LLMs can change their "voice" with a simple instruction. If detecting LLM-generated text was as easy as you think it is then services in this space wouldn't suffer any false-positives or any false-negatives.
Oh I don't think it's easy or that one can be completely sure about it. And in my experience, LLMs aren't very good at changing their voice. If you have any examples to the contrary, I'd like to see em.
I don't have examples, I'm asking you to provide evidence which supports your extraordinary claim.
You claimed, with a high level of confidence, that the text isn't written by AI because it lacks obvious "tells" which you believe to be present in any LLM generated text. But if the absence of these "tells" reliably indicated human writing then LLM detectors would have false negative rate of approximately 0%, do they?
It's not necessary for the tell to be easily computable. It's not something that's true everywhere, but it has been true more often than not when I've tried to use LLMs. And I haven't seen a LLM-generated piece of writing that's sufficiently long and complicated enough to rule this tell out.