Why are all LLMs Obsessed with Japanese Culture? On the Hidden Cultural and Regional Biases of LLMs

The surprising part isn't that LLMs have cultural bias — that's well-established — it's that the bias runs counter to the expected direction. Most prior critique assumed English-language, Western-centric training data would produce Western-centric outputs, so finding Japanese cultural content systematically overrepresented complicates that tidy story and suggests training data composition interacts with language-specific fine-tuning in ways researchers haven't fully mapped.

This paper sits largely disconnected from recent Modelwire coverage, which has focused on commercial dynamics and developer tooling rather than evaluation methodology. The closest thread is the SemEval-2026 narrative similarity task (also from arXiv cs.CL, same date), which shares the same underlying concern: that benchmarks and datasets shape what behaviors we can even see in models. Both papers are essentially arguing that how you measure matters as much as what you measure. The cultural bias finding also has quiet relevance to the WIRED newsroom AI piece from April 17, where editorial voice and cultural defaults in AI-assisted writing were left largely unexamined.

Watch whether the dataset gets adopted as an evaluation component in any major multilingual model release in the next six months. If it does, that signals the field is taking output-level cultural auditing seriously rather than treating training data provenance as a sufficient proxy.