New York City requires an automated employment decision tool to have had a bias audit within a year of use, with the results publicly available.
Why it matters when the plan changes
An audit answers the question a procedure's designers cannot answer about themselves: whether the tool, applied consistently, produces unequal outcomes. Because adverse impact is measured on results rather than intent, only measurement finds it, typically against a threshold such as the four-fifths rule set out in the Uniform Guidelines on Employee Selection Procedures. Jurisdictions are beginning to require the measurement, as New York City now does for automated employment decision tools, and employers are beginning to ask suppliers for it before the law does.
The tension is that an audit tests outcomes, not validity. A tool can pass a bias audit and still measure nothing useful, or show disparities and still be job-related and defensible. The field's own standard, the Standards for Educational and Psychological Testing, treats reliability, validity for the specific use, and fairness testing as three separate requirements, not one. The audit is one half of the evidence; a supplier that offers it without the other has answered a different question.
In practice
An employer receives a bias audit showing no material disparity by sex or ethnicity and treats the tool as safe. Nobody asks whether the tool predicts anything about performance in the role. Two years later the question is whether decisions rested on an instrument that was fair and irrelevant, which is a harder position than either fair or relevant alone.
Evidence
New York City requires that an automated employment decision tool has had a bias audit within a year of use and that information about it is publicly available.
New York City Department of Consumer and Worker Protection, Automated Employment Decision Tools (2023)A selection rate below four-fifths of the highest group's rate is generally regarded as evidence of adverse impact.
Uniform Guidelines on Employee Selection Procedures, 29 CFR Part 1607 (1978)
What it cannot tell you
A bias audit tests outcomes on a defined population at a defined moment. It is silent on whether the tool measures anything job-related, and a passing result on last year's applicants does not describe this year's. Absence of measured disparity is not evidence that a tool is valid or that its scores predict anything useful.
Questions
Selection or scoring rates by protected group, and the ratio between them, usually against a threshold such as four-fifths of the highest group's rate. Some audits also examine intersections of groups. It is a measurement of outcomes produced by the tool on a defined population over a defined period.
New York City requires one for automated employment decision tools, published and repeated annually. Elsewhere the requirement is emerging through the EU AI Act's data governance and bias examination duties and through general discrimination law, which makes the measurement prudent before any statute names it.
No. An audit shows outcomes are not materially unequal by group, often judged against the four-fifths threshold set out in the 1978 Uniform Guidelines on Employee Selection Procedures. Validity shows the tool predicts what the job requires. A tool can pass one test and fail the other, and both are produced by different methods.
An independent party with no stake in the result, using data from the actual deployed population where possible. A supplier's internal analysis is useful and is not an audit. The independence and the population are what make the result something a regulator or a works council will accept.
At least annually where a statute requires it, such as New York City's Automated Employment Decision Tools rule from 2023, and whenever the tool, the population or the use changes materially. A tool that showed no disparity on one year's applicants can show one on the next; a single audit at launch describes only that moment.