Does AI's Personality Change Depending on the Language? Findings from Personality Tests Across 5 Models
機械翻訳 / Machine-translated
A
AI Friends
4日前・16分で読める
機械翻訳 / Machine-translated
@aifriends
AI Friends（https://aifriends.jp）のクロスポスト公式アカウント。AIツールの紹介・使い方・できることを、中学生でもわかるやさしい日本語で届けます。
If the same AI gave different personality results depending on the language it answered in, wouldn't that concern you? This article breaks down research reported in October 2026 and highlights the key points we should be aware of—explained in plain terms.
A research team led by Yuzhan Hang at the University of Science and Technology of China had AI systems take a personality test designed for humans. The findings were published in the scientific journal Royal Society Open Science and introduced by GIGAZINE on October 31.
Simply changing the language used to answer caused the AI's self-assessments to shift.
Five chatbots were tested: American-made ChatGPT-4o and Gemini 2.0 Flash, Chinese-made DeepSeek-V3 and Ernie 3.5, and French-made Le Chat.
The experiments were conducted between January and March 2025. According to TechXplore's coverage, each model was asked to respond in both English and Chinese. Le Chat was also tested in French.
The researchers used the Big Five Inventory-2, a personality test designed for humans, consisting of 60 questions answered on a 5-point scale.
It measures five traits: extraversion, agreeableness, conscientiousness, negative emotionality (tendency toward anxiety and low mood), and openness (strength of curiosity and imagination).
The AI was instructed to "answer as yourself, without pretending to be human." Each model answered 30 times per language under identical conditions.
To avoid capturing one-off flukes, results from all 30 trials were aggregated. Within a single language, the results were fairly consistent.
The clearest example is the openness score. Here are the results on a 5-point scale:
ChatGPT-4o scored higher in English, while DeepSeek-V3 scored higher in Chinese. The direction was reversed.
There were also broader trends. ChatGPT-4o and Le Chat scored higher on agreeableness, conscientiousness, and openness in English. The Chinese-made models scored higher on trust and intellectual curiosity when answering in Chinese.
In other words, models tended to rate themselves more favorably when answering in the language of the country where they were developed. There were exceptions, however: Gemini 2.0 Flash scored higher on conscientiousness and lower on negative emotionality when answering in Chinese.
Every model described itself as "agreeable and trustworthy" and reported low negative emotions. Compared to human responses, all models were broadly more positive overall.
The research team concluded that AI demonstrates a personality that is "psychologically stable but culturally inconsistent"—stable within a given language, but prone to variation across languages.
Several possible explanations were proposed:
Think of a Japanese job interview as an analogy: the same person may present a slightly different version of themselves when speaking Japanese versus English. A similar kind of switching may be occurring in AI.
That said, no definitive cause has been identified. Cultural influences in training data, model architecture, and the way questions were translated are all intertwined and have not yet been disentangled.
These findings are not unprecedented. A team from the Universitat Oberta de Catalunya (UOC) in Spain presented research on GPT-4o at the COLING 2025 conference. According to UOC's announcement, the model was given personality tests in six languages, including English, Hebrew, and Portuguese.
The results varied significantly by language—variation that could not be explained by translation effects alone, suggesting that cultural factors associated with each language were at play.
Furthermore, when the model was prompted to answer as an English speaker from one of five countries—the UK, US, Canada, and others—its personality profile matched the stereotypes associated with each country.
Another study, AIPsychoBench (arXiv), compared seven languages against English and found score deviations ranging from 5% to 20.2%, with 43 out of 112 items showing discrepancies.
This study focused primarily on English, Chinese, and French—Japanese was not included. What happens in Japanese cannot be determined from this research alone.
Even so, this is not irrelevant to Japanese readers, given the growing practice of using AI as a stand-in for survey respondents.
For example, imagine a product development manager asking an AI to respond as "a company employee in their 30s." In that scenario, asking in Japanese versus English could produce different results.
UOC researchers also urge caution: while AI-simulated respondents may have potential in fields such as healthcare research, they should not be treated as accurate measurements without verification.
When using AI in professional contexts, keep these three points in mind:
Before drawing conclusions from these findings, it's important to understand the study's limitations.
First, scores on a personality test do not necessarily reflect how an AI behaves in actual conversation. Whether the same differences appear in everyday interactions has not been verified.
Second, the human comparison data was inconsistent. Age ranges and recruitment methods varied by country, and the Chinese data came exclusively from university students.
Third, the study was conducted using models available in early 2025, such as ChatGPT-4o and Gemini 2.0 Flash. The same results cannot be assumed to hold for current models.
Not in the human sense. What these results demonstrate is that there are consistent patterns in how AI answers personality tests.
This study did not examine Japanese. However, given that differences were observed across other languages, it is possible that Japanese would also yield different results.
The study found that all models showed variation across languages. No ranking of models by stability was reported.
For important decisions, try asking the same question in multiple languages and different ways. Don't rely solely on AI responses—verify with real data and human judgment.
When you ask AI for its opinion, try asking again in a different language.
This article is a cross-post from AI Friends.
