Face value: How avatar identity shapes epistemic trust in AI-mediated learning

Synthesis: Anthis & Kyriakidou-Zacharoudiou investigate how avatar identity cues (race, gender, age) shape epistemic trust and enacted reliance in AI-mediated learning. Two experiments isolated these effects: Study 1 (N=102) used a within-subjects laboratory design with tightly controlled avatars; Study 2 (N=294) used a between-subjects online design across varied instructional domains. Across both studies, social identity cues strongly influenced evaluations and behavior: White avatars, and in STEM contexts Asian male avatars, were rated more credible and competent, while older Black female avatars faced compounded penalties on all measures. Participants were more likely to adopt guidance from avatars aligned with their racial ingroup or stereotypical expectations of expertise. Domain moderated effects, with STEM and procedural tasks amplifying bias while reflective and interpersonal tasks attenuated it.

Key Findings

• Avatar identity shapes trust and uptake. Anthropomorphic avatars designed to increase engagement also activate social heuristics that reproduce offline patterns of bias — influencing who is trusted and who is ignored.
• Four outcomes + behavioral uptake. The studies measured credibility, warmth, competence, and willingness to act, alongside a behavioral index of epistemic uptake (incorporating AI guidance into learners' own work).
• Perceptual hierarchies. White avatars, and in STEM contexts Asian male avatars, were rated more credible and competent; older Black female avatars faced compounded penalties across all measures.
• Ingroup alignment drives adoption. Participants were more likely to adopt guidance from avatars aligned with their racial ingroup or stereotypical expectations of expertise.
• Domain moderation. STEM and procedural tasks amplified bias, while reflective and interpersonal tasks attenuated it — bias is context-dependent, not uniform.
• Trust and AI tutors is a design variable. The findings show that trust in AI tutors is shaped by surface identity cues, not just underlying capability.

What this means for practice

• Designers. Calibrate avatar design by domain instead of fixing it globally: bias was amplified on STEM and procedural items — where the largest competence gap appeared between young White and older Black female avatars — and attenuated on reflective and interpersonal tasks.
• Reduce the salience of demographic cues by using stylized or less anthropomorphic avatars and rotating diverse avatars across tasks, while pairing those changes with consistent behavioral reliability so engagement does not drop.
• Evaluate avatar interventions by their effect on enacted guidance uptake, not just attitudes or self-reported trust, because identity bias operated at the level of action as well as evaluation.
• Administrators. Treat default avatar configurations, representation policies, and deployment contexts as pedagogical and ethical decisions, and build in design guidelines, reliance monitoring, and vendor accountability for AI tutor procurement.
• Researchers. Model trust dynamically, since this study modeled trust and epistemic uptake as static outcomes and could not show whether the measured dimensions are causal intermediaries or co-occurring correlates of behavior.

Limitations

• Two convenience samples totaling N = 396 — Study 1 (N = 102) in a within-subjects laboratory design at one university and Study 2 (N = 294) in a between-subjects online design — drawn largely from an English-speaking, WEIRD university population.
• Only a limited set of visible identity cues (race, gender, age) was varied, and interactions with avatars were brief and controlled, so cultural and ecological generalizability is constrained.
• Trust and epistemic uptake were modeled as static outcomes rather than processes, so the design supports association but not evidence about mechanisms.
• Both samples were unpaid volunteers, with Study 2 recruited from a vetted online participant pool (e.g., Prolific, Qualtrics Panels), so participants may differ from typical classroom learners.

Citation

Anthis, Z., & Kyriakidou-Zacharoudiou, A. (2026). Face value: How avatar identity shapes epistemic trust in AI-mediated learning. Computers and Education: Artificial Intelligence, 10, 100610.