And he’s correct. AI “detectors” as he describes are both wrong, useless, and doomed to fail.
But there are better ways to get hints. Specific models tend to overuse certain tokens: they’ll pick the same character names for a story, or overuse certain phrases, and if you play when them long enough you start to recognize “oh, that’s an OpenAI” model, or “this is dry like Nvidia Nemotron,” or whatever.
Yeah, that’s sad.
And he’s correct. AI “detectors” as he describes are both wrong, useless, and doomed to fail.
But there are better ways to get hints. Specific models tend to overuse certain tokens: they’ll pick the same character names for a story, or overuse certain phrases, and if you play when them long enough you start to recognize “oh, that’s an OpenAI” model, or “this is dry like Nvidia Nemotron,” or whatever.
See EQBench’s slop profiles: https://eqbench.com/
But, ultimately, the way to fix this is to go back to trusted sources and citations.
Back in the early days, my friend and I were surprised to find the same two part character names appearing each of our AI dungeon games.
Yeah, and it’s still an ongoing problem, sadly.