I was wondering if there’s any Blockchain project that could be integrated with and fingerprint all GenAI content. Every Blockchain token node could be linked to an image, or video produced in their meta data. And companies could verify if the generated text appearing on their platforms is AI or not.
I think with government regulations we could reduce the number of scams, and misinfo.
Though I do understand it’s limitations and that it would not be super difficult getting past such restrictions if the party were willing enough.
And it wouldn’t work for text, and code I suppose.
But it would be very helpful for indentify AI content much more easily.
Privacy could be maintained as it is with bitcoin and crypto.
But I wonder how you would get everyone to integrate such a system. Both the governments and the companies.
I remember reading about this discussion in some thread months ago, but I can’t find it.
I didn’t know where to post this, but I assume everyone on lemmy is some kind of wiz, so even in this subreddit someone might have the answer.
fingerprint all GenAI content
People can generate a billion images every seconds. Also what about images generated locally and not put in that blockchain?
Once again, the blockchain is a solution looking for a problem to solve.
Yeah all this will do is increase the power consumption of ai again.
No no. I have nothing to do with Blockchain. I barely understand it.
I was just thinking about methods of mitigating harm from misinfo and scam.
And since Blockchain provides privacy but at the same time provides some information to the public I thought it might have worked.
There is no privacy in a blockchain. Its meant to share information.
You cannot fix society with technology, the end.
Hey, what if we tried to cure cancer by mixing it with aggressive herpes?
deleted by creator
How would you enforce the entry of all generated AI content into said blockchain? Big providers might be persuaded or forced to participate, but it’s not like someone - specifically well-funded bad actors - couldn’t just set up their own servers at will.
Ahh yes, my intention was for such tech to make it difficult and mitigate harm not entirely stop it
If something lacked the fingerprint, it would immediately be suspicious.
But yeah I don’t completely understand the tech, so I want to know if there’s any such project or research on the same.
Block chain is an interesting form of fingerprinting that allows privacy for actors but not complete privacy.
If something lacked the fingerprint, it would immediately be suspicious.
Suspicious of what? Being non-AI?
It would be another variable to work with when trying to distinguish if the image/video is genAI or not.
People are already witch-hunting real artists because they think their work is too good to be true or whatever. You’d be giving them even more ammunition, putting real art under general suspicion.
I think you’d have better luck doing it the other way around: fingerprint known non-AI content, and treat everything else as potential AI.
I did think about that. That was the original idea.
But I thought it was less feasible to get everyone on board for it than just the AI gen companies.
But again, I don’t know enough about it. I was wondering if there are any projects try to do something like this.
Exactly. Use blockchain to trace image/video origins - ie media data plus camera signature and camera operator id with timestamps, and track the sources through every edit, along with edit transform info and editor id. That way any produced clip can be traced back to origins through all edits and then verified to particular real sources. A provable chain of custody of every image. If a video clip came from a trusted known camera operator and stays under the control of entities you trust, you can trust the video. If any links in the ownership or edits are sus, an author can track down and produce the in/out of any edit link to prove the authenticity.
Oh damn. It’s starting to sound like something that can be used to justify complete control over people’s lives.
Governments and corporates might use the “security reasons” reason to get even more info and control over us.
Yeah that was not my intention.
Anything that relies on external factors for flagging AI images is doomed to fail.
There will quickly be tools created to strip the metadata from the image, or to create AI pictures but skipping adding the blockchain stuff
Images, text etc can be generated entirely offline and independently. There is nothing to force the image to be attached to the block chain either directly or as a fingerprint.
You would have to do the opposite: when you take a picture or video (or write some text?), as it is recorded, the camera chipset signs the image/video using TPM-esque hardware, proving (ish) that it was captured by a real camera sensor.
The issue is that it’s pretty close to mandatory doxxing.
The point was to mitigate the harm I suppose, not entirely stop it.
As for the mandatory dpxxing, you’d just be registering on the block chain, but wouldn’t give away any other information.
So the only info that you give away is - whether something is real or not.
Are you talking about the AI generator registering on the blockchain? Because there is essentially no incentive for them to do so and every incentive for them not to.
If you mean genuine camera images being registered on the blockchain, that would give away at minimum the time the image was taken, and probably what kind of device it was taken with and all other images taken by the same user. That’s a lot of data.
Yeah I guess it is not feasible.
And you can’t ever post process images.
Theoretically you could include the original signed unprocessed image (or make it available NFT-style) and let the viewer decide whether the difference to post-processed image is reasonable or unreasonable.
It would however make it impossible to partially censor images without giving away the non-AI proof, unless you had a trusted third party ™ verify the original and re-sign the censored version.
A ‘view cryptographically signed original’ button next to every instagram post would be complete LOL, though.
That seriously can’t ever work on the technical level. You can simply remove metadata from an image.
To be able to even remotely be able to do what you want, you’d have to create a new image format where this is unremovable and make it impossible to convert such images to other formats.
That would probably need to involve DRM, I don’t think it’s possible without that. And still there would be people who’d either crack your DRM or simply caught the direct output of the graphics card and they’ll have your image in an editable format.
In short it’s impossible to do.
The point is that any unsigned image is assumed to be AI generated. You can absolutely strip the metadata or convert it to some other format (there’s always the analog hole and it has to become a bitmap to be displayed) but then you’ve lost the proof you took it.
You’d still need secure key storage hardware and trust roots in the camera like TPMs but every phone has that already…
(This is referring to the ‘signed in camera’ model)
