AI Interviews Don't Score You the Way You Think They Do

What a human interviewer is actually reacting to

Start with the human baseline, because it's the standard most "sound confident" interview advice is quietly optimized for. In 2019, Yale researcher Michael Kraus and coauthors Brittany Torrez, Jun Won Park, and Fariba Ghayebi published a five-experiment study in the Proceedings of the National Academy of Sciences on how accurately people judge social class from speech alone.

The first four experiments established that people can identify a stranger's socioeconomic background from as little as seven spoken words. The fifth went further: 274 people with real hiring experience rated actual job candidates — competence, a starting salary figure, a signing bonus — based only on audio or text transcripts of short self-introductions, with no resumes and no visibility into how the candidates actually performed.

The candidates judged as sounding higher social class were rated more competent and offered more money, regardless of their actual qualifications for the job. Pronunciation carried more signal to the raters than what candidates actually said. It's a close cousin of the perception gap covered in what the HR round is actually scoring — evaluators reading confidence and fit off surface cues rather than off the content of an answer.

“Hiring managers rated competence and set salary figures off a few seconds of speech — before a single answer was actually evaluated.”

What actually changed when the interviewer was AI

In 2025, economists Brian Jabarian (University of Chicago Booth) and Luca Henkel (Erasmus University Rotterdam) ran a real field experiment, not a lab simulation: 70,000+ applicants for a customer-service role were randomly assigned to a screening interview with either a human recruiter or a voice-AI interviewer, at a real company making real hiring decisions.

The results ran in one direction. Candidates interviewed by the AI were 12% more likely to receive a job offer, 18% more likely to actually start the job, and 17% more likely to still be employed 30 days later, with 6% higher 30-day retention among those hired. Candidates also spoke differently in the AI interviews: more clearly, more directly, with fewer filler words and more job-relevant information per answer.

One finding matters most for the question this article is asking: self-reported gender discrimination was roughly half as common among candidates who went through the AI interview compared to the human one. That doesn't mean AI interviews are free of bias — see the honest caveats below — but it's real evidence the AI wasn't reacting to the same surface signals a human evaluator's gut read leans on. It behaves less like a snap judgment and more like a rubric, the same idea covered in how a rubric-scored interview round actually works.

“AI interviewers didn't reward the same things a human evaluator's gut reaction rewards — and the gap shows up in who gets hired.”

So what is an AI interview actually scoring

Neither study reveals the exact scoring model behind any single vendor's product — those are proprietary, and AI now shapes hiring well before an interview even happens, from resume screening to candidate sourcing. But the pattern across the research on interview scoring specifically points at the same few things: whether an answer is specific rather than vague, whether the reasoning holds together across the whole conversation rather than just within one answer, and whether a claim survives a follow-up question rather than just getting restated.

That last part is the one candidates most often get backwards. A generically rehearsed, polished-sounding story clears the bar with some human interviewers on delivery alone — practiced pacing, a confident pause, a tidy conclusion. An AI evaluator checking consistency across an entire conversation isn't grading whether an answer sounded rehearsed. It's checking whether the same story still holds up two follow-up questions later, which a memorized script often doesn't.

This is exactly the failure mode covered in why memorized STAR answers fall apart under an AI interviewer's follow-up questions — the framework itself isn't the problem; treating structure as a substitute for a real, specific, defensible answer is.

In practice, that makes an AI interview closer in spirit to a structured, rubric-scored technical round than to the free-flowing, rapport-driven conversation a lot of candidates picture when they hear "behavioral interview." The scoring criteria are more explicit and applied more consistently, because that's what a system built to run the same evaluation thousands of times over has to do to work at all.

AI interviews aren't automatically unbiased

It would be a mistake to read the Jabarian and Henkel result as "AI interviews are unbiased" — that's a bigger claim than the study makes. It measured one specific outcome (self-reported gender discrimination) in one specific context (a voice-AI screening interview for a customer-service role, at one company). It didn't measure race, disability, or accent-based bias against non-native speakers, or a dozen other dimensions bias can show up on.

AI hiring tools have their own well-documented history of getting this wrong in different ways. HireVue, one of the largest AI-interview vendors, discontinued the facial-analysis component of its scoring in 2021 after researchers and regulators raised exactly this concern: a model trained on past hiring patterns can encode the same biases those patterns contain, just less visibly than a person would. That's a different failure mode from the one the Kraus study documents in humans, not a rebuttal of it.

The honest summary: the research suggests AI interviews score differently than human ones, and in at least one real experiment that difference reduced one specific kind of self-reported bias. It doesn't mean every AI interview tool is fair, and it isn't a reason to stop preparing carefully.

What actually changes about how you should prepare

If an AI interview is scoring specificity, consistency, and whether your claims survive a follow-up question, the preparation that helps is different from, and in most cases more useful than, preparation aimed at sounding confident and warm.

• Answer with specific, checkable detail instead of a smooth, general summary — a number, a decision you actually made, a tradeoff you actually weighed.
• Expect your own earlier answers to get cross-referenced. If you called yourself highly collaborative in one answer, be ready to explain a moment where you made a unilateral call.
• Treat a framework like STAR or CARL as a way to organize a real answer, not a script to recite — the structure helps you; it isn't what's being scored on its own.
• Don't perform confidence you don't feel. A steady, specific answer delivered with a normal amount of hesitation reads fine to a system checking content, in a way it sometimes doesn't to a distracted human.
• Rehearse out loud, more than once, on real questions — not to sound smoother, but because the second and third pass through an answer is usually when the vague version turns into the specific, defensible one.

That last point is worth taking seriously: how many mock interviews it actually takes before an answer stops feeling rehearsed and starts feeling like your own has real numbers behind it, and the mechanism is the same one that matters here — repetition is what turns a general answer into a specific one, and specific is what an AI evaluator is actually built to notice.

If the format you're actually facing has no live interviewer at all — a recorded, one-way video interview — the same content-over-charisma scoring shift applies, with its own specific mechanics and a real regulatory right worth knowing about.

Key Takeaways

If you remember only a few things, remember these:

• A 2019 PNAS study (Kraus, Torrez, Park & Ghayebi) found human evaluators judged competence and set salary offers from just a few seconds of speech, based on perceived social class, before any answer was actually evaluated.
• A 2025 field experiment with 70,000+ real applicants (Jabarian & Henkel) found candidates interviewed by AI got more job offers, started and stayed in jobs more often, and reported about half as much gender discrimination as those interviewed by humans.
• The evidence points to AI interviews scoring specificity, internal consistency across the whole conversation, and how an answer holds up under follow-up questions — not vocal tone, accent, or polish.
• This isn't proof every AI interview tool is unbiased. It's evidence for one specific kind of bias reduction, in one real study, not a blanket guarantee.
• The preparation that actually helps is specific, checkable detail and answers that survive follow-up questions, not a more confident-sounding delivery.