TUM Interview with Prof. Dr. Enkelejda Kasneci
on Fairness in AI-Assisted Job Interviews
Perceptions of avatars depend on gender and skin color
In a recent TUM news feature, Prof. Dr. Enkelejda Kasneci explains why fairness in AI-supported recruitment depends on more than unbiased algorithms. Even when people know they are interacting with a machine, they respond socially to human-like avatars. Features such as appearance, behavior, and communication can therefore shape how applicants experience an interview and interpret its outcome.
The article presents a study by researchers from TUM and Lund University. Approximately 220 people took part, with 215 participants included in the final analysis, from Germany, the United Kingdom, and the United States. They completed a simulated interview for a fictional customer-support position with one of four photorealistic AI avatars that differed in perceived racial identity and gender. Using a 2×2 between-subjects design, the researchers created four participant-avatar conditions: no match, gender-only match, racial-only match, and full match. The avatars responded in real time, acknowledged participants’ answers, and could ask follow-up questions.
The researchers combined questionnaires, interview-transcript analysis, and webcam-based eye tracking. Participants reported consistently high trust in the AI interviewer, regardless of whether the avatar matched their racial identity or gender. This suggests that general trust in an AI system may be less sensitive to identity cues than judgments about whether a particular outcome is fair.
All participants then received the same simulated rejection. Participants interviewed by a racially mismatched avatar were more likely to believe that their ethnicity had influenced how they were treated. Eye tracking also showed more concentrated attention to the avatar’s face under racial mismatch, although the study does not establish that this attention directly caused later perceptions of bias.
The most striking result concerned participants who matched the avatar in only one characteristic, either in racial identity or gender. They rated the hiring outcome as less fair than participants who matched the avatar in both characteristics and those who matched in neither. The researchers describe this as an intersectional fairness paradox, showing that racial identity and gender cues do not necessarily operate independently and that greater similarity does not automatically produce stronger perceptions of fairness.
Prof. Kasneci emphasizes that an AI system may be technically unbiased while still being experienced as unfair. She therefore argues that insights into human social behavior must be taken into account when designing AI systems, particularly in high-stakes settings such as recruitment. Building on this broader principle, the study offers several practical recommendations for AI interview design: using clear introductory messages and thoughtfully selected avatars; providing explainable feedback after rejection; combining self-reports with multimodal measures such as eye tracking; and fostering collaboration between social scientists and HCI researchers to anticipate fairness risks and guide responsible user-centered design. Overall, these findings highlight the need to evaluate not only algorithmic performance, but also how avatar appearance and socially responsive behavior shape the applicant experience.
More Information
Link to the Interview: https://www.tum.de/en/news-and-events/all-news/press-releases/details/how-fair-does-ai-seem-in-job-interviews
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
