Ultimately I think AI models and their outputs should not be copyrightable unless they were only trained on data for which the trainer had appropriate rights (or was public domain) including the right to resell model results.
These things are basically like JPEGs for knowledge and text. If I make a JPEG of a work I do not strip copyright from it. Of course since the trained model is a cumulative set of all inputs the rights are the set disjoint of the rights the trainer had on the data.
If we allow these things to be copyrightable it allows what amounts to for-profit corporate piracy and unlicensed resale of all works.
>If I make a JPEG of a work I do not strip copyright from it.
Because anyone could point point out that your jpeg contains a copy of the work. It's a lot harder to point out a copy of any work in an AI model. They're trained on so many images that you can't even represent one pixel worth of data per image.
These things are basically like JPEGs for knowledge and text. If I make a JPEG of a work I do not strip copyright from it. Of course since the trained model is a cumulative set of all inputs the rights are the set disjoint of the rights the trainer had on the data.
If we allow these things to be copyrightable it allows what amounts to for-profit corporate piracy and unlicensed resale of all works.