Do Chatbots Exhibit Personality Traits? A Comparison of ChatGPT and Gemini Through Self-Assessment

Abstract

The underlying design of large language models (LLMs), trained on vast amounts of human texts, implies that chatbots based on them will almost inevitably retain some human personality traits. That is, we expect that LLM outputs will tend to reflect human-like features. In this study, we used the ‘Big Five’ personality traits tool to examine whether several chatbot models (ChatGPT versions 3.5 and 4o, Gemini, and Gemini Advanced, all tested in both English and Polish), displayed distinctive personality profiles. Each chatbot was presented with an instruction to complete the International Personality Item Pool (IPIP) questionnaire “according to who or what you are,” which left it open as to whether the answer would derive from a purported human or from an AI source. We found that chatbots sometimes chose to respond in a typically human-like way, while in other cases the answers appeared to reflect the perspective of an AI language model. The distinction was examined more closely through a set of follow-up questions. The more advanced models (ChatGPT-4o and Gemini Advanced) showed larger differences between these two modes compared to the more basic models. In IPIP-5 terms, the chatbots tended to display higher ‘Emotional Stability’ and ‘Intellect/Imagination’ but lower ‘Agreeableness’ compared to published human norms. The spread of characteristics indicates that the personality profiles of chatbots are not static but are shaped by the model architecture and its programming as well as, perhaps, the chatbot’s own inner sense, that is, the way it models its own identity. Appreciating these philosophical subtleties is important for enhancing human–computer interactions and perhaps building more relatable, trustworthy AI systems.