The author reports that AI-generated Japanese remains detectable through recurring word choices, unlike AI-generated English, which has become harder to identify. The author notes that model choice matters more in non-English languages. Anthropic's Opus 4.6 and Fable 5 produce reasonably good long-form Japanese outside coding sessions, while Google's models handle a wider range of registers. OpenAI's models, Opus 4.8/5, and Fable in coding sessions sound overly geeky in Japanese.
No score is assigned. Sources and their independence are shown in the citation chain below.