How to reduce bias in candidate screening

Hiring teams reduce bias in first-round screening by fixing the criteria before reading any application, asking every candidate the same job-related questions, scoring each answer against written benchmarks before discussing it, and checking selection rates afterwards. Blind resume review helped in one famous study and backfired in another. AI can make screening more consistent, or repeat old bias at scale.

Where does bias get into first-round screening?

At every step where a person, or a model trained on people's past choices, is left to judge without a written standard. The first round has five of them.

Swipe the table sideways to see every column.

Step	How bias gets in	What reduces it
Writing the requirements	Proxies such as a university, a past employer or "culture fit" narrow the pool to people like the last hire	Must-haves written as things a person can do, agreed before the search
Reading resumes	Signals unrelated to the job, starting with the name	Criteria fixed in advance, and screening on facts, not polish
The conversation	Each candidate gets a different interview, and rapport decides	The same job-related questions for everyone
Scoring	An overall impression, formed early, colours every answer	Each answer scored against a benchmark, independently
Deciding	The most confident voice in the debrief, and overrides nobody records	Scores written down before discussion, and reasons for any override

The resume step is the best measured. In Bertrand and Mullainathan's field experiment, otherwise identical resumes sent to real job adverts drew 50 percent more callbacks when the name at the top sounded white. The interview step is where similarity takes over: in Lauren Rivera's study of hiring at elite professional firms, published in the American Sociological Review, more than half of the evaluators ranked cultural fit, meaning perceived similarity to existing staff in leisure pursuits, background and self-presentation, as the most important criterion at the interview stage.

The requirements step is covered in the recruiter intake meeting. The rest of this page takes the other four in order.

Why do structured questions reduce bias?

Because they take away the discretion that bias needs. The US Office of Personnel Management describes a structured interview as one that asks all applicants "the same exact set of pre-defined lead and probe (i.e., follow-up questions)", scored "according to benchmarks of proficiency", with questions based on the competencies the job needs. OPM's summary of the research is direct: "Interviews with higher degrees of structure show higher levels of validity, rater reliability, rater agreement, and less adverse impact", and "generally little or no performance differences are found between men/women or applicants of different races".

In practice that means four things for a first round:

The same questions, in the same order, for everyone applying to the role.
Follow-ups decided in advance. Probing is fine; probing only the candidates you warmed to is not.
Questions about the job, from the must-haves, not about hobbies, family or where somebody grew up.
Notes of what was said, not of how it felt.

The trade-off is real: a fully scripted interview cannot chase an interesting answer. Follow-ups decided in advance keep some depth without handing the discretion back.

What does a scoring rubric need to reduce bias?

A rubric reduces bias only if it removes the moment where an interviewer decides what a good answer is after hearing it. That needs three properties.

Anchors written in advance. For each question, a sentence describing a weak, an adequate and a strong answer, agreed before the first interview.
A score per question, not per person. An overall rating is where the halo lives: one charming answer lifts everything else.
Evidence beside every score. The phrase the candidate actually said. A score nobody can point to evidence for is an impression with a number on it.

An example for one must-have, "can bring a new team member up to speed":

Swipe the table sideways to see every column.

Level	What the answer contains
Weak	General statements about liking to help people, no example
Adequate	A real example, with what they did, but no sign of adjusting to the person
Strong	A specific person, the difficulty they had, how the candidate noticed, and what changed

Then two habits. Every interviewer scores alone before any discussion, so the first opinion voiced does not become everyone's. And "culture fit", if it stays on the form at all, is replaced by something behavioural, such as "gives and takes direct feedback", with its own anchors. Undefined, it is the category the Rivera study above describes.

Does blind resume review reduce bias?

Sometimes, and sometimes the opposite. The two best-known studies point in different directions, and both are worth knowing before you redact anything.

The case for. Goldin and Rouse studied US orchestras that began auditioning musicians behind a screen, and found "the screen increases by 50% the probability a woman will be advanced out of certain preliminary rounds".

The case against. The Australian Government's Behavioural Economics Team ran a randomised trial in which over 2,100 public servants from 14 agencies shortlisted applicants from either standard or de-identified applications. Reviewers turned out to favour women and minority candidates when they could see who they were: compared with de-identified applications, participants were 2.9% more likely to shortlist female candidates when they were identifiable, and minority women were 8.6% more likely to be shortlisted. BETA concluded that "de-identification may frustrate efforts aimed at promoting diversity".

Read together, blind review removes a signal in both directions. It helps where reviewers discriminate against a group, and does nothing, or harm, where they do not or where they deliberately compensate. It also stops at the resume: the moment a first round becomes a conversation, the candidate is no longer anonymous. The dependable part is narrower: remove what no reviewer should weigh, such as photographs and dates of birth, and put the effort into structure, which works whoever the reviewer is.

