AI needs good training sets - that is large corpuses of data about a domain, which have been accurately labelled so the AI can learn safely. Curated content basically. Image recognition software is suffering from this as all the large corpuses are basically the same three or so.
When we move into the realm of knowledge about the human world, economics, politics etc. then having a corpus is easy. Just download reddit, twitter and 4chan. But having a labelled corpus is really hard and trusting one. I mean 4chan might be able to teach an AI but you would not want it to date your sister afterwards.
Journalism is basically a means of labelling the corpus of "everything" - and it is only trusted because "most people" trust what WaPo or NYT say. And the new generation of non traditional journalists (err, Captain Disillusion?) are doing something similar.
But there is no real means to decide what is a correctly labelled corpus or what is not. We have a long process of discussing the hard stuff - from flame wars to editorial letters to courts and scientific papers.
Ultimately if we think AI is going to help us manage the flood of information (and I am not sure how else) we need to put effort into training it - and that basically means journalism. Proper Nournalism.
It's not Facebooks job to decide truth. But is it the governments job? Is it everyone's? how do we build this labelled corpus at scale public ally and with trust and verification?