AI Models Trained in China Prioritize English Too

Why this is here: AI models developed in China perform nearly identically to Western models on languages like French and German, despite the potential to prioritize Chinese dialects.

Researchers at an unnamed institution in China compared the language skills of AI models developed there with those created in the West. They examined 21 languages, including various Chinese dialects, European languages, and other Asian regional languages.

The study found Chinese-built AI models perform similarly to Western models across most languages. Mandarin Chinese is the one area where they notably excel.

The team used tests of information recall and reading comprehension to assess performance. They discovered a strong connection—a correlation of 0.93—between how Chinese and Western AI models handle most languages. This suggests shared training data and a focus on globally recognized benchmarks.

However, Chinese models sometimes struggle to correctly identify languages spoken by minority groups within China, such as Kazakh and Uyghur. This indicates a potential trade-off between supporting linguistic diversity within China and optimizing for English-focused international standards.

The research was limited to open-weight large language models. Further study could explore proprietary models and a wider range of languages.