It's a starting point is what I think they're getting at. Preclassification which a human then corrects--we're effectively talking about a labor saving device for an otherwise tedious task.
When I did "The Valley of Fear" as my first project, the PG text was used as the base, but if I encountered any kind of ambiguity in the text, I consulted at least a half-dozen other versions of the text via Google Books scans for agreement.
The team is also very particular about only using editions that have entered into the public domain. So if the first edition of a book just entered public domain, you must make sure that what you have produced only uses text from the first edition, and that you haven't inadvertently used a later edition as a base that may have included subsequent editorial changes.