Latest / AI Ethics with Fexingo: Bias, Safety, and Responsible Artificial Intelligence / When Your AI Job Interviewer Has a Hidden Agenda
Transcript
- Lucas: If you've ever done a video job interview where an AI analyzed your facial expressions and tone of voice, you might assume it's just reading how well you communicate. But a study published in February 2026 by researchers at the University of Toronto tested five commercial AI interview platforms and found that four of them were inferring proxy measures for race and gender behind the scenes. Luna: Wait, how do you infer race or gender from a video interview without explicitly asking? Lucas: That's the alarming part. The researchers used headshots of actors from diverse backgrounds—controlled for lighting, angle, expression—and ran them through systems like HireVue and Retorio. The AI didn't just score their 'employability.' It consistently assigned lower scores to women and people of color, even when the actors delivered identical responses. Luna: So the bias isn't in the words, it's baked into the visual and audio analysis layer. The algorithm is essentially looking at skin tone or facial structure and making a hiring judgment. Lucas: Exactly. And these systems are being used by some of the largest employers in the world. HireVue alone claims to have processed over 15 million interviews. The vendors say they test for bias, but the study showed that even after 'de-biasing,' the scores still correlated with race and gender at statistically significant levels. Luna: Let me play that back: vendors claim they've removed bias, and independent researchers still find it. That's a transparency gap the size of a canyon. Lucas: Right. And the study didn't stop at just the big players. They tested a smaller platform called Retorio, which markets itself as 'ai powered personality assessment.' Retorio's model inferred not just conscientiousness and emotional stability, but also sexual orientation and political affiliation from a 30-second video clip. Luna: That feels like an enormous privacy violation. If I'm applying for a job, I'm not consenting to having my political leanings guessed from my voice tone. Lucas: You're not consenting, and until very recently, you weren't even told it was happening. The study found that none of the platforms' privacy policies disclosed that they inferred protected characteristics or personality traits beyond what was stated. That's a fundamental informed consent problem. Luna: Before we go deeper on consent, let's talk about the real-world impact. Did the researchers find evidence that actual job candidates were rejected based on these inferred traits? Lucas: They did. One case study in the paper: a university applicant in Canada used a platform called VidCruiter for a scholarship interview. The system flagged her as 'low conscientiousness' because she paused for two seconds before answering a question about teamwork. She was rejected. The researchers later interviewed her and found she was simply considering the question carefully. The AI had no context to distinguish thoughtfulness from hesitation. Luna: And that pause—in any human interview, a hiring manager would see that as reflective, not negative. The AI doesn't have that human judgment. Lucas: Which brings us to the regulatory landscape. In March 2026, the Equal Employment Opportunity Commission issued new guidance on AI hiring tools. It says employers must conduct an independent audit of any AI system that screens candidates, and they must provide applicants with a clear explanation of what the AI measures. Luna: But here's the catch: the EEOC guidance doesn't have the force of law. It's a 'best practices' document. There's no penalty for non-compliance unless the AI system produces a disparate impact that violates existing anti-discrimination law. Lucas: And proving disparate impact requires extensive data and a lawsuit. Most rejected candidates never know the AI made the call, let alone how it made it. That's the core of the problem: you can't challenge a decision you can't see. Luna: So what would real accountability look like? Should these platforms be required to open up their black boxes? Lucas: The researchers argue for three things. First, full disclosure: every candidate should receive a report showing exactly what traits were measured and how they were scored. Second, a right to human review: any automated rejection should be appealable to a person. And third, a ban on inferring any trait that isn't directly job-relevant—so no personality guesses, no emotion detection, no political leaning. Luna: A couple of dollars a month is genuinely what keeps these going—buy me a coffee dot com slash fexingo, if you've gotten something out of them. Lucas: Yeah, honestly, listener support is what lets us stay independent and dig into stories like this. If you've found the show useful, it really makes a difference. Luna: Back to the research—the study also found that the AI systems trained on Western data performed worse on candidates from non-Western backgrounds. Accent, speech rhythm, even eye contact norms varied enough to skew scores. Lucas: That's a huge fairness issue in a globalized job market. A candidate from Nigeria applying to a U.S. company might be penalized for not looking directly at the camera, which is considered respectful in their culture. Luna: And the platforms aren't exactly rushing to fix this. Retorio's CEO responded to the study saying their model is 'trained to predict job performance, not demographics,' but the researchers showed that the demographic correlations were statistically inseparable from the performance predictions. Lucas: That's the key technical problem: if the model learns that certain speech patterns or facial features correlate with success in past hires, and those past hires were predominantly white and male, then the model will encode that demographic correlation as a 'job performance' signal. It's a textbook case of algorithmic redlining. Luna: And it's not just hiring. These same AI interview platforms are being used for internal promotions, performance reviews, and even customer service evaluations. The bias scales across the entire employee lifecycle. Lucas: One company in the study, an unnamed Fortune 500 retailer, used an AI interview tool to screen for customer-facing roles. The AI consistently gave lower scores to candidates who spoke with a Southern U.S. accent—dinging them on 'professionalism.' That's a proxy for region and class, not job ability. Luna: So what can a job applicant do right now to protect themselves? Is there any way to game the system or request an alternative? Lucas: Some advocates suggest applying with a neutral background, minimal makeup, and deliberate pauses to avoid triggering emotional inference. But that's a terrible workaround. The real solution is legal protection. A few states, including Illinois and Maryland, have passed laws requiring employers to disclose AI use in hiring and to offer an alternative human interview. Luna: But most states haven't. And even where laws exist, enforcement is weak. The Illinois law has been in effect since 2020, and as of 2025, fewer than 50 complaints had been filed. Lucas: Part of that is because applicants don't know their rights. The disclosure requirements are often buried in fine print. And if you don't know the AI made a decision, you can't challenge it. The whole system relies on transparency that simply isn't there. Luna: So we have a technology that is opaque, biased, and widely deployed, with regulation that's mostly advisory and lightly enforced. That feels like a recipe for systemic discrimination. Lucas: It is. And the stakes are only rising. With remote work becoming permanent for many industries, video interviews are now standard. The AI layer is the new gatekeeper, and most job seekers don't even know they're being judged by a machine. Luna: I want to end on a question: if you're an employer listening to this, what should you do? Because many of them genuinely want to hire fairly. Lucas: The researchers recommend three steps. First, conduct your own audit—don't just trust the vendor's bias report. Second, require the vendor to provide a full list of every feature the model uses, including any inferred traits. Third, offer every candidate a human alternative. If your AI is truly fair, a human interview should confirm the same result. If it doesn't, you have a problem. Luna: That's a practical test. And it puts the burden back on employers to verify, not just assume. Lucas: Exactly. Because hiring is a human decision, even when a machine makes the first cut. The question is whether we're building better gates or just faster ones.