Comment by 0cf8612b2e1e

At the token, not word level, it would be possible for a Markov chain. It never has to know about Trump or XSS, only that it sees tokens like “ing”, “ed”, “is”, and so forth. Given a LLM size corpus, which will have ~all token-to-token pairs with some non-zero frequency, the above could be generated.

The actual probabilities will be terrible, but it is not impossible.