OK, I agree, translation is useful and is fundamentally something it makes sense for.
Disagree about “arbitrary”, you need a huge enough dataset for every language, and it’s not going to be that much better than 90’s machine translators though.
It’s closer, but in practice still requires a human to check the whole text. Which raises the question, why use it at all instead of a machine translator with more modest requirements.
And also this may poison smaller languages with translation artifacts becoming norm. Calque is one thing, here one can expect stuff of the “medieval monks mixing up Armorica and Armenia” kind (I fucking hate those of Armenians still perpetuating that single known mistake), only better masqueraded.