Hacker News Reader

Comment by phyrex | Hacker News Reader

How would you detect that?

Detection of generated images works well, as reported by lots of works on arxiv. It's not catching them all, but lots of them. One way to understand it is that DDPMs and their underlying neural network do leave patterns in the high frequencies, often invisible to the eye. Deconvolutions typically generate low noise "marbles".

False.

And if true, is only temporary.

Any "tell" can be inferred by the descriminator which will the lead the generative process towards a work around

Once again, it's not quite as simple as that.

This debate did not start just last year, lol, it has been going on for years. There's a reason you can't just simply slap a discriminator on things and get away scot free.

It'll be an ongoing cat and mouse game for....as long as we can tell, probably.

(Also, the author mentioning the deconvolution checkerboard should be a dead giveaway that this debate has been raging for years, if you've had experience in the field over the last several years.)

It's not a cat and mouse game, it a mouse game. Detection only works up to the certain point where the difference is too small to tell reliably, and the false positive rate is unacceptably high. In fact we're already at this point.

"AI detection" is a scam. Already such detection in education has been treated like "the computer says your work is AI, so that's final and you're getting sanctioned" rather than "the computer alleges you are submitting AI-generated work, can you show your work process?". It unfairly targets neurodivergent individuals who sometimes have a tendency to write in such a way that is closer to what these AI detection tools claim is AI, than neurotypical individuals.

AI detectors will (intentionally and unintentionally) end up being used in heavily ableist ways. Once it's ingrained in the zeitgeist, then we're stuck with it until it causes a scandal so big that it can't be ignored, and in the meantime it will harm millions, if not billions, of people.

AI image detection, even when it does work, will have an extremely short shelf life, as the model continue to improve. Remember when SD couldn't even reliably render hands with five fingers? Go try some image generation on the state of the art models and inspect the hands, you'll find that there are far fewer impossible hands than there used to be. This will only improve over time. What do you do when the state of the art models eventually overcome all these issues? What about when they become capable of rendering flawless text? What then?

It's like people who claim that they can tell you with 100% certainty whether an image was photoshopped. Bullshit.

Reply on news.ycombinator.com