Yeah, trivial text samples are obviously insufficient. At the other extreme (say millions of tokens of output), there's probably a reasonable chance that you co

Yeah, trivial text samples are obviously insufficient. At the other extreme (say millions of tokens of output), there's probably a reasonable chance that you could reliably distinguish human and LLM output. I suspect there might be some interesting questions down the rabbit hole in the middle though.