It's interesting that someone could write an article about AI writing detectors without mentioning the stylistic cues that humans use to identify LLM output in practice, which are completely different from statistical methods like perplexity: em dash spam, overused patterns like "not just X, but Y", tendency towards making every single sentence sound like an earth-shattering mic-drop moment, et cetera.
It's interesting that someone could write an article about AI writing detectors without mentioning the stylistic cues that humans use to identify LLM output in practice, which are completely different from statistical methods like perplexity: em dash spam, overused patterns like "not just X, but Y", tendency towards making every single sentence sound like an earth-shattering mic-drop moment, et cetera.
A friend of mine who forwarded this to me thinks it's proof that the writers of the U.S. constitution were time travelers ;<)