Interview bias: where it enters and how to catch it

On 24 October 2026 we shipped a report that compares HR’s own decisions with the AI’s recommendations, job by job: for every candidate, did the recruiter and the AI agree, and where they did not, which way did they split? We built it because a customer’s HR team wanted to know how often the AI got it right.

The first thing a report like that teaches is that it cannot answer that question on its own. When a recruiter advances a candidate the AI marked On Hold, either the AI missed something or the recruiter saw something that was not in the interview. Sometimes what the recruiter saw was a real strength. Sometimes it was a familiar college, a confident voice or a candidate who reminded them of someone. The disagreement does not tell you who was right. It tells you exactly where to look, and that is most of what fighting interview bias involves in practice.

What is interview bias?

Interview bias is any influence on an interviewer’s judgment that has nothing to do with whether the candidate can do the job: a first impression, a shared background, the previous candidate, an accent, or one strong trait that colours every other score. It affects human and AI interviewers, and structure reduces it more than good intentions do.

Bias is usually not a decision to be unfair. It is the brain taking shortcuts when it has to judge a stranger quickly, which describes every screening interview ever run. The trouble is that the shortcuts are consistent. They push the same kinds of candidates up and the same kinds down, drive after drive.

Large studies show how big that can get. In a correspondence experiment covering more than 83,000 applications to 108 of the largest US employers, Kline, Rose and Walters found that applications with distinctively Black names were 2.1 percentage points less likely to hear back, and that a minority of firms accounted for much of the gap. That study looked at résumés, not interviews, but the lesson carries over: bias is measurable, it is concentrated, and it varies between teams in the same industry.

Nine types of interview bias, and what each looks like in a screen

Two of these do most of the damage in high-volume screening. First impressions, because a recruiter on their fortieth call of the day has every incentive to decide early and coast. Accent and speech, because so many of the jobs we screen for need clear communication, and “clear” quietly turns into “sounds like me”. We wrote about the second one at length in AI screening and how candidates talk.

Why awareness training does not fix it on its own

The default response to interview bias is a training session: here are the biases, now notice them. Interviewers leave the session agreeing with every slide, and their scores change very little. Knowing that the halo effect exists does not stop it while you are listening to a confident, articulate candidate answer a question vaguely.

What works is changing the conditions the judgment is made in. Structured interviews (identical job-related questions for every candidate, each answer rated against one agreed scale) come out as the most predictive selection method in the most cited recent review of the research, a 2022 meta-analysis by Paul Sackett and colleagues summarised by HumRRO. Part of the reason they predict better is that they leave less room for everything else. Google’s guide to structured interviewing puts the method in one line: use the same questions and grade answers on an identical scale.

Training is still useful as the reason for the structure. People follow a scoring rule more faithfully when they understand what it protects against.

Seven structural fixes that reduce interview bias

1. Same questions, same order, for every candidate in the role. The fastest way bias gets in is through questions that differ: a friendly chat with one candidate, a grilling for the next. Fix the core questions per role and allow only planned follow-ups.
2. Write the anchors before the first interview. For each question, write what a 3, a 2, a 1 and a 0 answer contains. Anchors written after the first ten interviews describe the first ten candidates. Our behavioral interview questions guide has a 0-3 example.
3. Score each answer when it ends. Not at the end of the interview, and never at the end of the day. A score written two hours later is mostly memory, and memory is where the halo lives.
4. Separate what was said from how it was said. If the job needs spoken English, score communication as its own item with its own anchor. Do not let it leak into the score for problem solving.
5. Score independently, then discuss. In a panel or a debrief, everyone records their scores before anyone speaks. The most senior person talking first sets the number for the room.
6. Hide what the interviewer does not need. Photo, age, address and college name rarely change what a screening question can show. Where you can, keep them off the screen until the scores are in.
7. Audit decisions, not intentions. Every month, look at who advanced and who did not, by role and by interviewer. Look for one interviewer whose scores run far above or below the others, and for groups who pass the résumé stage and then fall away at the interview.

None of these needs software. All of them are easier to keep up at volume when the structure is built into the tool rather than into a policy document nobody opens during a drive.

