How an algorithm audit finds unfairness
An algorithm auditor does not usually decide whether a system is fair by reading its code alone. The key question is what happens when the system makes decisions about people—and whether those outcomes differ in ways that cannot be justified by the task.
An auditor might compare how often a résumé-screening tool advances applicants from different groups, or examine whether a risk score leads to different rates of searches or arrests. That requires tracing the system’s inputs, outputs, and role in the larger process. A model can appear neutral while relying on information—such as a past address or employment history—that reflects unequal opportunities.
The comparison depends on the decision
There is no single fairness test that fits every use. For a hiring tool, auditors may ask whether qualified applicants are wrongly filtered out at different rates. For a system estimating risk, they may compare the accuracy of its predictions across groups. Those goals can conflict: improving one measure of fairness can worsen another, especially when groups have different historical data or the system makes different kinds of errors.
So an audit begins by making the decision explicit: who is affected, what counts as a mistake, and which differences matter. Auditors can then test records, inspect how the model was built, and compare results under different conditions. If the data are incomplete or the system’s decision process is opaque, the limits of the audit become part of the finding.
A test is not the same as a fix
When an audit uncovers a problem, possible responses include changing the training data, adjusting the model, or altering how people use its recommendations. But a technical change cannot settle every question. A company may need to decide which errors it is willing to reduce, while people affected by the system may have little control over that choice.
Auditing can expose risks before they become public failures, but it matters most when someone can act on the results. A report that identifies unequal outcomes without changing the tool or its use leaves the underlying decision untouched.
