Obviously this is pure speculation, but I suspect it'll be a lot like cloning someone's voice. The first systems required the participant to read thousands of lines of text. Now you can do it over a phone call without the participant noticing that you're cloning their voice.

Why would bilingualism be a defense? Try throwing ChatGPT or whatever a line which combines several languages. e.g., If you ask it “¿Quel est the resultado of 四*6.1, in inglese?”, it will probably reply “4 times 6.1 equals 24.4.” AI doesn’t think in words, but expects bits of information to be related to X meaning, and from there it can predict a reply and translate that information into words we understand. I haven’t tried the above line, but if it’s not possible already there’s nothing really stopping us from making it work soon enough.