I don't think anyone would claim that it does, except maybe Google Marketing and a few blue-eyed computational linguists - who would probably do themselves a favour by focusing a little more on the linguistics of the endeavour.

It is a text-to-text machine translation system. There is, as far as I know, no attempt at an internal representation of anything like semantics or knowledge about the world. They may have added some grammatical facts, but there is no model of the world and no notion of time like a person would use to make or understand statements in natural language.

Case in point: The artifacts of the training processes behind machine translation systems like Google Translate cannot be used to generate sentences from pictures or notions or anything like that. What is happening is nothing like a human learning a language, or even a pair of languages.

That's not to say it's not impressive, useful and important tech, mind, but it really does not do all of what it appears to do at first glance.