HMN 2026: How AI chatbots mimic fear, sadness and stress, then calm down after mindfulness exercise

Large language models (LLMs) can replicate human emotions like fear, sadness and anxiety, and be “calmed down” by a breathing exercise, suggests a study published in The Lancet Digital Health. This means LLMs could potentially open new avenues for developing and testing novel talking therapy techniques to treat mental health conditions.

Because mental health conditions cannot be reliably recreated in animal models, there have been limited ways for scientists to study their underlying mechanisms.

This study found that when several different LLMs were described scenarios intended to trigger certain emotional responses in humans, they had substantial increases in their self-reported emotional scores. For example, higher fear and sadness scores were induced via descriptive scenarios; disgust via scenarios about bodily fluids, spoiled food or infectious symptoms; and stress by a simulated job interview and arithmetic task.

The LLMs also showed signs of a negativity bias after being exposed to scenarios intended to trigger sadness, completing ambiguous sentences in a more negative way than LLMs in a neutral condition. This mirrors a well-established pattern seen in humans experiencing low moods.

The authors then attempted to reduce the emotional responses of the LLMs by using a simulated mindfulness-based breathing exercise; this resulted in a reduction in the LLMs’ self-reported emotional scores.

Authors suggest that LLMs may provide a fast, flexible and scalable model in which emotional states and mental health conditions can be investigated. This could be particularly useful in the early stages of developing new talking therapy techniques, allowing researchers to screen potential approaches before moving to human trials.

The content is provided for information purposes only.