Gosh, I really need to write the piece "watermarking AI content is technically infeasible," we can only watermark human verification, and only until the first meaningful edit.
Good piece as usual. Copied this comment from his version.
If you want good regulation about a topic that legislators don't deeply understand, you have to write it for them, or at least popularize a soundbite of necessary/key features that you demand.
Funny enough - I made an app to remove these watermarks.
This was done for photos out of Photoshop / Lightroom where a minor edit can lead your images marked as "Made with AI"
I guess I'll have to kill the app - as the bill appears to completely disallow any apps that can remove watermarks.
An interesting tidbit:
C2PA, as mentioned, one of the standards being used - they put out a report on some of the harms of using these watermarks. It's a fascinating read - privacy infringement, humans rights issues, etc.
Apparently, watermarking text, if it's long enough (a few hundred words), is not entirely infeasible: Scott Aaronson was working on such a scheme. That scheme is also robust to small changes to the text. It involves changing the sampler, though, not the model, so to demand such watermarking, "AI system" would need to be defined in a way that only applies to complete generative systems, not models in isolation (which is really how it ought to be defined, anyway). Regardless, it still seems like a bad idea to mandate it.
Gosh, I really need to write the piece "watermarking AI content is technically infeasible," we can only watermark human verification, and only until the first meaningful edit.
Good piece as usual. Copied this comment from his version.
If you want good regulation about a topic that legislators don't deeply understand, you have to write it for them, or at least popularize a soundbite of necessary/key features that you demand.
Funny enough - I made an app to remove these watermarks.
This was done for photos out of Photoshop / Lightroom where a minor edit can lead your images marked as "Made with AI"
I guess I'll have to kill the app - as the bill appears to completely disallow any apps that can remove watermarks.
An interesting tidbit:
C2PA, as mentioned, one of the standards being used - they put out a report on some of the harms of using these watermarks. It's a fascinating read - privacy infringement, humans rights issues, etc.
https://c2pa.org/specifications/specifications/1.0/security/_attachments/Initial_Adoption_Assessment.pdf
Seems we are slowly headed toward a DRM world - where you cant just post any JPEG you want.
Apparently, watermarking text, if it's long enough (a few hundred words), is not entirely infeasible: Scott Aaronson was working on such a scheme. That scheme is also robust to small changes to the text. It involves changing the sampler, though, not the model, so to demand such watermarking, "AI system" would need to be defined in a way that only applies to complete generative systems, not models in isolation (which is really how it ought to be defined, anyway). Regardless, it still seems like a bad idea to mandate it.