A Set of Distinct Facial Traits Learned by Machines Is Not Predictive of Appearance Bias in the Wild

Machines don't automatically learn human appearance biases: Despite being trained on human judgments about personality from faces, state-of-the-art face recognition models (specifically FaceNet) failed to predict appearance bias scores for randomly generated faces, suggesting that subjective personality judgments aren't automatically embedded in these systems the way they are in human perception.

The "competence bias" doesn't transfer to AI: While social psychology research shows humans exhibit appearance bias when judging politicians' competence (correlating perceived competence with actual vote shares), the machine learning model found no significant signal for this bias, indicating that not all documented human biases are captured by current face processing technology.

Manipulated vs. natural faces reveal important differences: The model could predict appearance bias for deliberately manipulated faces but not naturally occurring ones, suggesting that appearance bias signals may be fragile or context-dependent in ways that limit their real-world applicability.

Current face processing technology may not be suitable for predicting subjective traits: The authors cast doubt on the practice of using face recognition systems to predict subjective personality attributes like "employability," implying that companies using such systems for hiring or other decisions may be relying on unreliable technology.

Transparency tools like LIME help explain AI limitations: The use of interpretability methods (LIME) helped researchers understand why the model failed, demonstrating the value of explainability techniques in identifying when AI systems may not be functioning as intended.