Currently it seems like the proofread mode is just doing English BUT if the text is obv another language, it'll do it properly.
The problem is with edge cases where you have similar languages and the prompt for proofread probably doesn't take language into account much.
so for example:
https://translate.kagi.com/?text=physioclogisch%20water%201L&target=proofread
it's too close to english so even though that the Translate ui detects it as Dutch correctly, the actual llm api call doesn't contain 'dutch', so either the prompt should be adjusted to do better language detection OR you could just take the correct language detection and insert it into the prompt
