I wonder if this will turn out to be something of an antipattern (for lack of a better term). The set of vocabulary in a given language is fairly limited by the human mind, enough such that it's relatively uniform enough that subtle alterations are achievable enough to the individual that they could conceivably, of their own volition, intentionally fool emergent AIs leading to false attribution. I mean, if the AI can alter the text subtly enough to fool other AIs, given the limited number of words known to other humans (which is kind of the point of speech - what use is a massive vocabulary if nobody else knows those words?), why couldn't humans do it too?

So, here we have AIs erroneously pinning the author of heretical text X as human Y, thus triggering punishment Z. All the while, it was really written by bastard A, who is really a member of political party B: the opposition of human Y's party, the C's.