OpenAI just released a tool (web only, no API) for detecting AI-written text:
Conversation
Replying to
If the classifier is open to everyone, now folks who care about getting detected will just run it through it
Seems like a losing battle
2
7
Show replies
Replying to
Meh. :)
Quote Tweet
Hm. AI-text detection accuracy doesn’t seem so high out of the gate, and I’m guessing there will be a bunch of re-wording tools that will pop up to avoid detection further.
Seems like a bit of a fool’s errand? twitter.com/DeveloperHarri…
1
1
Replying to
Reminds me of halting problem. Also - an obvious accelerator for ever improved AI quality. Just use this machinery to anti-train your new models :)
Replying to
If the accuracy is good enough, a chrome ext that auto highlights/blocks gpt content when web surfing would be kinda interesting. Unsure what UI would look like
Replying to
OAI’s classifier “correctly identifies 26% of AI-written text (true positives) as “likely AI-written,” while incorrectly labeling human-written text as AI-written 9% of the time (false positives).” Why would you offer something unreliable to the public?
1
6
Replying to
*undocumented API. It's actually just the same ol' /completions but with a new (unlisted) model. The website checks logprobs on "!" token.
6
