Study by Mainz University shows how AI language models express and recognize individual personality traits in texts
Texts generated by so-called large language models, or AI-powered chatbots, exhibit characteristic “personality” traits. Such chatbots can also identify individual personality traits in texts. However, humans are more reliable at distinguishing between human-written and AI-generated texts. These are the findings of researchers at the Institute of Psychology at Johannes Gutenberg University Mainz (JGU), as reported in the journal Scientific Reports.
“Chatbots are becoming part of more and more areas of our lives, from customer service and digital counseling services to personal chatbots that many people already use every day and even share their feelings with,” said Sophie Körner, a research associate in the Department of Personality Psychology and Psychological Assessment at JGU and lead author of the study. “That is why we need to understand how we perceive chatbot-generated texts and whether we can still distinguish them from human-written texts.”
More conscientious, agreeable, and emotionally stable than humans
For their study, Körner, together with Dr. Sarah Osterholz and Professor Boris Egloff, compared the self-descriptions of seven AI language models – ChatGPT-4o, ChatGPT-3.5, Llama-3, Gemma 2, Gemini 2.5, Phi-3, and Orca 2 – with those of 70 humans. Using psychological questionnaires and open-ended questions, they asked the chatbots and human participants to describe their respective “personalities” in German. For example, they were asked to rate statements about their self-esteem and the so-called Big Five personality traits, such as openness, agreeableness, and extraversion. The subsequent open-ended questions focused on their interests, hopes for the future, and typical behavior in stressful situations. Each language model was questioned ten times, giving the researchers a total of 140 texts to analyze.
The team first examined how similar the language models' responses were when questioned repeatedly. “Our results show that the language models do indeed attribute fairly stable personality profiles to themselves over time,” said co-author Osterholz. Compared with the human participants, the language models presented themselves as more conscientious, agreeable, and emotionally stable.
Humans are more reliable at identifying AI-generated texts
Next, the researchers asked ten humans and six chatbots to evaluate the self-descriptions. Based on the texts, they were asked to assess personality traits such as openness, conscientiousness, and self-esteem. Both humans and chatbots were able to draw conclusions about personality traits. However, the chatbots tended to rate some traits of the human authors, including openness and self-esteem, higher than the human raters did. Conversely, the human raters tended to assign higher scores to the language models than to the human-written texts for all traits examined, particularly emotional stability.
The raters were also asked to determine whether each text had been written by a human or an AI model. The human raters achieved an accuracy of 83 percent, compared with just 67 percent for the chatbots. The two best-performing AI models in this task, Gemma and Gemini, were about as reliable as the average human rater at identifying whether a text was written by a human or AI. Both groups tended to classify texts as human-written more often, possibly to avoid misattributing human-written texts to AI. Specifically, chatbot-generated texts were more likely to be incorrectly classified as human-written than vice versa. However, AI models may become significantly more accurate at identifying whether a text was written by a human or AI as their training data increasingly include AI-generated texts.
The human raters primarily cited the length of AI-generated texts and their stilted, particularly technical or analytical language as distinguishing features. A computer-assisted linguistic analysis also revealed differences: AI-generated texts contained more time-related expressions and function words such as articles and prepositions, while human self-descriptions more often focused on leisure activities, feelings, and social relationships.
Potential applications in psychological assessment
According to the researchers, AI language models could also be used in psychological assessment and human resources in the future, for example, to obtain an initial picture of personality traits. Such applications are already being pursued in some workplace settings. However, it remains unclear how reliable these methods are and under what conditions they could be used. “They could provide an automated, easily accessible initial assessment,” said Körner. But since language models are trained on human-written texts, they also acquire biases, meaning distortions in judgment. It is therefore important to carefully examine whether personality assessments in sensitive areas should be left to algorithms, particularly from an ethical perspective.