Beyond Performance Disparities: A Three-Level Audit of Representational Harm in CelebA

This collection of papers documents multiple risks and failure modes in modern AI systems spanning datasets, model internals, human-AI interaction, and sociotechnical deployment. Key findings: (1) CelebA exhibits entrenched cultural double standards that manifest across dataset structure, model feature weights (SHAP/XGBoost), and attention maps (Grad-CAM), producing hyper-scrutiny of women and categorical exclusion of older men; (2) evaluation frameworks for agent human-likeness (HumanStudy-Bench) show agent behavior can either fully replicate or completely fail human-validated effects, and “w