disparate impact theory, per EEOC guidance.

How Should HR Teams Prepare for a Bias Audit?

Preparation starts with data access: auditors need demographic outcome data broken down by the specific stage the AEDT touches, not just overall hiring statistics. HR should also maintain a running log of which tools are in use, when each was last audited, and where results are published, since Local Law 144 and similar laws require public disclosure of a summary, not just internal recordkeeping.

Why Does This Matter Beyond Legal Compliance?

Bias audits are one of the few ways an organization can catch a hiring algorithm quietly narrowing its talent pool before it becomes a pattern. Employers that treat the audit as a genuine diagnostic — not a checkbox — tend to catch and fix problems long before a regulator or plaintiff's attorney does.

Discover how our HR solutions streamline onboarding, boost employee engagement, and simplify HR management

Request a Demo

Frequently Asked Questions

Q: Is a bias audit the same as an EEO-1 report?
A: No. EEO-1 reports overall workforce demographics; a bias audit specifically measures a tool's selection rates by group at the point it screens candidates.

Q: Can a company audit its own AI tool internally?
A: Most legal definitions require an independent auditor with no role in developing the tool, so a purely internal audit usually won't satisfy the requirement.

Q: What's the minimum sample size for a valid audit?
A: There's no universal number, but auditors generally need enough historical selection data per category to produce statistically meaningful rates, not a single hiring cycle.

Q: Do bias audits cover disability or age discrimination?
A: Most current laws focus on race, ethnicity, and sex; disability and age bias in AI tools are typically addressed separately under existing ADA and ADEA frameworks.

Q: Where do audit results have to be published?
A: Under NYC Local Law 144, a summary must be posted publicly, often on the employer's or vendor's website, before the tool is used on covered candidates.

Q: Does a passed audit protect an employer from all liability?
A: No. It reduces risk and demonstrates good faith, but a passed audit doesn't eliminate liability if the tool is later shown to produce discriminatory outcomes in practice.