4 ms·
Given f(prompt, model) -> text is there some h(model, text) -> [0,1) which tells you if the the text was generated by the model? Crucially, could you publish h
by gorgoiler 3y ago
Given f(prompt, model) -> text is there some h(model, text) -> [0,1) which tells you if the the text was generated by the model? Crucially, could you publish h without publishing the model?
It’s a bit like public key cryptography — if generating text is using sign() to sign a prompt with your model then is there a publicly available verify() that verifies the output came from
the model but which doesn’t leak the private model itself?
What sort of things exist like this (other than private, API models recording all text they’ve generated)?
- jraph 3y agoInteresting mathematics problem aside. I think OpenAI sells a subscription to a ChatGPT detector. So you can pay for h. Does it work correctly? I don't know. Selling the poison and the antidote seems like a good business move though (I know, you asked "other than private API" - not sure they record the generated text, not sure they don't). Now, a ChatGPT-generated text (for current versions of ChatGPT) is more or less recognizable so for moderation purpose I would guess you don't really need h, you can smell the bullshit. It has a specific way to be overconfident and it feels like it's giving you a lesson in a specific impersonal way without emotions. Something like this. I have the same kind of feeling when reading a WikiHow page, WikiHow has a very specific and recognizable style to explain things. I guess you can recognize patterns / specific behaviors on accounts posting ChatGPT texts too, which can help for particularly short texts.
- timmaxw 3y agoScott Aaronson has worked on this with OpenAI. Ctrl-F for "watermarking" in https://scottaaronson.blog/?p=6823 https://scottaaronson.blog/?p=6823. I don't know if they've actually deployed it in ChatGPT or not.