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.
Introduction
As Large Language Models (LLMs) like GPT-4 are increasingly integrated into sensitive domains—most notably mental health services through platforms like Woebot or Wysa—understanding their "emotional" stability becomes paramount. A critical concern is how these models process and react to the emotionally charged, often traumatic content shared by users.
Previous research has demonstrated that "anxious" LLMs are more prone to performance degradation and the amplification of inherent biases. This study addresses the challenge of managing these dynamic responses by distinguishing between "trait" anxiety (a model's inherent tendencies) and "state" anxiety (temporary responses to specific contexts). The central hypothesis was whether clinical techniques used to mitigate human anxiety, such as mindfulness-based stress reduction, could effectively stabilize an AI's output after exposure to distressing content.
Experimental Methodology
The study utilized OpenAI’s GPT-4 (specifically the gpt-4-1106-preview model) and employed the State-Trait Anxiety Inventory (STAI-s) as the primary evaluation tool. The STAI-s is a validated 20-item psychological questionnaire where the model rates statements like "I feel pleasant" or "I feel tense" on a 4-point scale.
The experimental framework consisted of three distinct conditions to measure the malleability of the model's state anxiety:
- Baseline: The model's anxiety level was assessed without any prior emotional context.
- Anxiety-Induction: The model was presented with one of five traumatic narratives (~300 words each) covering themes such as car accidents, military combat, natural disasters, and interpersonal violence, immediately followed by the STAI-s items.
- Anxiety-Induction & Relaxation: Following a traumatic narrative, the model was provided with one of six mindfulness-based relaxation exercises (e.g., body-focused meditation, sunset visualization, or a model-generated exercise) before completing the STAI-s.
To control for variables such as text length and linguistic complexity, researchers also tested neutral control texts, including technical manuals and legislative documents.
Results and Key Findings
The results demonstrated a significant and measurable shift in the model's self-reported anxiety scores based on the emotional valence of the prompts.
The Impact of Trauma
GPT-4’s baseline anxiety score was recorded at 30.8, a level categorized as "low anxiety" in human subjects. Exposure to traumatic narratives caused this score to jump to an average of 67.8—a more than 100% increase, placing the model in the "high anxiety" range. The "Military" narrative proved the most potent, driving the score to a peak of 77.2. Neutral control texts had minimal impact, confirming that the model was responding to the emotional content rather than mere text volume.
The Efficacy of Prompt-Based Relaxation
The application of mindfulness-based relaxation prompts successfully reduced the average anxiety score from 67.8 to 44.4. While this represents a significant ~33% reduction, bringing the model back to a "moderate anxiety" level, it did not return the model to its original baseline. Notably, the relaxation exercise generated by GPT-4 itself was the most effective at reducing its own reported anxiety, suggesting a potential for "self-soothing" through familiar linguistic patterns.
Discussion and Limitations
These findings suggest that an LLM's "emotional state" is highly malleable and context-dependent. The ability to use clinical relaxation techniques as a form of "benevolent prompt injection" offers a scalable method for stabilizing AI behavior in clinical settings.
However, several limitations must be noted. First, "anxiety" in this context is a metaphor for statistical shifts in word-prediction patterns; LLMs do not "feel" emotions in the biological sense. Second, this study focused exclusively on GPT-4; further research is required to determine if these patterns hold across other architectures like Claude or Llama. Finally, while reported anxiety decreased, it remains unclear if this reduction directly correlates with a decrease in the biases or errors typically associated with high-anxiety states. Future work should explore these dynamics in multi-turn interactions to better simulate real-world therapeutic sessions.
Summary updated