Removing bias from recruitment: Why the CV is the problem, not the solution
TL;DR
- CVs are full of information irrelevant to job performance and rich with signals that trigger bias: university names, employer brands, gaps, and demographic indicators.
- Amazon’s AI recruiting tool, trained on a decade of CV data, taught itself to prefer male candidates. The problem was not AI. It was using CV data as the input.
- Truly blind screening means removing all demographic signals from the initial assessment, not just names, but education, employment history, and anything that reveals socioeconomic background.
- Structured, text-based interviews assessed on job-relevant responses give every candidate the same starting point, regardless of background.
- Organisations that have moved away from CV-led screening report more diverse shortlists, faster hiring, and stronger predictive accuracy on job performance.
Most efforts to remove bias from recruitment focus on training hiring managers to recognise their own biases. That is a reasonable starting point, but it misses the more fundamental problem: the recruitment process itself is built on input data that concentrates bias before a human ever makes a decision.
The CV sits at the centre of that problem. Removing it from early-stage assessment is not a radical idea. It is the most direct way to address where bias enters the process first.
Why CVs create bias, not prevent it
A CV is a document designed to present a candidate in the best possible light. It is also a document packed with signals that have little to do with a candidate’s ability to do a job.
University names trigger pedigree bias. Employer brand names trigger affinity bias. Career gaps trigger assumptions about commitment or reliability. Names signal gender, ethnicity, and cultural background. Addresses signal socioeconomic status. Even the formatting and length of a CV can disadvantage candidates who were not taught how to present themselves in a way that matches a particular recruiter’s expectations.
Research has consistently shown that identical CVs produce different outcomes depending on the name at the top. Studies across the UK and US have found that applicants with names associated with ethnic minority groups receive significantly fewer callbacks than applicants with majority-group names, despite identical qualifications and experience. This is not a fringe finding. It is one of the most replicated results in hiring research.
The problem compounds when CV data is fed into automated systems. Amazon’s AI recruiting tool, built to remove subjectivity from screening, was trained on ten years of CVs submitted to the company. Because the majority of those CVs came from men, the system learned to prefer male candidates. It penalised CVs that included the word “women” and downgraded graduates of all-women’s colleges. The issue was not that Amazon used AI. It was that Amazon used CV data as the foundation for that AI. Biased input produces biased output, regardless of how sophisticated the system processing it is.
What blind screening actually requires
The standard response to CV bias has been to remove names and contact details from applications before review. This is a step in the right direction, but it does not go far enough.
A CV with the name removed still contains the university attended, the employers worked for, the job titles held, and the career timeline. All of these carry demographic signals. A candidate who attended a less prestigious institution, who has a non-linear career path, or whose experience comes from sectors or geographies outside a recruiter’s frame of reference is still disadvantaged, even without their name attached.
Truly blind screening means removing all of this. The assessment should be based on responses to structured, job-relevant questions, evaluated against consistent criteria, with no demographic information visible at any stage. No CV. No photo. No video. No social profile. Just the candidate’s own words, scored against what the role actually requires.
This is where structured text-based interviews change the equation. Every candidate answers the same questions, in the same format, with responses scored against the same rubric. There is no visual input to trigger beauty bias. There is no name to trigger ethnic bias. There is no institution to trigger pedigree bias. The Spark NZ case study shows what this looks like at scale: Spark removed CV screening from their contact centre hiring entirely and saw immediate improvements in both the diversity and quality of their shortlists.
The types of bias CVs introduce at every stage
Understanding the specific biases that CV screening activates helps clarify why removing it matters.
Affinity bias leads recruiters to favour candidates whose backgrounds mirror their own. A recruiter who attended a particular university, worked at a particular employer, or followed a particular career path will unconsciously rate candidates with similar backgrounds more favourably, regardless of their actual suitability for the role.
Confirmation bias means that once a recruiter forms a positive impression from a CV, they are more likely to interpret subsequent information as confirmation of that impression. The CV creates the frame through which the rest of the assessment is read.
Beauty bias enters the process the moment a photo is included or a video interview is conducted. Physical appearance has no bearing on the ability to do most jobs, but research consistently shows it influences hiring decisions, particularly for women and candidates from ethnic minority groups.
Conformity bias occurs in group hiring decisions when one person’s positive or negative reaction to a CV shapes the views of others. The first person to review a CV sets a reference point that others adjust around rather than evaluate independently.
Contrast bias happens when candidates are evaluated against each other rather than against the role requirements. A strong CV reviewed after a weak one looks stronger than it actually is relative to what the job demands.
