I regret to report that the AIs are weebs, too (www.pcgamer.com)

A recent study by researchers from the University of the Basque Country and Cardiff University revealed that frontier language models (LLMs) like Claude, Gemini, and DeepSeek exhibit a notable bias towards Japanese culture when responding to open-ended cultural questions. In a multilingual analysis involving 31,680 culturally grounded prompts across 24 languages, Japan emerged as the favored reference point for LLMs when answering questions about culture, even when they were instructed to choose examples from regions unrelated to the language of the prompt. This trend persisted despite a general preference among the models to reference the country associated with the language used. The implications of these findings are significant for the AI/ML community. The research indicates that instruction tuning—the process of refining models to improve responses—tends to homogenize cultural representations, favoring responses from culturally dominant regions like the United States and Japan while reducing the diversity of cultural perspectives in model outputs. This pattern highlights a potential limitation for LLMs in global applications, where diverse cultural viewpoints are essential, raising concerns about the broader impacts of cultural bias in AI systems and their alignment with user expectations in multicultural contexts.