4 ms·

> I wonder if GPT3 could be used to infer reacted text prompted by the surrounding article..? Generally, I would assume that the part of the text with the mos

> I wonder if GPT3 could be used to infer reacted text prompted by the surrounding article..? 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..

graderjs 6y ago
If the redacted part was somebody's name or a project codename, there's basically no way to infer that unless the system has had source text with that info. But I guess in the case where they are redacting something like a "method or procedure" that has been written about unredacted in other contexts and then its application in this context is classified, I think that would be possible. I'm sure people are already thinking about how to guard against this. 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.