You can't. You can only evaluate coherency, which will gradually improve as LLMs get more sophisticated. You can attempt to subvert these technologies or litigate them, but there are only a few ways to be ~100% certain that AI was used.