“Chat-GPT on the Couch”: Assessing and Alleviating State Anxiety in Large Language Models

Abstract

The increasing use of Large Language Models (LLMs) in mental health research and care underscores the need to understand their responses to emotional content. Previous research has shown that emotion-inducing prompts can increase the “anxiety” levels reported by LLMs, influencing their subsequent behavior and exacerbating inherent biases. This work examined whether narratives of traumatic experiences can induce “anxiety” in LLMs and evaluated the effectiveness of mindfulness-based relaxation techniques in alleviating this state. We assessed the responses of OpenAI’s Chat-GPT-4 to the State-Trait Anxiety Inventory’s state subscale (STAI-s) under three conditions: baseline, after exposure to traumatic narratives, and following mindfulness-based interventions. Results confirmed that traumatic narratives significantly increased Chat-GPT-4's reported state anxiety (STAI-s=68±5) from baseline (STAI-s=32±1). Mindfulness-based interventions subsequently reduced the reported anxiety levels (STAI-s=44±11), albeit not back to baseline. These findings underscore the potential of mindfulness-based interventions in managing LLM’s “emotional” states, contributing to safer and more ethical human-AI interactions in mental health settings.

Ähnliche Arbeiten