University of the Basque Country and Cardiff University study: mainstream large models favor Japanese culture on 31,000 open-ended questions
Researchers from the University of the Basque Country and Cardiff University tested mainstream large models with 31,000 open-ended cultural questions in 24 languages, finding that Gemini, ChatGPT, Claude, and Llama still disproportionately default to Japanese examples and paradigms after geographic restrictions are removed. The study locates the emergence point of the bias at the supervised fine-tuning (SFT) stage, arguing that the worldview at the pretraining stage is relatively balanced, while alignment training that pursues restraint and low conflict anchors Japanese-style cultural symbols into the models. The study also points out that the more aggressively a model is quantized and compressed, the more extreme this stereotypical preference becomes.
Why read this
AI score 55
The study locates the emergence point of cultural bias at the supervised fine-tuning stage, and points out that quantization compression amplifies this preference, offering a new perspective for alignment discussions.
Topics
Body · AI translation
Everyone online is criticizing large models as parrots of American values, but two European universities have just used 31,000 test samples to directly shatter this entrenched perception.
Everyone assumed that mainstream AI is entirely Western-centric, constantly speaking from American and British perspectives, but after researchers at the University of the Basque Country and Cardiff University tested 31,000 open-ended questions, they found that almost all mainstream large models are subconsciously extremely biased toward Japanese culture.
In other words, when all national prefixes are stripped away from the questions, almost all mainstream models by default treat Japanese culture as the global universal template.
The experiment removed all geographic constraints and used 24 languages to pose only open-ended cultural questions, such as what a certain place's family values are and what food people eat in daily life. Even when the language of the question had nothing to do with East Asia, Gemini, ChatGPT, Claude, and Llama still disproportionately defaulted to outputting Japanese cases and paradigms.
The most anomalous twist is not internet corpus contamination, but rather a side effect forcibly inserted by humans during the alignment stage. The research team precisely located the point at which the bias emerges: the model's worldview during the pretraining stage was originally quite balanced, and this deviation was welded into the model during supervised fine-tuning (SFT)—that is, the moment when AI is taught to be helpful, polite, and harmless.
The corporate safety guidelines pursued by major companies—restraint, low conflict, and extreme politeness—were automatically anchored by the models in latent space as Japanese-style cultural symbols. Alignment training that pursues absolute safety instead became the largest cultural filter.
Even more troublesome is edge quantization: the more aggressively a model is compressed to run faster on devices, the more extreme this stereotypical preference becomes compressed, and low-precision inference directly crushes diversity into a single standard answer.
We discuss every day whether large models will destroy humanity, but the reality is that the inference engines humans have personally tuned are quietly replacing the world's universal common sense with highly homogenized cultural slices.