I think that depends on how you calibrate the ship of Theseus: If someone uses an LLM to "proofread" and it suggests one word instead of another, does it stop being the author's work? Two words? Three? Switching the order of two sentences?

I think the key concern here is that a statistical "watermarking" system could insert a positive signal even when the gradual changes are--to any human watching the editing process--not things we would consider disqualifying.