We present the first quantitative analysis of excess vocabulary in Japanese AI-generated text. Using 350 samples from 7 LLMs compared against 977 human articles from Qiita and Zenn (2020-2026), we identified 651 statistically significant Japanese excess words. Key findings: (1) Japanese excess vocabulary is distinct from English; (2) newer Claude generations show increasing excess scores; (3) 68% of top AI excess words show increased frequency in post-LLM human articles; (4) cross-domain classifier evaluation reveals complete transfer failure.
Ken Imoto (Thu,) studied this question.