TBH, I’m cool with personal use LM-based translation as long as the training data is ethical sourced and the use of process disclosed, maybe with original message as context so it’s clear that it’s a translation rather than an outright generation.

After all, isn’t that what language models are made for? Modelling language?