What would happen if you made a lossy image compression format derived from tons of scraped, copyrighted images.
There’s no generative ability, but anytime you compress/decompress your image the model uses weights and biases learned from copyrighted works.
Is that a violation?
Nothing.
and no:
Only distributing it to others is when copyright is an issue. Private translations/transformations are unenforceable. You can mark up a book you've bought as much as you want.
copyright makes no sense;
It can not be treated just as a Xerox machine, but it can be treated as a Xerox machine that has within it all the copywritten works (that a user can inventively request combinations there within) which it has trained on (and saved in the form of weights/bias). In this case the AI model itself is the distribution of works under copyright. Encrypting/transforming copywritten works and transmitting it is a violation of copyright (afaik; ianal).
This is all to say, copyright - as it stands - needs heavy reform. I'm rather copyleft. Because all of this is vestigial nonsense from an age where printers from the 1800's setting the rules, and our thinking hasn't updated yet.