This is an article from 2024, when open weights models like llama were only beginning to emerge. With those you basically cannot reliably do any detection (as the authors admit by the end).

Which is really boiling down to text having statistically very similar properties to human generated one. Introduce a more motivated attacker and the text would be indistinguishable from real (with occasional typos, no use of "delve", "it's not x its y", emdashes and so on).

It really is a lost battle: you cannot embed extra information in the text that will survive even basic postprocessing (in contrast to, say, steganography)

▲

piperswe an hour ago | parent | next [-]

Ultimately it shouldn’t be too surprising that the machine that works by generating the most statistically likely text, generates text that’s statistically identical to human-generated text

	▲	userbinator 4 minutes ago \| parent \| next [-]
		I've never seen the word "delve" show up with such frequency in the pre-AI era, but now it's an overwhelmingly large signal of LLM-generated text, so I'm not sure where that came from. Ditto for vomiting emojis everywhere.
	▲	slopinthebag 7 minutes ago \| parent \| prev [-]
		I'm not so sure I buy that. AI written text is fairly obvious to good writers with exposure to LLM output. Is it a case where it's sort of an average of writing styles, but that average is not human and thus humans can detect it?

▲

nylonstrung 2 hours ago | parent | prev [-]

It sounds like a "cursed problem". Are there any contemporary techniques that show any promise?