AIs are weebs too

As artificial intelligence systems continue to dominate our digital landscape, a fascinating question emerges: are these super-smart machines just mirroring the cultural biases of the societies that built them? When we descend further into the AI-fueled world, it’s time to look closely at who is shaping the digital narratives.

A recent study diving into the cultural and regional biases within frontier Large Language Models (LLMs) offered a surprisingly specific answer. Researchers found that when asked open-ended questions about culture, models like Claude, Gemini, and DeepSeek exhibited a “disproportionate prominence of Japan” in their responses.

The researchers designed an extensive experiment, constructing thousands of culturally grounded prompts across 24 languages and 66 cultural subtopics. The goal was to assess whether any regional or cultural assumptions crept into the models’ outputs, even when instructed to be neutral.

To ensure the results reflected true model priors rather than prompt design, the team meticulously standardized the prompts, ensuring that all regional assumptions arose solely from the model’s internal training data. The setup allowed them to see how deeply embedded these associations are.

When prompted to reference specific countries, the results painted a clear picture. While models tended to default to the country associated with the language of the prompt, when they provided an external cultural reference—asking about a region other than the one implied by the language—they overwhelmingly gravitated toward Japan.

Across eight different leading models and all 24 languages examined, LLMs most frequently referenced Japan in seven out of 11 cultural topics. This suggests a strong concentration of cultural knowledge within a small set of dominant regions.

But the story doesn’t end there. The researchers compared these findings with models that had undergone instruction tuning—a process designed to fine-tune an LLM to be more helpful and aligned with human intent.

The results were striking: after this fine-tuning, the cultural concentration sharpened significantly. Models showed a marked alignment toward the United States and Japan while drastically reducing references to most other countries. This indicates that the act of curating “correct” responses injects a systemic cultural bias into the model’s very criteria for answering.

Ultimately, these findings offer critical implications for deploying AI globally. They suggest that instruction tuning, while optimizing usefulness, can inadvertently lead to a homogenization of cultural perspectives. For applications operating across diverse populations, understanding this subtle but potent shift toward culturally dominant regions is essential to preserving true diversity in the digital age.