The interviewer is the bias

The first question a research-literate buyer asks about an AI interviewer is whether it biases what people say.

It is the right question. It is also aimed at the wrong interviewer.

People edit what they say to a person

Tourangeau and Yan reviewed three decades of survey research on sensitive questions for Psychological Bulletin. Their conclusion was that misreporting is common, situational, and motivated: people edit what they report to avoid embarrassing themselves in front of an interviewer or to avoid consequences from third parties.

That is not a technology finding. It is a finding about what happens when another person is listening.

The fix has been known for thirty years

Richman, Kiesler, Weisband and Drasgow pooled 61 studies and 673 effect sizes in the Journal of Applied Psychology. Two results matter here. Computer questionnaires and paper questionnaires produced almost no difference in social desirability distortion. Computerised interviews and face-to-face interviews did: less distortion without the person, and least of all when respondents were alone, anonymous, and could go back and change an answer.

Kreuter, Presser and Tourangeau then did the study most of this field cannot: they checked the answers. Recent university graduates were randomly assigned to a human telephone interviewer, an automated voice system, or a web form, and asked about things like low grades and academic warnings. The answers were compared to university records. Reporting of sensitive information and accuracy were highest on the web, lowest with the human interviewer, and the automated voice sat in between.

It is the belief that matters

Lucas, Gratch, King and Morency ran the cleanest test of the mechanism. Two hundred and thirty-nine people were interviewed by a virtual interviewer about their health. Half were told it was automated. Half were told a person was operating it. Separately, and without their knowledge, half actually were automated and half actually were operated by a person.

The people who believed no human was involved reported less fear of disclosing, managed their impression less, showed more sadness on their faces, and were rated by independent observers as more willing to disclose. Whether a person was actually behind the screen changed nothing except how usable they found the system.

So the honest sentence is this: an AI interviewer does not remove bias by being clever. It removes bias by not being a person the participant has to face afterward.

What about the questions it asks?

That is the second half of the objection, and it is fair. A leading question biases an answer no matter who asks it.

Chopra and Haaland had an AI conduct 381 qualitative interviews and coded every question it asked. About 95 percent were open-ended, relevant, and non-leading, and the interview data predicted what participants actually did eight months later.

Jabarian and Henkel ran the largest test so far: 70,000 job applicants randomly assigned to a human recruiter or an AI voice agent, with humans making every hiring decision. Applicants interviewed by the AI were 12 percent more likely to receive an offer and more likely to stay in the job. The transcripts show why. The AI spoke less, asked more consistently, and collected more decision-relevant information. Given a choice, 78 percent of applicants picked the AI.

Where the evidence stops

Three boundaries, because this is the study a sceptic will read closest.

The effect is about the perceived absence of a human judge, not about screens. Dodou and de Winter found no difference in social desirability across paper, offline and online surveys. If a participant believes a person will read their answer with their name attached, the benefit shrinks.

Kreuter's automated voice was a set of recorded prompts, not a conversation, so it is a floor for what a voice interviewer can do rather than a description of one. What carries over to a conversational agent is Lucas's mechanism: the participant's belief that no person is judging. Anonymity on request strengthens that belief.

Candour is not the same as accuracy. A machine that people speak to more freely still cannot verify what they say. That is what the human review and the verbatim quotes are for.

What Candidly AI does with this

Participants are told plainly that they are speaking to an AI. Calls can be fully anonymous on request. The questions are written with you in the Design phase and approved before launch, so what gets asked is on the record before anyone answers. Every conversation is read by a person, and the Ground Truth Report puts the verbatim quotes beside the themes, every participant quoted at least once.

The bias a human interviewer introduces is the participant's own editing. The research says take the person out of the room and the editing drops. That is not a claim about what our software can do. It is a claim about what people do, and it has held for thirty years.

An AI interviewer does not remove bias by being clever. It removes bias by not being a person the participant has to face afterward.

Do people disclose more to an AI interviewer than to a human?

The evidence says yes, when they believe no person is listening. Lucas, Gratch, King and Morency found that participants who believed a virtual interviewer was automated reported lower fear of disclosure and were rated by observers as more willing to disclose, and whether it was actually automated made no difference. Richman and colleagues' meta-analysis of 61 studies found less social desirability distortion in computerised interviews than face-to-face interviews.

What is social desirability bias?

The tendency to answer in the way that makes you look better to whoever is asking. Tourangeau and Yan describe it as a motivated process in which respondents edit what they report to avoid embarrassing themselves in front of an interviewer or to avoid consequences from third parties.

Does an automated voice interviewer reduce bias compared to a human on the phone?

Kreuter, Presser and Tourangeau randomly assigned people to a human telephone interviewer, an automated voice system, or a web form, and checked answers against university records. Reporting of sensitive information and accuracy were lowest with the human interviewer and highest on the web, with the automated voice in between.

Does an AI interviewer ask biased or leading questions?

Chopra and Haaland coded every question asked across 381 AI-led interviews and found about 95 percent were open-ended, relevant, and non-leading. Jabarian and Henkel's field experiment with 70,000 applicants found the AI interviewer more structured and consistent than human recruiters, and applicants interviewed by AI were 12 percent more likely to receive an offer.

What is Candidly AI?

Candidly AI is a Calgary-based Canadian platform that holds hundreds of one on one voice conversations with the people whose perspective a decision depends on. Participants call a dedicated number on their own schedule, the conversation adapts with follow-up questions, and the organisation receives the Ground Truth Report: themed findings backed by verbatim quotes, with every participant quoted at least once. Candidly AI is ISO 27001 Certified and SOC 2 Audited, runs on fully Canadian infrastructure with self-hosted models, and makes no third-party model API calls.