4 ms·
Would be very interesting to require generative AI companies to log their created images for that purpose, it isn't gonna eliminate self hosted alternative, but
by Vilian 1y ago
Would be very interesting to require generative AI companies to log their created images for that purpose, it isn't gonna eliminate self hosted alternative, but it can destroy any fake evidence faster
- jdiff 1y agoI don't think this would solve the problem, unfortunately. AI can certainly be coerced into reproducing an existing image, leaving plenty of room for probable deniability.
- A_D_E_P_T 1y agoLogging them would be rather cost prohibitive, but images can be hashed + (invisibly) watermarked and video can be hashed frame by frame, in such a way that each frame authenticates the one before it. Surely there's a way to durably mark generated content.
- Iulioh 1y agoIt works untill someone takes a screenshoot of the image
- A_D_E_P_T 1y agoThe watermark would still be there. You can take a photograph of the image with an old Polaroid and, if the resolution is high enough, the watermark would still be there.
- brookst 1y agoIf it can be detected, it can be removed. Asking that false things carry claims of falseness is a dead end.
- A_D_E_P_T 1y agoLet's say you want to release a deepfake image. Your AI image, as generated, is not watermarked or hashed. Difficulty: Story-mode. You can release your image in two minutes flat. Let's now let's say your AI deepfake image is watermarked/hashed upon generation. Difficulty: Intermediate to hard. You might be able to get it done, but not without real effort, and removing all traces of the watermark without leaving artefacts might be so difficult as to be nearly impossible. ...So it doesn't eliminate the possibility of fakes, but it hugely raises the cost and effort bar. Only the most motivated would bother.
- brookst 1y agoGood thing the detection and removal can’t be automated and zero effort, eh?
- A_D_E_P_T 1y agoCan that be done today, in a "zero effort" manner, even with watermarked stock images? I tried some "AI watermark removers" on a Shutterstock image, and they did a pretty good job, but they left traces that would be obvious to any investigator.
- airstrike 1y agoIf an image has no watermark, we could distrust it outright The problem then is repeated compressing causing that to occur inadvertently
- bogtog 1y agoI think such watermarks are one of DeepMind's goals: https://deepmind.google/science/synthid/ https://deepmind.google/science/synthid/
- A_D_E_P_T 1y agoInteresting! I wonder how they watermark audio. The obvious way would be to do it in some frequency inaudible to humans (say 30kHz), but most conventional file formats can't handle that. (You'd probably need to make a modification that contains an additional ultra-high frequency.)
- JimDabell 1y agoThis rules out any form of self-hosted generative AI. It’s not going to work. It needs to be the other way around; we need to prove authenticity.