多語言訓練數據分佈不均

Risk Description

LLM training data is predominantly English (typically 80%+), causing significantly degraded performance in non-English languages. Understanding of Chinese, Taiwanese, Hakka and other languages may exhibit cultural misinterpretation and semantic errors. Low-resource language users cannot access AI services of equal quality to English users, exacerbating global digital inequality.