> If these things don't need to be trained on the work of others for whom they don't have consent and have not paid a license fee to, then why are they? You ma

> If these things don't need to be trained on the work of others for whom they don't have consent and have not paid a license fee to, then why are they? You made a point about meta training on non-publicly viewable/freely available works, such as pirated movies. That's why I made the point that using non-publicly viewable data is sufficient. At the same time I said that, using data for which you have neither consent nor license is indeed necessary, and for the training on such data the copyright/fair use considerations are complicated. > I am pointing out that search is consensual. If I train a diffusion model on images from google search,and make a paid latent-space search where I just show people some pictures without any claims of selling them rights to anything, then I claim that this is just as consensual as google images. That is because I took the source images from google images where they were consensually posted, and in the worst case my model is exactly reproducing them. More likely my model is only approximately/interpolatively reproducing them. Sure, you might change your robots.txt to update google image behavior and while model won't update, but I think generally if your images has been reposted, it is quite likely that your image is still searchable on google images, even if you change the robots.txt for your personal blog. So it's not like google is very good at updating consent either. Perhaps there is a point to be made that my diffusion-search may not just reproduce your works approximately, for which I claim that consent is kind of given, but furthermore reworks or recombines them with other works. There it is perhaps true that new consent should be necessary.