But what about an imaginary specialized AI that really just focuses on reproducing its input training material in a nearly verbatim way?
The current reality is somewhere in between these two extremes, and defining the boundaries sounds difficult.
The output should certainly not be copyrightable if it does not produce an existing copyrighted work or if fair use is in play, which is why many people seem to be attacking the training set itself, an argument which is also complicated but has little merit.
e.g.: What happens to your MIT licensed code when an AI assistant generates some code for you which contains some GPL, BUSL or Source Available licensed parts? Moreover, do these code bases scraped with consent of the owners?
The image itself is copyrighted, or it isn't. Fair use may or may not come into play for things like replication of known copyrighted characters, depending on commercial use.
None of this has changed from the traditional copyright process. Even 15 years ago I could find some of my old works rehosted on obscure Chinese websites, and my options for recourse were the same as they are today. Many of my own works were already derivative.
So if I make a program that takes an image and changes a single pixel, I can use it to launder copyrighted images?
Here's what you need to know about derivative works in US law: https://www.copyright.gov/circs/circ14.pdf
2. You can't train people exclusively on books. If you asked someone to explain a concept they learned in those books, sure that knowledge will help, but their response will also affected by every moment in their life so far. They're bringing other unique things to their response.
[1] - https://www.nytimes.com/2023/07/10/arts/sarah-silverman-laws...