Cultural bias or universal traits? Exploring personality-like profiles of large language models

Abstract

Do large language models (LLMs) exhibit coherent personality profiles, and are these profiles culturally neutral? This study used two validated personality inventories, the Big Five Inventory–2 and the Dark Triad Dirty Dozen, to assess five state-of-the-art LLMs (ChatGPT-4o, Gemini 2.0 Flash, DeepSeek-V3, Ernie 3.5 and Le Chat) across English, Mandarin and French. Models were evaluated at both domain and facet levels for consistency, linguistic variation and cultural alignment by comparing their responses to human data from the USA, People's Republic of China and France. Results showed that LLMs consistently exhibited prosocial profiles (high Agreeableness and Conscientiousness, low Negative Emotionality and Dark Triad traits). However, personality expression varied with both the language of assessment and the cultural origin of the model. In general, Western models aligned more closely with English inputs, while the Chinese models displayed stronger traits in Mandarin. No model fully mirrored its respective human population, and cultural divergence was especially pronounced in the Chinese models. These findings suggest that LLM personality-like response profiles are relatively stable but vary systematically with assessment language and model development context, raising questions about fairness, neutrality and the psychological imprint of artificial intelligence systems.