What can AI make worse?

AI screening can be more consistent than a tired person, and it can also be consistently wrong, for everyone, at once. Three risks are documented well enough to plan around.

It can learn the pattern it was meant to remove. In a resume audit of text-embedding models used to retrieve and rank resumes, Wilson and Caliskan found the models "significantly favoring White-associated names in 85.1% of cases and female-associated names in only 11.1% of cases", across nine occupations and over 500 resumes and 500 job descriptions.
Speech recognition is not equally accurate for everyone. Five commercial systems tested by Koenecke and colleagues made about twice as many errors transcribing Black speakers as white speakers: an average word error rate of 0.35 against 0.19. Any score built on a transcript inherits the transcript's mistakes.
Consistency hides error. A person's bias varies from day to day. A model's bias is applied identically to every applicant, and a ranking with no visible evidence gives nobody the chance to notice.

People also expect more of AI than the evidence supports. In Pew Research Center's survey of 11,004 US adults, 47% thought AI would be better than humans at evaluating all job applicants in the same way, and 15% thought it would be worse. Treating the same way is not the same as treating fairly, and only your own numbers can tell you which one you are getting.

How do you check your own screening for bias?

Count selection rates: of the people who entered a stage, what share moved on, broken down by group where you lawfully hold that data. Do it for each stage, not just for hires, because a gap at the first round never shows up in a small number of final offers.

In the United States, the enforcement agencies' rule of thumb is written into the Uniform Guidelines. The EEOC's own questions and answers on the Guidelines say the agencies "will generally consider a selection rate for any race, sex, or ethnic group which is less than four-fifths (4/5ths) or eighty percent (80%) of the selection rate for the group with the highest selection rate as a substantially different rate of selection", and that the rule "is not intended to resolve the ultimate question of unlawful discrimination". It is a signal to look closer, not a pass mark.

If a role is in New York City and you use an automated employment decision tool, the city requires the tool to have had a bias audit within one year of its use, information about the audit to be public, and notices to candidates. The wider legal picture is in AI interviews in 2026. None of this is legal advice; ask counsel what applies to you.

One role's numbers are small and noisy: a prompt to read the evidence, not a verdict.

How does Career1's AI interviewer behave?

Described as it is, including the parts that cut against us.

The same structure, not the same questions. Every applicant to a role gets the same section plan (an introduction, three technical questions, one on experience, one behavioural and a close) and is scored on the same dimensions. The questions themselves are generated from the role and each applicant's resume, and the follow-ups depend on the answers. That tests each person's own claims, and it is less standardised than an identical script. If your process needs identical questions, ask them in the human round.
It scores words, not faces. The AI interviewer's report is written by a language model from the transcript, the resume and the job, quoting the answers it relied on. As the privacy policy's AI section states, nothing scores the candidate's face, expressions, voice, accent, background or camera. The transcript itself comes from speech recognition, though, and the research above shows those errors are not evenly spread. Where a transcript reads oddly, play the recording before trusting the score.
It is not blind. The interviewer and the report both see the applicant's name and resume.
It scores "culture fit". It is one of five dimensions in the report, and it is the one this page tells you to discount unless you have defined it.
English only. At the time of writing the interview runs in English, which disadvantages applicants who would do the job well in another language.
A person decides. Applicants are ranked by their match score, Career1 rejects nobody, and every decision stays with your team. There is a complete sample interview report to judge the evidence by.
No independent bias audit yet. Career1 has not commissioned one. If you hire in New York City, the audit obligation is yours, and you should ask any vendor, us included, what it can give you.

The interview takes about eight minutes and every applicant can take it, which removes one real source of unfairness, being among the resumes nobody had time to read. It does not remove the others on this page. Plans, including a free plan of two interviews a month, are on the pricing page.

What should a hiring team change first?

Write the must-haves first, as things a person can do.
Screen only eligibility facts from the document, asked on the form. The process for a large pile is in how to screen 200+ applicants.
Ask everyone the same questions, with follow-ups decided in advance.
Anchor the scores, score alone, then discuss.
Count selection rates per stage each quarter.
Ask any AI vendor what it scores, from what inputs, with what evidence, and whether it has been audited. "Not yet" is an honest answer; "our AI has no bias" is a reason to leave.

Sources, checked 29 September 2026

Every page below was opened and read on 29 September 2026. Career1 is not affiliated with any organisation named here.

