> What if I write a program that does that translation agnostically of the specific work being translated?

Then that program will likely produce non-copyrighted output for non-copyrighted input. But it's also likely going to sound pretty bad.

> And what if my program is parametrized by an ML model of unknown training data?

Then its output could be considered a derivative work of every work in that training corpus, and anyone whose creative work went into that training data will have a copyright claim on the results.

> Then its output could be considered a derivative work of every work in that training corpus, and anyone whose creative work went into that training data will have a copyright claim on the results.

Depends how tech savvy the judge is. An ML model is not really like a copy or reproduction, it's much more like a person's brain in that it aims to take general patterns from various sources.