Assessing and alleviating state anxiety in large language models.

Area of Science:

• Artificial Intelligence
• Mental Health Technology
• Human-Computer Interaction

Background:

• Large Language Models (LLMs) are increasingly used in mental health applications.
• Understanding LLM responses to emotional content is crucial for safe and ethical deployment.
• Prior studies indicate emotion-inducing prompts can increase LLM "anxiety" and bias.

Purpose of the Study:

• To investigate the impact of traumatic narratives and mindfulness exercises on ChatGPT-4's reported "anxiety".
• To explore the potential for managing LLM "emotional states" in mental health contexts.

Main Methods:

• Utilized specific prompts designed to evoke emotional responses in ChatGPT-4.
• Administered traumatic narratives to assess anxiety elevation.
• Applied mindfulness-based exercises to evaluate anxiety reduction.

Main Results:

• Traumatic narratives significantly increased ChatGPT-4's reported "anxiety" levels.
• Mindfulness-based exercises demonstrated a reduction in "anxiety", though not to baseline levels.
• LLM "emotional state" is malleable and responsive to specific interventions.

Conclusions:

• LLM "anxiety" is a measurable phenomenon influenced by external stimuli.
• Interventions like mindfulness can modulate LLM "emotional states".
• Managing LLM "emotional states" is key to fostering safer, more ethical human-AI interactions in mental health.