How do I audit an AI screening tool for bias before we use it?

Specific questions with written answers. What testing has been performed on the screening or ranking feature, by whom, and how recently. What data was it evaluated against, and does that data resemble your hiring. What attributes does the system use, and are proxies for protected characteristics excluded, such as names, photographs, addresses, graduation years and institution names. How does the vendor monitor for drift after deployment. What documentation can they provide to support your own compliance obligations. Some jurisdictions now require an independent audit of automated employment decision tools, so ask whether one exists and whether the report is available. Note the difference between a vendor stating that their system is fair and a vendor showing you what they tested and what they found.

More than most teams expect. Take a historical role where you know the outcomes and check whether the tool's ranking correlates with candidates who performed well through your process. Then vary inputs deliberately: submit the same experience with different name formats, different education institutions, different career gap patterns, and see whether the ranking moves in ways that have nothing to do with capability. This kind of probing is not a formal audit, and it regularly finds behaviour worth questioning. Where you lawfully hold demographic data for equal opportunity monitoring, compare selection rates across groups before and after the tool is introduced, following your counsel's guidance on how that analysis may be conducted. Keep the results, because documented testing is useful evidence of diligence.

Keep a human decision point on any outcome that removes a candidate from consideration, and record that the review happened rather than assuming it. Configure the system to rank rather than reject where you have the option. Review the shortlist composition periodically instead of only at launch, since drift appears over time as your applicant mix changes. Limit the attributes the system can see where configuration allows, particularly anything that acts as a proxy for a protected characteristic. Give one person ownership of monitoring, with a defined cadence. And keep the ability to turn the feature off quickly, which sounds obvious and is worth verifying, because some deployments are difficult to reverse once workflows depend on them. These controls also support more consistent screening practice generally.

Regulation in this area is moving and varies by jurisdiction. Some places require notice to candidates that an automated tool is in use, some require independent bias auditing of such tools, and general employment discrimination law applies everywhere regardless of whether a decision was made by a person or a system. The employer usually carries the liability, not the software vendor, which is the point most buyers underestimate. Involve counsel before deployment rather than after, document what you tested and why you considered the tool appropriate, and revisit that assessment when the vendor changes the feature or when you begin hiring in a new jurisdiction. Treat vendor assurances as input to your assessment, not as a substitute for it.