Simple.
Copyright owners have to be asked for agreement
BEFORE the use of ai training.

How is the training issue (not seeking permission or giving compensation to the copyright holder) different from using samples in music? There've been lots of lawsuits over this and I don't think the notion that the sample is short gets you off the hook.

The fact that you can't create these images without a large database to "train" with is not at issue, as far as I know. The fact that there are lots of people's copyrighted work that you're only stealing a very little bit from doesn't really change the basics of the transaction. Even images lifted from social media have copyright - the person who snapped the image holds it.....

thanks for those links!
Sampling music takes one work & adds/modifies it; similar to extracting sky, background from an image & pasting it into your image. but there some relevant issues:


==========
C. Fair Use Defense
Exceptions exist to the exclusive rights granted to copyright owners. The
fair use doctrine allows someone other than the copyright owner to use the
copyrighted work in a reasonable manner without permission.   

...
If the trier of fact determines that the level
of unauthorized use does not rise above the threshold of substantial similarity,
then the trier should find the unauthorized use to be de minimis. [too small to matter]

...
 Problems arise when the
appropriation is quantitatively small and not the heart of the work.

the difference again is that these sampling cases address specific, detectable copying. in ML the individual images are
so, if the data used is only one of thousands tested, it isnt the heart of the work? (very different than transforming one image)

for ML training, the images are analyzed and classified. the result is not copying or storing the image per se, but storing a digital analysis of the image retrievable by its tags. such a process is one way, not commutative --  you can't reconstruct the image from the new collection [for some a wonky details:  https://nanonets.com/blog/machine-learning-image-processing/#working-of-machine-learning-image-processing ]  analogous to netflix scraping copyrighted works and producing recommendations?

 In Sandoval v. New Line Cinema,

 artist Sandoval alleged that New Line
Cinema�s theatrical release of Seven infringed his copyrighted work when his
photographs appeared in a movie scene without his permission.  The court
reiterated the de minimis analysis in Ringgold, but held as a matter of law, that
the use of the photographs was de minimis because the photographs were not
quantitatively observable.

similarly, no individual image is observable in images created de Novo? To create a new image, the generative phrase is compared with those non-image tables.  the training results are a novel expression.

quotes from https://cpb-us-e1.wpmucdn.com/sites.suffolk.edu/dist/5/1153/files/2018/01/SRWILSONV1N1N-rp2cve.pdf (hopefully fair use!)

 the copyright questions arise in the collection & use of the imagers in the training phase, not in the later creative phase.  should infringement in the first area contaminate the later use of that resource?   if you use a thesaurus which was prepared using copyrighted collections/compilations, are you liable if you use that book to write the next great novel?  similarly, the results of training are usable in creating new images. IANAL