This does not have to be true in order to translate languages.
But in the case of current programming languages, they all compile to the same standard instruction set architectures following the same basic laws of boolean logic, so it should not be surprising that PLs are translatable. Likely, the model has discovered some basic statistical representation of those things.
I'm unclear on what doesn't have to be true?
I intended to observe that a system capable of recognizing idioms in multiple languages, built on the large training sets of those languages' existing code, suggests the possibility of idiom-level combinational synthesis of a novel language. Most easily as a course-grain polyglot (combining large chunks of such code), and increasingly speculatively as grain size decreases (eg different languages in the same file, or function, or expression), or as language options are pruned (a manticore - representational redundancy is removed as a feature present in multiple languages becomes only available in some/one of its forms).