Visual Cues of Human-Likeness, Not Salience, Impact Trust-Related Human-Computer Interaction

As interactions with digital agents become increasingly integrated into daily life, understanding how visual representations influence social decision-making is critical. Previous research in human-computer interaction has frequently confounded the psychological effects of an agent's perceived human-likeness with the underlying visual salience of the stimuli. To address these persistent gaps, the present study systematically isolated the effects of human-likeness and visual cue trustworthiness on trust behavior while controlling for objective image properties. The present study expanded on and normed the Virtual Avatar Facial Stimuli Set (VAFSS), a comprehensive database comprising hundreds of identity-matched photographs and computer-generated avatars varying across a spectrum of human-likeness. Experiment 1 aimed to measure visual properties of the stimuli. To control for visual properties, image salience was objectively quantified through behavioral metrics (a dot-probe task) and computer vision models (OpenCV for bottom-up physical conspicuity and ResMem for top-down cognitive memorability). Utilizing these validated stimuli, subsequent experiments investigated the impact of facial trustworthiness and human-likeness on trust behavior using a Single-Shot Trust Game. Experiment 2 manipulated human-likeness along a morphing continuum, while Experiment 3 used a diverse, non-morphed set of naturalistic avatars rated post hoc for perceptions of human-likeness. Across both behavioral studies, results demonstrated a robust main effect of visual cue trustworthiness, where participants consistently invested more in trustworthy-looking agents. Crucially, human-likeness significantly moderated this behavior such that effect of facial trustworthiness on trust behavior was greater for highly human-like stimuli. Hierarchical linear mixed-effects modeling confirmed that this moderation persisted even after statistically controlling for both physical conspicuity and algorithmic memorability. These findings demonstrate that behavioral shifts in human-computer trust are not mere artifacts of an image's structural sharpness or memorability, but are driven by complex, top-down social evaluations inherently tied to the perceived "humanness" of the digital agent.