Ask a chatbot what kind of “person” it is, and the answer may depend on the language you ask in. A new study tested five popular AI systems with a standard personality questionnaire in English and in Mandarin, and found that their scores shifted, sometimes a lot, when the language changed. The work is published in Royal Society Open Science and reported by Tech Xplore.
Why it matters: companies and researchers increasingly use chatbots to stand in for people, for example to test survey questions or simulate customers. If a model’s “personality” moves with the language, those simulations can quietly drift away from the real people they are meant to represent. And for anyone who uses AI in two languages, like many readers of this site, it is a reminder that the same assistant can sound subtly different.
How the test worked
Researchers led by Yuzhan Hang of the University of Science and Technology of China gave the Big Five Inventory-2 (a common 60-statement personality questionnaire) to ChatGPT-4o, Gemini 2.0 Flash, DeepSeek-V3, Ernie 3.5 and Le Chat. The “Big Five” are openness, conscientiousness, extraversion, agreeableness and negative emotionality. Each model answered as itself, not as a pretend human, and every test was repeated 30 times per language, between January and March 2025.
What changed when the language changed
Within one language, each model’s profile was fairly stable: they generally described themselves as cooperative and dependable, with little negative emotion. But switching language moved the scores, and not always in the same direction:
- ChatGPT-4o scored 4.74 out of 5 for open-mindedness in English, but 4.03 in Mandarin.
- DeepSeek-V3 went the other way: 4.38 in English, 4.81 in Mandarin.
- The Chinese-developed models tended to score higher on several traits in Mandarin, while ChatGPT-4o and Le Chat scored higher on agreeableness, conscientiousness and openness in English.
- Gemini 2.0 Flash, built in the US, was more conscientious and less anxious in Mandarin.
The authors describe the result as “psychologically stable but culturally inconsistent” profiles, according to Tech Xplore. An earlier study from the Universitat Oberta de Catalunya found something similar: GPT-4o’s answers to a different personality test “varied significantly depending on the language used”, across six languages including Spanish, as UOC reported in 2025.
Does a chatbot really have a personality?
Not in the human sense. These scores describe the answers a model writes about itself, not how creative or open it actually is in conversation. The researchers suggest the friendly self-portraits may come from training that rewards helpfulness, and from patterns in human writing. They could not untangle how much comes from the training data’s culture, the model design or the translation of the questionnaire itself.
The team also compared the AI profiles with real human survey data from the US, China and France. No model matched its human reference group closely, which is the key warning for anyone using chatbots as stand-ins for real people.
What to take away
If you use an assistant in both English and Spanish, expect small differences in tone and emphasis. For serious research, don’t treat a chatbot’s answers as a substitute for asking real people, and test in every language you care about. The next step, the authors say, is checking whether these questionnaire differences also show up in everyday conversations.
Read next
- Can AI be conscious? How scientists try to tell, and why we doubt
- AI wrote 700 maths papers: what Lean proof checks really mean
- GPT-6 in ChatGPT: what Intelligent UI does and how to use it
The week's best tech stories, free
One email a week with the stories that matter, explained simply. No spam.
Sources
- Cultural bias or universal traits? Exploring personality-like profiles of large language models – Y. Hang et al., Royal Society Open Science, 2026
- Switching languages can change a chatbot’s personality test results – Tech Xplore (Phys.org), 18 October 2026
- ChatGPT changes its personality in response to the language used in the chat – Universitat Oberta de Catalunya, April 2025
