A study involving simulated AI interviews found that applicants who were rejected perceived the fairness of the process differently depending on the race and gender combination of the avatar delivering the outcome, showing that interface design can shape candidate trust and fairness perceptions independently of the AI’s scoring.
- Candidates' fairness perceptions shift with AI interviewer avatar's race and gender
- Partial demographic matching causes stronger feelings of unfairness
- Interface design is critical alongside algorithmic fairness in AI hiring
What happened
Researchers conducted a study involving around 220 participants who completed a simulated AI interview for a fictional customer support role. All candidates received a rejection outcome, but the study varied the photorealistic AI interviewer's avatar across four race and gender combinations. Despite identical decisions, participants’ perceptions of fairness differed based on the avatar’s appearance.
Interestingly, candidates who matched the avatar in only one characteristic — either race or gender — judged the interview process as less fair than those who matched fully or not at all. Eye tracking showed that participants paid more attention to faces of a different skin color from their own, and participants were more likely to suspect bias when the avatar's race differed from theirs.
Why it matters
The study demonstrates that fairness issues in AI hiring tools may originate from the interface design before any algorithmic bias is considered. Candidates do not interact with raw algorithms; they experience a face delivering outcomes, which adds social context and emotional reactions to automated decisions.
This finding challenges companies to broaden their fairness assessments beyond audits of AI scoring to include how avatar design and representation may influence candidate perceptions, especially after receiving negative decisions. A fair algorithm alone does not guarantee that the hiring process feels unbiased to applicants.
What to watch next
Organizations deploying AI interview systems should include diverse demographic groups in fairness testing to evaluate reactions across different avatar designs and outcomes. Comparing human-like photorealistic avatars with less humanized interfaces may reveal alternative approaches that reduce perceived unfairness and skepticism.
Developers may need to prioritize clear expectation-setting in AI interviewer design rather than striving solely for relatability or realism. Future research could explore why partial demographic matching produces stronger fairness concerns and identify avatar configurations that balance candidate comfort with transparent communication.