Assessing AI Fairness in Job Interview Decisions

In the rapidly evolving landscape of recruitment technology, artificial intelligence (AI) avatars are increasingly conducting job interviews, promising efficiency and impartiality. Contrary to the widespread assumption that AI eliminates bias, new research led by Enkelejda Kasneci from the Technical University of Munich (TUM) reveals that applicants’ perceptions of fairness are significantly influenced by the avatars’ appearance, particularly their gender and skin color.

The study involved over 220 participants from Germany, the United Kingdom, and the United States who engaged in a simulated job interview via a photorealistic AI avatar. These avatars were programmed to display human-like interactive behavior, asking follow-up questions based on participants’ responses. The research team developed four avatar variants differing in gender (male or female) and skin color (light or dark) to investigate how these social cues affect trust and fairness perceptions.

Eye-tracking technology monitored where participants focused during the interviews, revealing that they scrutinized the avatar’s face more closely when its skin color differed from their own. Intriguingly, despite this increased attention, initial trust in the AI remained high across all participants regardless of demographic matches with the avatars. This suggests that the human-like interaction style of the AI fosters a baseline level of confidence in the system at first encounter.

However, the dynamic shifted once participants received rejection notices for the fictional job. At this juncture, applicants’ views on the fairness of the decision diverged, revealing complex biases shaped by social identification. Participants who saw avatars with different skin tones were more prone to suspect that bias influenced the rejection. Yet, most strikingly, those who matched the avatar in either gender or skin color—not both—felt the greatest sense of unfairness. These individuals judged the AI’s decisions as less fair than those who matched the avatar completely or not at all.

Kasneci underscores that these findings challenge the narrow focus on algorithmic bias alone. “Even when AI systems operate free from programmed prejudice, human social behavior can color perceptions of fairness,” she explains. The social identity that applicants unconsciously associate with the AI avatar influences their interpretation of its decisions, blurring the lines between technical impartiality and subjective experience.

This study highlights the crucial need for AI developers to integrate social psychological factors into system design, particularly for applications like recruitment where acceptance by all stakeholders is vital. Simply correcting datasets or models may be insufficient if the avatars conveying AI decisions trigger complex social responses.

In light of these insights, future AI recruitment tools might benefit from customizable avatars or designs that transcend traditional social categories to mitigate misperceptions. As AI continues to mediate more human interactions, understanding and addressing these subtler dimensions of bias will be essential to building trust and fairness into technological decision-making.

By exposing how surface-level avatar traits can shape applicant trust and justice evaluations, this research paves the way for more nuanced, human-centered AI interfaces that respect both ethical standards and social acceptance.