The Invisible Embedded “Values” Within Large Language Models: Implications for Mental Health Use

AbstractValues are an integral part of any mental health intervention, profoundly shaping definitions of psychopathology and treatment approaches. As large language models (LLMs) hold promises for mental health applications, it is prudent to evaluate their embedded “values-like” abilities prior to implementation. This study uses Schwartz's Theory of Basic Values (STBV) to quantify and compare the motivational “values-like” abilities underpinning four leading LLMs. The results suggest that Schwartz’s theory can reliably and validly measure “values-like” abilities within LLMs. However, apparent divergence from published human values data emerged, with each LLM exhibiting a distinct motivational profile, potentially reflecting opaque alignment choices. Such apparent mismatches with human values diversity might negatively impact global LLM mental health implementations. The appropriate transparency and refinement of alignment processes may be vital for instilling comprehensive human values into LLMs before this sensitive implementation in mental healthcare. Overall, the study provides a framework for rigorously evaluating and improving LLMs’ embodiment of diverse cultural values to promote mental health equity.