What about copyleft analogues? Plenty of important work in our field is licensed with e.g. apache license, which only came into existence because people needed new kinds of licenses for new kinds of purposes.
Copyright enables copyleft. Does copyright in relation to AI also enable copyleft in relation to AI?
The logic seems to be that "some of the training input was copyright, therefore the inference output is copyright." The same logic says "some of the training input was copyleft, therefore the inference output is copyleft."
Maybe future training sets will have to worry about the license compatibility between all the data in the training set?
The idea copyleft doesn’t need copyright has been dead since TiVo, you need copyright to ensure people’s freedom, if you simply repealed copyright laws you would still have your rights as a user trampled.
The problem in this case is the AI companies aren’t ignoring copyright and then releasing their output with no copyright, which would happen in a copyleft or copyrightless world.
They instead want the benefits of closing their output but they want the benefits of taking other peoples work to make that output.
