Looking for noise generation tech to protect IP from A.I

It is simply not possible nor worth the effort. Beware of people who want to sell you snake oil, which are aplenty in the AI business.

The thing is, as far as I can tell, most cases of generative models reproducing something 1:1 was something which was very likely to occur many times in the dataset, like stock images or passages from well-known newspapers.

Technically - the image generation models are already trained as denoisers and so you can't effectively thwart them without compromising perceptual quality of protected media. "Protection" presented so far relied on some quirks of particular architectures which aren't bound to occur in newer releases.

Eventually, it'll become almost universally true that platforms will require you to allow them to use/sell your data for AI. Anti-AI training tech doesn't work because it either affects the quality of the data to humans, or it can be easily removed before it goes into the training dataset. I think we'll eventually have models that aren't affected by data poisoning, making the efforts doubly pointless.

bullfrogs want wings, so they don't bump they little butts when they hop.

I find it interesting that seemingly the same people willing to embrace "When buying isn't owning, copying isn't stealing" approaches to Intellectual Property are worried about protecting their IP from AI.

Right now "AI" is pretty much "Big Tech / Big Media" bad guys, sure. Anyone suspect that's the way it stays? "Model training is too expensive" so no one will have home built or shared, sharded, federated public model building?

Certainly not, if there's no legal way to train without massive involvement in the universe of IP law.

Will "publication" come to mean more than "performance" and include some compulsory licensing scheme? Will we need to pin flags to our lapels so that everyone can know the legal status of our every utterance?

https://news.ycombinator.com/item?id=39406709