Tried Japanese: it's better than Google Translator, but almost equally useless.
Perhaps the training corpus is too small (not very likely).
Perhaps the edit distance from JP to IT via EN is too big?
Perhaps the used pattern recognition algorithms aren't suited for character-based languages?
All three together I think, but especially with long sentences. In the end, it takes longer than translating from scratch with the help of manual electronic dictionaries.
The next CTE update will include the language changes in DeepL API.
Good to hear!
Seldom. Life's too short for that.