AI Bias in Hiring and the Interviews Nobody Audits

People tend to worry most about the risks they can measure. Usually that's a sensible habit. It goes wrong when the risk you can't measure is the bigger one, and I think that's what is happening with bias in hiring right now.

AI bias has become the regulatory topic of 2026. The EU AI Act's provisions on HR are pulling compliance teams into recruiting for the first time, and a lot of that attention is going to screening algorithms. The concern is fair. But I suspect most of the legal exposure in a typical hiring process sits somewhere else, in the unstructured interview, the "tell me about yourself" conversation that still runs most of the middle of the funnel.

Human judgment in hiring has been studied for decades, and the results aren't flattering. One well-known experiment sent out identical resumes under different names and found that white-sounding names got roughly 50% more callbacks than names associated with Black applicants. An often-quoted finding says interviewers form their impression in the first four minutes, before much real information about qualifications has come up. Research on accents has found that candidates with regional or foreign accents tend to be rated as less competent, whatever they actually say. None of this is new. It just barely comes up in the current compliance conversation.

Part of the reason is that it's hard to audit. You can take a screening model's decisions and run a disparate-impact analysis on them. You can't pull the reasoning out of a hiring manager's head and do the same. So regulators and compliance teams look where they're able to look. That makes sense, but being easy to inspect and being where the risk is are different things, and the harder-to-measure problem is often the larger one.

The worry about AI is real. If a model learned from past hiring decisions that were biased, it will repeat those decisions, and at scale. That has to be fixed. It can be fixed, though, because a model can be tested, adjusted, and monitored on an ongoing basis. There are established methods for auditing algorithmic systems for bias, and a growing set of rules that require it.

Interviews by people have no comparable correction loop. I suspect most companies do little formal calibration of their interviewers, and hiring managers get only light training in evaluating against set criteria. Suppose two managers interview the same candidate for the same role. It's easy to imagine them reaching opposite conclusions, simply because nobody ever wrote down what a good answer looks like. And nobody files an adverse impact report when a manager passes on someone for not seeming like a culture fit. Repeat that kind of judgment across thousands of interviews in a year and you can get demographic outcomes far outside what an audited system would produce.

The EEOC's technical assistance on AI and algorithmic decision-making tools framed the liability plainly: an employer is responsible for disparate impact whether it comes from an AI tool or from a process run by people. That applies in both directions. If your interviews produce skewed outcomes, and unstructured interviews often do, the fact that humans made the calls won't protect you. It only means you're less likely to notice until it shows up in a lawsuit.

Consistency is a large part of why Asendia works the way it does. It phones each applicant and runs a screening conversation built on criteria written down in advance for that specific role, which is exactly what the "tell me about yourself" interview lacks. Everyone who applies for a given role gets the same core questions, judged by the same rubric and scored the same way. There's no version of it that had a bad morning, or that asks different follow-ups because a candidate reminds it of someone it liked at a previous company.

What comes out of each screen is a written record: a summary of the conversation with the candidate's key statements, and an assessment against the criteria that were set in advance. Each decision point is logged and can be explained. If a candidate asks why they weren't moved forward, you can show which criteria were used and how their answers measured up. That record goes into the ATS alongside everything else, so there's nothing to reconstruct later. It's the kind of audit trail most companies can't produce for their human interviews, which is where most candidates are actually turned down. It also makes candidates easier to compare, since you no longer have one recruiter screening hard and another screening loose. (If you're wondering how this changes what you should measure, there's a separate post on what recruitment KPIs to track in a post-AI world.)

The bias in most hiring pipelines is older than any AI tool. What AI has done is put consistent, documented screening next to inconsistent, undocumented human judgment, so the difference is finally easy to see. I don't think the companies that come out of this period in the best shape will be the ones that kept AI out of hiring. They'll be the ones that used it to get the consistency interviewer training never quite delivered, and that can show their work.

A simple test: take the last ten candidates your hiring managers turned down and ask whether you could explain, in writing, why each one was rejected. If you can't, that's where I'd look first.

Ready to transform your hiring strategy? Schedule a Demo with our founders today!