3 ms·
This is a real issue, but it's solvable with work. It's claimed that ML models' output isn't copyrightable because it's fair use, but that's hard to believe; a
by astrange 4y ago
This is a real issue, but it's solvable with work.
It's claimed that ML models' output isn't copyrightable because it's fair use, but that's hard to believe; a large model can easily memorise and output exactly one of its inputs again. This is easier to see with text, where GPT and Copilot both do it, but images can do it too.
> So how do you get access to hundreds of millions of images and use them to create derivative works? Did they get consent from millions of authors?
Build the model out of Creative Commons images only. There's a lot of 'em and it's good enough. You may need to exclude CC-BY since they currently can't follow the attribution requirement.
> Or is something like that only available to the rich with access to lawyers on tap?
More likely companies willing to license a stock photography database.