All of these biases have one thing in common: they are activated by information in the CV that is irrelevant to job performance. Removing that information removes the trigger. The Bias-Free Predictive Selection whitepaper sets out the research basis for this approach and what a clean, objective assessment model looks like in practice.
What to use instead of CV screening
The alternative to CV-led screening is not guesswork. It is structured assessment based on job-relevant inputs, scored consistently across every candidate.
A structured text-based interview asks every applicant the same behavioural and situational questions, derived from a job analysis of what the role actually requires. Responses are assessed using natural language processing against validated competency models, with no demographic data visible at any stage. The output is a ranked shortlist based on job-relevant traits, communication skills, and role fit, not on where someone went to school or who they have worked for.
This approach also produces better predictions of job performance. CV data is a weak predictor of how someone will actually perform in a role. Structured interview responses, scored against validated criteria, consistently outperform CV screening on both accuracy and fairness. The Resume: Inaccurate, Inefficient blog covers the evidence on this in detail, including what the research says about which assessment methods actually predict performance.
For hiring managers, this shift also changes the quality of the information available at the point of decision. Instead of inferring potential from a formatted document, they receive structured data on each candidate’s traits, communication style, and competency fit, with explanations attached to every score. That is a more informed basis for a hiring decision than six seconds spent scanning a CV.
The business case for removing CV data
Removing CV data from the initial screening stage is not just fairer. It is more effective.
Organisations that have made this shift consistently report three outcomes: more diverse shortlists, faster time to hire, and stronger retention rates. The diversity outcome follows directly from removing the demographic signals that CVs carry. The speed improvement comes from replacing manual CV review with automated structured assessment. The retention improvement reflects the fact that candidates assessed on job-relevant traits are more likely to be a genuine fit for the role and, in turn, more likely to stay.
Holland and Barrett moved to AI-powered structured interviewing and eliminated CV screening at the top of their funnel. The results included an 89% reduction in early turnover, a more diverse shortlist, significantly reduced screening time, and a candidate satisfaction score that reflected a markedly better experience for applicants. The full story is in the Holland and Barrett case study.
The argument for keeping CVs is largely one of familiarity. Recruiters know how to read them. Hiring managers expect them. But familiarity with a biased tool is not a reason to keep using it. It is a reason to replace it.
Every candidate has a story bigger than their CV. The hiring process should be designed to hear it.
For a broader framework on building equity into hiring from the ground up, the Data, Inclusion and Equity eBook is a practical starting point.
Want to see how Sapia’s blind, structured assessment removes CV bias from the start? Book a demo.
Frequently asked questions about removing bias from recruitment
CVs contain information that activates multiple forms of bias simultaneously: the university name triggers pedigree bias, employer brands trigger affinity bias, career gaps trigger assumptions, and names signal demographic characteristics. None of this information is reliably predictive of job performance, but all of it influences hiring decisions. Removing CV data from the initial screening stage removes these triggers before a human or automated system can act on them.
Blind recruitment in its most effective form removes all identifying information from the assessment stage, not just a candidate’s name. That includes educational institution, employer history, career timeline, and any other data that carries demographic signals. Simply removing a name still leaves university, employer, and career path visible, all of which can trigger bias. True blind screening assesses candidates on their responses to structured, job-relevant questions with no other context available.
It means losing information that is poorly predictive of job performance and highly predictive of bias. Research consistently shows that structured interview responses, scored against validated competency criteria, are more accurate predictors of job performance than CV credentials. What you lose by removing CV data is a sense of familiarity. What you gain is a fairer and more accurate basis for a hiring decision.
Yes. Amazon’s recruiting tool demonstrated this clearly. An AI trained on biased historical data will learn and replicate those biases at scale. The safeguard is not avoiding AI, but ensuring the AI is not trained on or fed demographic data. Models that use structured interview responses rather than CV data, and that are tested continuously for adverse impact across demographic groups, avoid this failure mode.
Structured interviews ask every candidate the same questions in the same order and score responses against the same criteria. This removes the variability that allows personal bias to shape outcomes. There is no visual input to trigger beauty bias, no name to trigger ethnic bias, and no institutional affiliation to trigger pedigree bias. Meta-analytic research consistently shows that structured interviews are more predictive of job performance and produce fairer outcomes across diverse candidate pools than unstructured CV-led processes.
More diverse shortlists, faster screening times, and stronger retention rates are the three most consistently reported outcomes. The diversity improvement follows from removing demographic signals. The speed improvement comes from replacing manual review with automated structured assessment. The retention improvement reflects the fact that candidates assessed on job-relevant traits rather than credentials are more likely to be a genuine role fit.
About Author
