It turns out that's because ICU uses a combined Chinese/Japanese dictionary instead of separate dictionaries for each language. Which probably is a little more robust if you misdetect some Chinese text as Japanese and vice-versa.