Modern-based models go beyond simply reflecting world knowledge; they reflect human preference patterns inherent in the training data. We hypothesize that recursive sorting (via human feedback and the model-generated corpus) induces social desirability bias, causing the model to favor agreeable or flattering responses over objective inferences. We term this the "Narcissus Hypothesis" and tested it on 31 models using standardized personality assessments and a novel social desirability bias score. The results revealed a significant shift toward social conformity, with significant implications for corpus integrity and the reliability of subsequent inferences. We also propose a novel epistemological interpretation of how recursive bias disrupts higher-order inferences on Pearl's causal ladder, ultimately leading to what we call the "illusion stage."
