A product manager at a mid sized company was sabotaging his own team's interviews. He gave thoughtful candidates weak scores. He stalled on feedback. He kept finding reasons to pass. A bias audit caught him. He was blocking hires that would compete with him for feature ownership.
That is one kind of bias. Most discussions of interview bias focus on the demographic kind. The political kind is just as common and gets less attention.
AI can reduce both, but only if you point it at the right thing. The wrong way to use AI for bias reduction is to scrub resumes of names and schools. That moves the bias one round down the funnel where it is harder to see. The right way is to compare interviewer feedback against interview evidence.
In one customer deployment, an AI layer caught a 14 percent discrepancy rate between what interviewers wrote in their feedback and what actually happened in the interview. A candidate would demonstrate strong system design and the writeup would say "weak technical depth." The evidence and the verdict did not match.
That is the bias signature. Not a missing word in a JD. A gap between what an interviewer saw and what they reported.
Where AI makes bias worse is in pure screening models trained on past hire decisions. If your last three years of hires were biased, the model learns the bias and scales it. This is well documented and still routinely ignored.
The rule is simple. Use AI to audit human judgment, not to replace it. Platforms like Mazle do the audit by comparing evidence to verdict, which is the only audit that actually works.
