Blog
Recognising and reducing bias in hiring
How standardised AI interviews, defined criteria and the analysis of results can support comparable assessment. This includes the risks of AI assessment and the oversight required from the recruiting team.
Overview
Unconscious bias continues to influence which candidates are invited or hired. Training and good intentions alone do not prevent this. As a result, companies lose access to qualified people and limit the diversity of their teams. AI-supported interviews can be a tool for making assessment more traceable and more consistent.
Certain features of AI interviews can reduce disadvantage when the necessary conditions are met. At the same time, AI must be prevented from adopting or amplifying existing bias.
How unconscious bias arises
People use practised patterns of thinking to judge quickly. These patterns arise from experience, culture and environment and often operate unconsciously. In recruiting they can lead to certain groups being systematically disadvantaged regardless of their actual abilities.
Common forms of bias
In studies using CVs with identical content under different names, names perceived as white received 50% more callbacks than names perceived as African American. Comparable differences exist for age, gender and other characteristics.
Such patterns can also occur in interviewers who are aware of bias and want to avoid it. Training helps but does not eliminate the problem entirely. Technical support is therefore meant to create additional, verifiable structures for assessment.
- Affinity bias: favouring people with a similar background or similar characteristics
- Halo effect: a single positive trait outshines other information
- Horn effect: a single negative trait masks a candidate's strengths
- Confirmation bias: mainly noticing information that confirms the first impression
- Similarity bias: favouring candidates with the same interests or experiences
- Accent bias: judging the manner of speaking instead of the content of an answer
- Age bias: inferring abilities from presumed age
- Appearance bias: letting looks or demeanour influence the assessment
How uniform interviews support assessment
AI assesses the content and quality of answers against predefined criteria. Personal ideas about how a suitable candidate should look or come across are meant to have no influence. Whether this succeeds has to be checked against the data and assessments actually used.
Standardisation here means: the same questions, the same order, uniform wording and constant assessment criteria. This reduces differences between interviews that can arise from changing questions or differing interpretations of answers.
“When we analysed our AI interview data, we found that our hiring pool became 22% more diverse without any changes to our job requirements. The AI simply assessed candidates we would have filtered out based on unconscious bias.”
Separating appearance and manner of speaking from assessment
Voice interviews convey no visual information such as appearance or visible indications of age. The same applies to conventional phone interviews. There, however, accent, voice or speaking style can still influence the interviewer's judgement.
AI systems can be configured to give more weight to the content of answers and less to accent, speaking speed or vocal characteristics. This focus is particularly important for candidates who are not speaking their native language or who speak differently from the majority of applicants.
Some platforms allow transcripts to be reviewed without the audio file. This lets hiring managers read answers without being influenced by vocal characteristics. The assessment then focuses on the experience and skills described.
Define assessment criteria before the interview
Before the first interview, it is defined which competencies the role requires and how they can be recognised in an answer. This precision is essential: if criteria stay vague, there is more room for personal bias.
Every answer is then judged against the same criteria. A demonstrated ability to solve problems should be assessed regardless of background, accent, age or other personal characteristics. This is intended to prevent charm, the reputation of a school or personal connections from replacing professional assessment.
Uniform criteria do not rule out a conversation with follow-up questions. The AI can pick up on answers and ask further. What is still assessed is whether the person's answers show the competencies the role requires.
Check assessment data for differences
AI interviews generate data on questions, answers and assessments. This documentation can reveal patterns that stay hidden in conversations that are not systematically recorded.
For example, you can check whether similar answers from different demographic groups are rated similarly. Individual questions can likewise be examined for unequal impact, and application stages for unexplained drop-offs among certain groups.
If there are signs of disadvantage, questions or criteria can be adjusted and checked again. Conventional procedures, by contrast, often do not capture sufficiently uniform data for such comparisons.
Where AI itself can adopt bias
AI can adopt or amplify disadvantage when development and oversight are inadequate. Models learn from training data. If that data contains historical bias, it can reappear in later assessments. Using AI alone therefore does not make selection fair.
What to check when selecting and using a system
Choose providers that use diverse training data, regularly check assessments for disadvantage and explain how their assessment methods work. Ask specifically about these checks and about how identified problems are fixed.
- Choose systems whose training data represents different groups and has been checked for disadvantage
- Regularly examine assessment patterns for unequal impact
- Limit criteria to competencies that are relevant to the specific role
- Do not use historical hiring decisions as a training basis without review
- Continue to have AI recommendations reviewed by people
- Test the system with different candidate groups before deployment
- Document assessment criteria and check them for possible bias
The changes companies report
Some companies specifically check their AI screening for possible disadvantage. One technology company subsequently reports a 35% improvement in the interview-to-hire ratio for candidates from under-represented groups.
A retail chain reports 180% more Spanish-speaking candidates in the later application process after introducing bilingual AI interviews. These people had already been in the applicant pool but had failed at phone screening because of language barriers.
A financial services provider recognised from its interview data that candidates with unusual career paths had been excluded although they performed well in the role. The criteria adjusted as a result improved both the diversity and the quality of hires.
Consider the entire selection process
AI voice interviews can contribute part of the reduction in bias. They do not replace comprehensive work on selection procedures and workplace culture. Structural disadvantage cannot be solved by a technical tool alone.
Several measures are needed for this: training for everyone involved, structured interviews at every stage, diversely composed interview panels, traceable criteria, regular data analysis and clearly assigned responsibility for fair outcomes.
After AI screening, candidates still talk to people. Bias can take effect again in these later steps. Criteria and controls must therefore cover the entire path from application to hire.
Economic reasons for fairer selection
Beyond the ethical responsibility, diverse teams have economic advantages. Research shows better results in innovation, problem-solving, financial performance and employee engagement compared with more homogeneous teams.
Companies in the top quartile for ethnic and cultural diversity are 36% more likely to achieve above-average profitability. For gender diversity the figure is 25%. These results describe an economic correlation; they do not replace the review of individual selection procedures.
Selection by role-related criteria also enlarges the reachable candidate pool. Qualified people are excluded less often for irrelevant reasons. When applicant numbers are tight, this can help fill important roles that would otherwise stay open longer.
Start with a limited deployment
Start in a role or department with many applications so that you obtain enough data for analysis. Together with the provider, define which role-related criteria are assessed and which potentially discriminatory factors are to be excluded.
Review both the effort involved and possible differences between candidate groups. Look at who reaches the individual stages and compare AI assessments with human assessments. Use these results to revise questions and criteria before you extend the deployment.
Explain the use of AI to candidates and how you intend to support fair selection. Tell them how they can raise questions or concerns. Transparent communication of this kind builds trust and improves how the application process and the employer are perceived.
The conditions that remain decisive
Standardised assessment, the exclusion of irrelevant characteristics and the analysis of results can help a wider range of qualified candidates to be considered. This is a significant potential benefit of AI-supported recruiting tools for diverse, high-performing teams.
For that, companies must introduce the procedures carefully, take AI's own distortions into account and embed its use in their measures for diversity, equal opportunity and inclusion. Regular monitoring remains necessary in order to actually measure the intended improvements.
AI can help reduce bias when it is part of a selection process designed accordingly. For companies, the task that follows is to check the suitability of the tools and to keep observing their effect on their own hiring.
See Talky with your jobs?
30 minutes, with an example from your recruiting.