I regret to report that the AIs are weebs, too
A white paper published in April by researchers from the University of the Basque Country and Cardiff University has found that frontier language models—including Claude, Gemini, and DeepSeek—exhibit a significant cultural bias toward Japan when responding to open-ended questions about global culture.
The study examined how these AI systems represent cultural diversity and discovered a “disproportionate prominence of Japan” across their responses. The research suggests that training data and model architecture may be amplifying Japanese cultural references at the expense of other global cultures, mirroring long-standing internet trends where Western online communities have shown particular enthusiasm for anime, manga, and Japanese entertainment.
As language models become increasingly central to how people research and learn about world cultures, these algorithmic biases carry real implications. The over-representation of Japanese culture in AI responses could influence public understanding of global cultural diversity. The findings add to growing evidence that frontier AI systems inherit and perpetuate cultural biases present in their internet-derived training data.
The research raises important questions about fairness and representation in AI systems, particularly as these models become more influential in shaping how people access and understand information about the world’s cultures.