April 17, 2026
Artificial intelligence (AI) is increasingly used in hiring processes, often through human-like avatars that simulate real interviews. While these systems are promoted as objective and fair, an open question remains: how do applicants actually perceive fairness, trust, and bias when interacting with AI interviewers – especially when identity cues like race and gender are involved?
A research team from the Technical University of Munich, led by Ka Hei Carrie Lau with our EdTech Fellow Dr. Efe Bozkir, and Ed Tech Director Prof. Dr. Enkelejda Kasneci, as well as Dr. Philipp Stark from Lund University, conducted a controlled experimental study to investigate how avatar appearance shapes applicants’ perceptions in AI-mediated hiring. Their findings provide new evidence that justice-related evaluations of AI are not only about algorithmic bias, but also being shaped by the unequal social responses that avatar appearances trigger in applicants.
Empirical Study
215 participants completed a simulated job interview with a real-time, AI-powered avatar.
To investigate the impact of identity cues on user perceptions of a simulated job interview, the study used a 2×2 between-subjects experiment, manipulating whether the avatar matched the participant’s race and/or gender:
- No match (neither race nor gender aligned)
- Gender match only
- Racial match only
- Full match (both race and gender aligned)
Participants answered interview questions verbally, filled a post-interview survey, received a standardized rejection, and finally a post-outcome survey, allowing researchers to examine reactions under negative outcomes. In addition to questionnaires, the study combined sentiment analysis and eye-tracking data to capture both explicit self-reported perceptions and implicit behavioral responses.
Results
The study examined how racial or gender matching affects trust, fairness, bias perception, and implicit behavioral responses.
Trust
Across all conditions, participants reported consistently high trust in the AI interviewer. Matching or mismatching identity traits had no significant effect on trust levels.
Perceived Bias
Racial mismatch led to higher perceptions of bias. Participants were more likely to believe the system treated them unfairly when the avatar’s race differed from their own.
Fairness (Distributive Justice)
A notable interaction effect emerged: partial matches (only race or only gender aligned) resulted in lower perceived fairness ratings than both full matches and complete mismatches. This creates an unexpected “partial match paradox”, where mixed identity cues invite greater critical evaluation.
Implicit Behavioral Measures
- Eye-tracking showed increased attention to the avatar’s face under racial mismatch, suggesting heightened scrutiny.
- Sentiment results were limited, showing only a small interaction effect for polarity, suggesting that emotional tone shifted slightly across identity conditions, but without overall effects of identity matching.
Takeaway Message
This study highlights that fairness in AI hiring cannot be understood purely from an algorithmic perspective, but must also account for social categorization processes triggered by avatar identity cues. Applicants interpret and evaluate AI systems through these social signals, meaning that even technically fair decisions can be perceived as unfair depending on how the system is presented. To address this, the researchers recommend designing AI interview systems with clear introductory messages and thoughtful avatar selection to manage expectations and reduce discomfort, providing explainable feedback (XAI) after rejection to clarify decisions and support applicants, leveraging multimodal measures such as eye-tracking to better understand how users process avatar identities beyond self-reports, and fostering interdisciplinary collaboration between social scientists and HCI researchers to anticipate fairness risks and guide responsible, user-centered design.
References
- Lau, K. H. C., Stark, P., Bozkir, E., & Kasneci, E. (2026). Skin-Deep Bias: How Avatar Appearances Shape Perceptions of AI Hiring. In Proceedings of the 2026 CHI Conference on Human Factors in Computing Systems. https://doi.org/10.1145/3772318.3790379
The project is opensourced: https://gitlab.lrz.de/hctl/skindeepbias
