@Micru: of course, as you say, machine learning is the elephant in the room.
I dream of something we could call "Wikisource as a platform":
meaning an environment with structured data and workflows where you can have APIs
and tools for interact with humans and machines, both for input and for output.
We could have OCR software that learn from our human proofreaders, and ideally we could
even have OCRs tailored for determined centuries or types of books.
We could ue machine learning to look for citations within books (for example other cited books or authors).¹
This could improve heavily our library:
on Internet Archive or Google Books we have millions of books that just wait for us to make them
readable and accessible, and, of course, connect them to Wikipedia, to Wikidata, to other Wikisource books.
IMHO, this is obviously important for GLAMs:
we could be much more usable and easy for libraries, archives and museums that want to upload into Wikisource their texts and books, and make them part of our hyperlinked library.