Bias in AI interviews: what changes and what does not

An AI interviewer removes some sources of bias completely. It is not tired at 6 PM, it does not remember the previous candidate, and it does not warm to a candidate from its own college. It asks the hiring team’s questions every time.

It does not remove bias as a category, and anyone who says it does is selling something. An AI can be stricter on accents it heard less often in its training, can reward a polished delivery over a specific answer, and can carry whatever bias is in the rubric a person wrote for it. Our own pipeline has places where this needs watching. Interview questions come from the job and from the candidate’s own résumé, so the résumé-based questions differ from one candidate to the next by design; the job-based questions and the scoring anchors are the parts that stay fixed.

What we do about that is the same list as above, built in:

• Fixed questions and anchors per job. The hiring team’s questions and the scoring guide are saved with the job, so every candidate for that role is scored against the same levels.
• Evidence behind every score. Every interview closes with a recommendation and a score for each question that quotes what the candidate said, next to the recording and transcript. A recruiter can check whether a low score came from the answer or from the way it was spoken.
• No guessing on deal-breakers. A deal-breaker answer counts only if it matches what the candidate actually said; otherwise the candidate goes to On Hold for a person to check.
• People decide. The AI recommends; no candidate is rejected automatically unless HR switches on auto-decide or sets deal-breaker rules for the job. The AI’s verdict and the recruiter’s decision are kept as separate records, so neither overwrites the other, and what happened after the decision can be recorded too, as we described in screener vs reality.
• Disagreement you can see. The 24 October report puts HR’s decisions and the AI’s recommendations side by side, job by job, which is where both kinds of bias show up first.

When you find a gap: what to do next

Sooner or later a check like the one below turns up something: one interviewer who passes half as many candidates as the others, or candidates from smaller towns clearing the résumé stage and then falling away at the interview. The instinct is either to ignore it (“the numbers are small”) or to overcorrect (“lower the bar for that group”). Neither fixes anything.

Work through it in this order instead:

1. Read before you decide. Pull ten interviews from the group that is falling away and ten from the group that is not, at similar résumé scores. Read the transcripts or listen to the recordings with the scorecard open. You are looking for the moment the scores diverge, and whether the answers justify it.
2. Find the question, not the person. Gaps usually come from one question or one anchor. A question that rewards a particular kind of previous employer, or an anchor that says “articulate” where it should say “specific”, will produce a gap for every interviewer who uses it.
3. Fix the instrument. Rewrite the question or the anchor so it describes the answer the job needs. Then re-score a sample of past interviews with the new anchor and see whether the gap closes. If it does not, the problem is elsewhere.
4. Talk to the interviewer with the outlier scores with their transcripts in front of you, not their pass rate. “On these three answers your scores were two points below the scorecard; walk me through what you heard” is a coaching conversation. “Your pass rate is low” is an accusation.
5. Write it down. Date, what you found, what you changed. If anyone ever asks how you check your hiring for fairness, that log is the answer, and it is far more convincing than a policy.

Do not lower the bar for one group, and do not add points for a background. That swaps one bias for another and leaves the broken question in place for the next drive.

A 15-minute bias check for your next drive

You do not need a study to start. Before the next drive, take one role and fifteen minutes:

1. Read the questions. Is every one about the job? Would you ask it, word for word, of every candidate? Delete any that are not.
2. Read the anchors. Does each level describe the answer, or the person? “Confident, polished speaker” describes a person. “Gives a specific example with their own action and a result” describes an answer.
3. Pull last month’s decisions. Sort by interviewer. Is one interviewer’s pass rate far above or below the rest on the same role?
4. Compare the stages. Of the candidates who cleared the résumé stage, which groups (college tier, city, gender) cleared the interview at a much lower rate? A gap is not proof of bias, but it is the place to start reading transcripts.
5. Read five disagreements. Five candidates where the interviewer’s decision and the scorecard pointed in different directions. For each, write one sentence: what did the decision rest on that the scores did not? If the answer is “the candidate seemed right”, that is the bias, written down.

Repeat it every drive. The numbers change slowly; the habit of reading disagreements changes how a team interviews within a month.

If you want to see what per-question evidence looks like before you trust a single score, take the live demo interview and read the scorecard it sends you. Every line in it quotes something you said.