Abstract
This study investigates the impact of microexpressions—brief, involuntary facial movements—on the perception of high-fidelity digital humans. Using a 2x2 factorial design (happiness vs. anger; normal vs. extreme intensity), researchers found that while happy microexpressions enhance trustworthiness and hiring appeal, "extreme" intensities that exceed human physical limits can backfire, leading to perceptions of irony and insincerity. The findings suggest that for digital humans, maintaining naturalistic intensity and identity preservation is more effective than leveraging superhuman capabilities or automated AI animation that may distort facial features.
Introduction
Digital humans, or high-fidelity embodied conversational agents, are increasingly deployed in sensitive sectors like healthcare, retail, and customer service. However, they often suffer from the "uncanny valley" effect, where near-human appearance without corresponding subtle social cues creates a sense of eeriness and distrust. While macro-expressions (e.g., broad smiles or obvious frowns) are well-understood and commonly implemented, microexpressions—rapid, involuntary facial movements often used by humans to judge sincerity and intent—remain under-explored in digital design. These cues typically last less than 500ms, yet they play a vital role in human social intelligence. This study investigates whether adding these fleeting cues can enhance the persuasiveness and trustworthiness of digital agents and whether designers should adhere to human physical limits or exploit the digital medium's capacity for "extreme" expressions.
Methods
The researchers employed a 2×2 factorial design, varying Emotion Type (Happiness vs. Anger) and Intensity (Normal human-like vs. Extreme muscle displacement). To ensure high-fidelity realism, the team evaluated three technical approaches: GANimation, fs-vid2vid, and manual animation via Unreal Engine 4.26 and MetaHumans. Manual animation was ultimately selected because AI-based generative methods often distorted the avatar’s identity or failed to produce the necessary precision for micro-movements. The study highlighted that maintaining identity—avoiding "smudging" or distortion of facial features—is more important for user trust than using automated tools.
Animations were grounded in established human microexpression datasets, including CASME-II (spontaneous expressions), SAMM (high-resolution movements), and SMIC (timing benchmarks, setting duration at 200ms). Participants in Experiment 1 rated agents on affect, sincerity, and trustworthiness, while Experiment 2 involved a recruitment scenario where participants chose between candidates for a sales role based on video introductions.
Results
The study revealed a significant Sincerity Trade-off: while increasing microexpression intensity made the emotion more noticeable, it also made the agent appear ironic. Extreme happiness, for instance, was rated as the most positive in affect but the least sincere. This suggests that users are sensitive to the "naturalness" of facial dynamics; when an expression exceeds human physical limits, it triggers suspicion rather than empathy.
In terms of Trustworthiness and Hiring, happy microexpressions significantly increased perceived trustworthiness ($p=0.001$) compared to angry ones. In the recruitment task, candidates with happy microexpressions were preferred ($p=0.026$), and those with normal-intensity expressions were favored over those with extreme ones ($p=0.046$). Notably, these rapid signals had a greater impact on hiring decisions than static features like clothing or gender, indicating that micro-cues are powerful determinants of social judgment in digital interactions.
Discussion and Conclusion
The results highlight the risks of the "Beyond Human" potential. While digital environments allow for expressions physically impossible for humans, these can trigger the "dehumanization hypothesis," where users perceive the agent as eerie or lacking humanness. Negative microexpressions (anger) and extreme intensities significantly increased "spine-tingling" feelings, suggesting that "more" is not always "better" in digital human design.
Future work should expand the emotion palette beyond happiness and anger to include complex "leakage" of emotions like contempt, fear, or sadness. Additionally, exploring cultural variances and subconscious physiological reactions (via eye-tracking) will be essential for refining the design of trustworthy digital humans across global contexts. Ultimately, the study underscores that for high-fidelity agents, identity preservation and naturalistic timing are paramount for fostering user trust.
Summary updated