Instead of building image generators off of images scraped from people's art without their consent, we could use openly licensed images, and intentionally push the public towards licensing more of their images for open datasets by encouraging twitter and instagram/meta to add image license options to image uploads, and running some public service campaigns on these platforms to encourage use of open licenses to help build better datasets.
At the same time the smaller image sets that would be available would encourage additional research in to sample efficiency, which I regularly hear would be a generally useful area for further research.
This approach would ensure several things: Public datasets would be available to all, not just the major institutions (assuming enough people cared about open licensed models to push the major players to use them), sample efficiency would be improved, more people would get used to licensing their images for public use, and artists who did not want their style copied by these technologies would have their rights respected.
That last point would have a HUGE effect on public opinion about AI, building trust between the public and the AI research community.
Ultimately I am in favor of a world where intellectual property restrictions are sharply curtailed, but to do that ethically we would need to put in place other systems to provide for those whose livelihood depends upon their intellectual works. And if someone created a work before AI existed, and they reserved their legal rights for their work at the time of publication, I think that should be respected such that the work would be left out of machine learning datasets.
This approach would take more work, but setting this expectation would encourage the tech industry to clarify image licenses on their platforms, ultimately promoting open culture and open image licenses.