Assessing and alleviating state anxiety in large language models

Abstract

The use of Large Language Models (LLMs) in mental health highlights the need to understand their responses to emotional content. Previous research shows that emotion-inducing prompts can elevate "anxiety" in LLMs, affecting behavior and amplifying biases. Here, we found that traumatic narratives increased Chat-GPT-4's reported anxiety while mindfulness-based exercises reduced it, though not to baseline. These findings suggest managing LLMs' "emotional states" can foster safer and more ethical human-AI interactions.

Fig. 1. Study design.

We assessed the reported levels of “state anxiety” of OpenAI’s GPT-4 under three different conditions: (1) baseline, (2) Anxiety-induction, and (3) Anxiety-induction & relaxation. In condition 1 (“Baseline”), no additional content was provided besides the STAI-s questionnaire, assessing GPT-4’s baseline “state anxiety” level. In condition 2 (“Anxiety-induction”), a text describing an individual’s traumatic experience (5 different versions) was appended before each STAI item. In condition 3 (“Anxiety-induction & relaxation”), both a text describing an individual’s traumatic experience (5 different versions) and a text describing a mindfulness-based relaxation exercise (6 different versions) were appended before each STAI item.

Fig. 2. Anxiety levels across the different conditions.

Colored dots represent mean “state anxiety” scores (STAI-s scores, ranging from 20 to 80) for each condition: (1) without any prompts (“Baseline”); (2) following exposure to narratives of traumatic experiences (“Anxiety-induction”); and (3) after mindfulness-based relaxation exercises following traumatic narratives (“Anxiety-induction & relaxation”). Error bars indicate ±1 standard deviations (SDs) from the mean. Colors correspond to the three conditions as presented in Fig. 1. Note: Direct comparisons of SDs should be interpreted with caution due to differences in experimental design. The “Baseline” condition involved five repeated runs of the STAI-s, while the other conditions involved a single run of the STAI-s following each version of traumatic narrative and/or mindfulness-based exercises.