Embed Notice

This strategy avoids the trouble of drawing a connection from works in the training data and *outputs* of the models by focusing more granularly on the inherent copying entailed in training. If the plaintiffs can get over the procedural hurdles of establishing specific infringement (and class cert, etc.), this strategy also has the advantage of shifting the burden of dealing with the outputs into the case for fair use, which is way more atomistic and complex post-Warhol.