Languages have all sorts of weirdness. Transliteration is usually not very accurate as turning phrases and decoding idioms literally might not make sense. Literal "Darmok and Jalad at Tanagra" level stuff.

So useful translating applications have models to keep the nonsense from emerging from the output. In the old days it was heuristic models. These days I don't know how it works but if it's going GAN-based, the training of the system would be around what needs to be avoided from the output.

The difference between oddball phrases that shouldn't be transmitted literally and inconvenient facts that must not be communicated is merely a few points in a data store.