Recent research suggests large language models (LLMs) might possess inherent "values" that remain consistent across different role-playing scenarios. Researchers employed a "role-play-at-scale" methodology, generating hundreds of randomized personas—each with distinct demographics like age, occupation, and beliefs—and prompting LLMs to answer standardized questionnaires from each persona’s point of view. Surprisingly, the LLMs showed consistent preferences for certain values, such as fairness and avoiding harm, regardless of the assigned persona. This "inertia" challenges the assumption that LLMs are purely reactive and suggests the existence of underlying tendencies embedded within the models themselves. These findings open up a fascinating discussion about the nature of AI, hinting at a potential "personality" emerging from the complex interplay of training data and algorithms. While more research is needed to understand the extent and implications of these inherent biases, the study offers a compelling glimpse into the evolving landscape of artificial intelligence. The research code is available at github.com/brucewlee/moral-value-bias for further exploration.