I think there's an argument to be made that neural nets are in some sense a form of compression. The model is a lossy compressed representation of the data. So

I think there's an argument to be made that neural nets are in some sense a form of compression. The model is a lossy compressed representation of the data. So training a model on copyrighted data is quite direct copyright infringement - you're compressing, then redistributing. Has this ever been used as an argument in a legal case?