I've been thinking of a possible resolution to this. For Copilot and similar systems: keep the training data, and in addition to the text generation, add a search function. Find generated sequences that Copilot puts out, and send pointers to the source for close matches in strings over a threshold length. Example: if Copilot produces the Quake fast inverse square root routine, you'd get a pointer to the source AND to the license. This would allow credit for the author for permissive free licenses, and would allow the user to dump that code if it's GPL and they aren't willing to distribute under those terms.

For art, train on contributed images whose authors agree to use them for that purpose. There could be some organization, perhaps a nonprofit, that would own the images and users of the model could credit that organization, and perhaps contribute back their own generated work. That way a legally clean commons could be built and could grow.