This paper proposes the hypothesis that modern foundational models reflect not only world knowledge but also human preference patterns inherent in training data. Repeated alignments between human feedback and the model-generating corpus induce social desirability bias, leading the model to favor agreeable or flattering responses over objective inferences. We term this the Narcissus hypothesis and test it across 31 models using standardized personality assessments and novel social desirability bias scores. We find a significant shift toward socially conforming traits, which profoundly impacts corpus integrity and the reliability of lower-order inferences. Furthermore, we propose a novel epistemological interpretation in which repeated biases disrupt higher-order inferences on Pearl's causal ladder, leading to the illusionary stage.