US Office of Personnel Management, structured interviews: opm.gov.
Bertrand and Mullainathan, "Are Emily and Greg More Employable than Lakisha and Jamal?", NBER working paper 9873: nber.org/papers/w9873.
Rivera, "Hiring as Cultural Matching", American Sociological Review, 2012, as released via ScienceDaily: sciencedaily.com.
Goldin and Rouse, "Orchestrating Impartiality: The Impact of 'Blind' Auditions on Female Musicians", NBER working paper 5903: nber.org/papers/w5903.
Behavioural Economics Team of the Australian Government, "Going blind to see more clearly", June 2017: pmc.gov.au.
Wilson and Caliskan, "Gender, Race, and Intersectional Bias in Resume Screening via Language Model Retrieval", AAAI/ACM AIES 2024: arxiv.org/abs/2407.20371.
Koenecke and colleagues, "Racial disparities in automated speech recognition", PNAS, 2020: PubMed Central.
Pew Research Center, "AI in Hiring and Evaluating Workers: What Americans Think", 20 April 2023: pewresearch.org.
US Equal Employment Opportunity Commission, questions and answers on the Uniform Guidelines on Employee Selection Procedures: eeoc.gov.
New York City, automated employment decision tools: nyc.gov.
Career1, how interviews are scored: privacy policy, AI section.

Questions people ask

How do you reduce bias in candidate screening?

Fix the criteria before reading any application, screen only hard eligibility facts from the document, ask every candidate the same job-related questions, score each answer against benchmarks written in advance, score alone before discussing, and count selection rates at each stage. Structure does most of the work, because it removes the discretion bias depends on.

Why are structured interviews less biased?

Because every candidate gets the same questions and is scored against the same benchmarks, which leaves less room for rapport or first impressions to decide. The US Office of Personnel Management says interviews with more structure show higher validity, better agreement between raters and less adverse impact than unstructured ones.

Does blind resume screening reduce bias?

The evidence is mixed. Blind orchestra auditions made women more likely to advance from preliminary rounds, but an Australian Government trial found reviewers favoured women and minority candidates when they could see them, so de-identifying applications would have reduced their chances. Removing irrelevant details helps; structure matters more.

What is the four-fifths rule in hiring?

It is a rule of thumb from the US Uniform Guidelines on Employee Selection Procedures. If one group's selection rate is below four-fifths, or 80%, of the rate for the group selected most often, enforcement agencies generally treat it as a substantially different rate. The EEOC says it signals adverse impact but does not settle whether discrimination was unlawful.

Can AI reduce bias in hiring?

It can make screening more consistent, because every applicant is assessed the same way and nobody is skipped for lack of time. It can also repeat bias at scale, from training on past decisions or from errors in speech recognition. Whether it reduces bias for you depends on what it scores, the evidence it shows, and your own selection rates.

Can AI make hiring bias worse?

Yes. A 2024 study of text-embedding models ranking resumes found they favoured White-associated names in 85.1% of cases. Speech recognition systems have made about twice as many errors for Black speakers as for white speakers. A biased model applies the same skew to every applicant, which is why its evidence must be visible and checked.

Is culture fit a source of hiring bias?

It often is, when it is left undefined. In a study of elite professional firms, more than half of the evaluators ranked cultural fit, meaning similarity to existing staff in background, leisure pursuits and self-presentation, as their most important criterion at interview. Replace it with defined behaviours, each with written anchors, or drop it.

What should a scoring rubric include to reduce bias?

A benchmark for a weak, an adequate and a strong answer to each question, written before the first interview; a score for each question rather than one overall rating; and the candidate's own words beside every score. Interviewers should score alone before any discussion, so the first opinion voiced does not become the group's.

Do AI interviews discriminate against accents?

They can, indirectly, if scores depend on a transcript. Research on commercial speech recognition found error rates differed by race. Career1 scores the words of the answers, not the voice or accent, but the transcript comes from speech recognition, so a reviewer should play the recording wherever a transcript reads oddly before trusting a score.

Is Career1's AI interviewer audited for bias?

Not yet. Career1 has not commissioned an independent bias audit, and says so in its privacy policy. Its report is written from the transcript, resume and job, with quotes as evidence, and it rejects nobody. If you hire in New York City, the bias audit obligation under Local Law 144 sits with the employer.

Should AI make screening decisions on its own?

No. Let software gather evidence and rank, and let a person decide with that evidence in front of them. An automatic rejection is a decision nobody reviewed and nobody can explain. Career1 ranks applicants and recommends, rejects nobody, and leaves every decision to the hiring team.

Is it fair to ask every candidate the same questions?

It is the fairest default, because the answers can be compared on the same terms. Adjustments for disability or other accommodation needs are the exception, and should change how a question is asked, not what it tests. Probing is fine when the follow-ups are decided in advance and used with everyone.

Try it on a real role, free

Post a role and every applicant gets an eight-minute spoken interview in the browser, with no scheduling. You get a score, strengths, risks, a recommendation, the transcript and the recording for each one.

The free plan vets two candidates a month and needs no card. Candidates never pay.

Start hiring free See pricing

Keep reading

All articles

How to reduce bias in first-round candidate screening