There is this frequently repeated claim that babies don't learn to speak a language from watching TV. People use this factoid to argue that grounding in interaction is crucial for acquiring a language for people (and maybe for machines too).

I think the factoid is probably true in the strict sense. But I do have some anecdotal observations that make me sceptical about *the amount* of such grounding that is needed.

A child I know was exposed to fair amounts of overheard adult-directed English speech since birth. She also regularly watched English-language children's cartoons.

Meanwhile she learned a couple of other languages in the usual way, without speaking any English, and showing little evidence of understanding it. Until at some point she very rapidly and spontaneously started speaking English and is at this point (~4 yo) reasonably fluent in it.

This makes me think that it is indeed possible to acquire a substantial amount of knowledge of language without grounding in interaction, and then, if you will, fine-tune this model to learn to use the language in question in the usual human way.