Hey, thanks I learned something (I think). Computer shit isn’t my area of nerdery, but from what I gather AI images gets stamped with invisible water marks and you can detect these with software (rather than AI) if that software knows what the watermark is (which is kept secret by the AI provider)
that’s exactly right! it’s kept secret because if someone figure out exactly how it’s done it becomes easy to remove. so it’s security through obscurity, but it’s what we have unfortunately. (and even today, anyone can generate images with an open-weight model, which will not have any watermarks).
So if it’s been steganographied with something the tool you use doesn’t check for, you’d never know. If someone’s post-processed the picture enough, the steganography will be destroyed and you’d never know. If the image was generated with a model that never embedded anything, you’d never know.
So as ever, AI detectors are unreliable snake oil and should not be mentioned except to point out how unreliable they are.
Is it though? It looks like shop to me. Look at the neck/body connection
I tried reverse image search of just the body to see if there was a “sexy toilet” costume. No luck, so I’m suspicious.
I did find this though…
see it for yourself: https://synthid.com/ https://openai.com/research/verify/
I tried but it didn’t work. Kept getting hung up when I tired to sign in.
Also doesn’t using AI to detect AI defeat the purpose of screening for AI?
the openai one does not require signin.
it’s not AI. it’s a kind of hidden watermark or steganography called SynthID
see also: https://en.wikipedia.org/wiki/AI_watermarking, although this one is about text watermarking.
Hey, thanks I learned something (I think). Computer shit isn’t my area of nerdery, but from what I gather AI images gets stamped with invisible water marks and you can detect these with software (rather than AI) if that software knows what the watermark is (which is kept secret by the AI provider)
that’s exactly right! it’s kept secret because if someone figure out exactly how it’s done it becomes easy to remove. so it’s security through obscurity, but it’s what we have unfortunately. (and even today, anyone can generate images with an open-weight model, which will not have any watermarks).
So if it’s been steganographied with something the tool you use doesn’t check for, you’d never know. If someone’s post-processed the picture enough, the steganography will be destroyed and you’d never know. If the image was generated with a model that never embedded anything, you’d never know.
So as ever, AI detectors are unreliable snake oil and should not be mentioned except to point out how unreliable they are.
Nothings perfect. I dont think anyone’s going that far out of their way to avoid detection with a sexy skibidi costume picture.
“That far”?
Never mind.