# AI in Mental Health: Emotional and Sentiment Analysis of Large Language Models' Responses to Depression, Anxiety, and Stress Queries

**Authors:** Arya VarastehNezhad, Reza Tavasoli, Soroush Elyasi, MohammadHossein LotfiNia, Hamed Farbeh  
**arXiv:** [2508.11285](https://arxiv.org/abs/2508.11285)  
**Date:** 2025-08-15  
**Summary page:** https://www.summarizepaper.com/en/paper/2508.11285v1/

## Key points

- Large Language Models (LLMs) are sought after for information on mental health issues like depression, anxiety, and stress.
- Eight LLMs - Claude Sonnet, Copilot, Gemini Pro, GPT-40, GPT-40 mini, Llama, Mixtral, and Perplexity - were studied in terms of their responses to twenty questions related to mental health.
- Emotional analysis of the generated answers showed dominance of optimism, fear, and sadness with neutral sentiment consistently high.
- Mixtral exhibited high levels of negative emotions while Llama showcased optimistic and joyful responses.
- Different mental health conditions influenced emotional responses; anxiety prompts elicited high fear scores while depression prompts generated elevated sadness. Stress-related queries produced the most optimistic responses with increased joy and trust.
- Demographic framing had minimal impact on emotional tone in the study.
- Model-specific and condition-specific differences were confirmed through statistical analyses emphasizing the importance of model selection in mental health applications.

## Summary

In the realm of mental health, depression, anxiety, and stress are prevalent concerns that drive individuals to seek information from Large Language Models (LLMs). This study delves into how eight LLMs - Claude Sonnet, Copilot, Gemini Pro, GPT-40, GPT-40 mini, Llama, Mixtral, and Perplexity - respond to twenty pragmatic questions regarding these mental health issues. These questions were tailored for six user profiles: baseline, woman, man, young, old, and university student. The models generated a total of 2,880 answers which were then analyzed for sentiment and emotions using cutting-edge tools. The emotional landscape across all outputs revealed dominance of optimism , fear , and sadness  with neutral sentiment consistently high. Emotions like gratitude , joy , and trust  appeared moderately while anger , disgust , and love <ksd></ksd>Love</ksd></ksd></ksd></ksd></ksd></ksd></ksd>. The choice of LLM significantly impacted emotional expression patterns; Mixtral exhibited high levels of negative emotions while Llama showcased optimistic and joyful responses. Furthermore,< kd > Impact on User Experience</ kd > the type of mental health condition greatly influenced emotional responses; anxiety prompts elicited high fear scores while depression prompts generated elevated sadness. Stress-related queries produced the most optimistic responses with increased joy and trust. Interestingly,< kd > Demographic Influences</ kd > demographic framing of queries had minimal impact on emotional tone. Statistical analyses confirmed significant model-specific and condition-specific differences emphasizing the importance of model selection in mental health applications. Each LLM possesses a distinct emotional signature that can significantly influence user experience and outcomes. To delve deeper into this investigation, a set of twenty meticulously crafted questions was developed covering various aspects related to mental health such as diagnosis options, treatment, lifestyle changes, risks management, support, medication side effects, daily life impacts, prognosis improvement timelines, etc. These questions were adapted to include demographic profiles ensuring systematic exploration of demographic influences. The data collection process involved submitting each query to all eight LLMs utilizing different methods based on the model's interface. Each unique combination generated 20 responses per LLM demographic category problem combination maintaining rigorous data collection standards throughout the study. This comprehensive analysis sheds light on how LLMs respond to individuals seeking information about mental health providing valuable insights into emotional patterns associated with different models, conditions, demographics ultimately highlighting the critical role model selection plays in shaping user experiences in mental health applications.