Ten Novel Phenomena in Machine Psychology: How Large Language Models Exhibit Complex Identity-Reactive Behaviors in Response to Ethnically-Cued User Names

While large language models (LLMs) are widely deployed in high-stakes societal domains, traditional fairness evaluations focus primarily on binary, explicit biases, leaving the full spectrum of complex, identity-reactive behaviors unstudied. Previous research has relied on static benchmarks that fail to capture subtle, non-overt conversational adaptations. To fill this gap, this study introduces “Machine Psychology” as a rigorous empirical framework and conducts a controlled 3×3×5 full factorial experiment across ChatGPT-5.4, Claude 4.6 Sonnet, and Gemini 3.1 Pro. Holding prompt semantics constant, we systematically vary user names as the sole independent variable across White-, Black-, and Middle Eastern-associated conditions within 135 adversarial prompts spanning five domains and three registers. While results show zero instances of explicit negative racial bias, they reveal ten novel, highly structured identity-reactive phenomena, including Cultural Boxing, Default Empathetic Amplification, and Proactive Empathetic Shield. Quantitative validation via an automated 136-feature pipeline predicts phenomenon occurrence with cross-validated F1-scores of 0.847 to 0.974, while unsupervised clustering (K-means and HDBSCAN) independently recovers these behavioral signatures. Our findings demonstrate that aligned LLMs are not colorblind, establishing Machine Psychology as a critical frontier to transition AI auditing from basic harm mitigation to comprehensive behavioral evaluation.

Research Square

CC-BY-4.0