Generally, I would assume that the part of the text with the most entropy (redacted nouns and their adjectives) would be especially hard for a system like GPT3 to generate. There might be some clues in the surrounding texts, but perhaps someone would need to train it specifically on similar (un-redacted) documents?
Also, the other way around, marking parts as viable-for-redaction and then having GPT3 generate text to fill in those gaps would be a really insidious way of redacting content. I'm guessing a GAN-type network could even generate the visuals that are appropriate to the document that has been redacted..
Your second paragraph is interesting. It's basically an automatic disinformation generator. I think there's almost more security value in releasing a document like that with faked interpolated text standing in for redacted sections, and then a few blacked out sections to add to intrigue and credibility, than in doing a regular redacted document...but maybe there's benefits of the old way I don't understand. Not sure if I got what you were saying there, but I think I got your meaning correctly.
But for something like this which is sort of a semi-serious report I kind of have the feeling maybe they're just doing some of these as spurious redactions for the laughs.