I think a problem with this is that if you're at all familiar with how someone speaks or writes you can figure out who they are rather easily.
So there's another weird problem (probably not so hard though): translating conversation within the same language in a way that doesn't lose meaning but obscures identifying patterns.
This was a problem that I briefly worked on it undergrad. Turns out authorship attribution is an active sub field in comp ling. So you’re not just trying to obscure identity from human readers, but the neural nets looking through all your public writing as well.
For the interested some paper keywords would be adversarial stylometry and authorship attribution.
Or multiple translations. This is a great idea that I’ve considered but want to leave for V2. V1 is maybe disappoingly simpler, but at least it would mean that your boss 10 years from now isn’t reading through your post history
So there's another weird problem (probably not so hard though): translating conversation within the same language in a way that doesn't lose meaning but obscures identifying patterns.