Comment by rincebrain

The simplest example that comes to mind of something frequency analysis might be useful for would be if you had simple ciphertext where you knew that the characters probably 1:1 mapped, but you didn't know anything about how.

It could also be useful for guessing whether someone might have been trying to do some kind of steganographic or additional encoding in their work, by telling you how abnormal compared to how many people write it is that someone happened to choose a very unusual construction in their work, or whether it's unlikely that two people chose the same unusual construction by coincidence or plagiarism.

You might also find statistical models interesting for things like noticing patterns in people for whom English or others are not their first language, and when they choose different constructions more often than speakers for whom it was their first language.

I'm not saying you can't use an LLM to do some or all of these, but they also have something of a scalar attached to them of how unusual the conclusion is - e.g. "I have never seen this construction of words in 50 million lines of text" versus "Yes, that's natural.", which can be useful for trying to inform how close to the noise floor the answer is, even ignoring the prospect of hallucinations.