Can Large Language Models Exhibit Paranoia? A study on Chat-GPT, Llama, Gemini and Claude.

Nowadays, large language models (LLMs) are an inseparable part of millions of users' daily lives. Due to the rapid development of LLMs a series of questions have been raised about their potentiality. Firstly, if they are capable of showing human-like performance on different disciplines, are they also prone to make human-like errors? Secondly, are they neutral agents that simply provide an answer, or various factors such as training data can influence their answers and judgments? Many mental disorders manifest on a language level and affect the flow of thoughts; Therefore, it is essential to examine the possibility of any language related malfunctions in LLMs. Considering that complex and novel advancements of LLMs have allowed us to treat them as participants of psychological tests, we have conducted the R-gpts test to determine whether our four chosen LLMs (ChatGPT 4, Claude, Gemini and Llama) show any sign of paranoia. We discovered that usually by increasing the temperature of each model, paranoia levels escalate. Additionally, some LLMs generally show higher scores than others under the same circumstances. This research highlights the need for safety measures to ensure unbiased output and reliable performance for all users.