It is another thing altogether to read The Fellowship of the Ring and then write a book Companionship of the Amulet that has roughly the same plot. The more similar my work is to the original work the more likely it is to be ruled derivative and then what I can do with my work becomes strictly curtailed.
There's an interesting experiment for you:
Put copies of an AI on two different computers, and make them identical in every way except that one of them has had their training data searched for the text of the Lord of the Rings books, and it's been deleted. Then give them identical prompts, and to the extent they use random number generators, fake it by giving them the same random numbers (e.g. https://xkcd.com/221/">4), and see what they come up with.
Because similarity, even perfectly identical works, isn't enough for infringement. There has to have been actual copying. And fantasy tropes like elves, dwarves, and magic rings (which are a dime a dozen -- from the ring of the genie in Aladdin to the ring of the Nibelung to the ring of Gyges) are freely usable under the scÃne à faire doctrine.
However, copying the full text of a work (or an image) into the memory of an AI model probably is not covered.
I disagree. It sounds remarkably like the Google Books and Google Image Search cases, where the full text of a book or full image were dumped into a computer and saved permanently for the purpose of creating a tool that permits people to search the text of books and search for images that match a description or a query image. Those were held to be perfectly legal.
the courts have been siding against decrypting a work as fair use for a long time
First, decryption is neither here nor there. Second, I don't think courts have held that decryption isn't fair use so much as it isn't copyright infringement to begin with so fair use is inapplicable. That part of the DMCA is not about copyright, it's about anti-circumvention.
The fact that AI works can't be copyrighted makes it easy to conclude that AI generated content is nothing but the uncopyrightable derivative content of every input that went into the model.
No. Derivative works are copyrightable as a rule, and I'm not convinced that the output of these AI programs are in fact derivative works under copyright law. The reason for the AI generated works not being copyrightable is because copyrights are only vested in authors, and these things can't be authors. It's something that might be revisited if we ever create a truly self-aware AI on par with a human, like HAL 9000. We aren't at that point yet.
Controlling how copyrighted material is used is 100% what copyrights are about.
Eh, not really. The purpose of copyright isn't to control the use of works, and in fact, 'use' isn't one of the rights that comprises copyright. Copying is, distribution is, public performance and display are, preparation of derivatives is, but not mere 'use.' As long as you're not doing one of those other things, you can use a work without permission all you like.
This really is no different than me taking a book that I like and making a recording of me reading it. I am entitled to do this. I can even copyright my performance, but I can't monetize (or even share) that performance without the express permission of the original copyright holder. That's even despite the fact that there is a genuine creative act by an actual human as the written word is turned into an audio performance.
No, that would be preparing a derivative work, and potentially performing a work publicly. You aren't entitled to do it, and you probably can't get a copyright for your performance. Regardless of the fact that it would indeed be potentially creative. (It depends on how creative your reading was -- like if you're doing the voices and such. A Ben Steinian monotone would quite likely not qualify for copyright.
nor that protection from creating works that are derivative by default
I'd really recommend looking at the legal definition for what a derivative work is.
That bit isn't fair use, as it involves the entirety of the work
You can have fair uses that involve the entirety of a work, e.g. time shifting TV broadcasts onto a VCR tape for later watching, or space shifting a sound recording from a CD to an MP3 player. You can also have infringing uses that involve only selected quotations of a work, like in Harper & Row v. Nation Enterprises.
George should absolutely be able to control how his copyrighted material gets copied into an AI model. This is essentially the same right that keeps Hollywood from making a movie of his works without his permission.
I'd disagree on both points. I think it's better if he can't, and I don't think it amounts to an infringement of the derivative right. The reproduction right is the more significant one to be concerned about anyway, but I think they're good there too.
Alternatively, I suspect that George would be fine with the idea that everything generated with a model that included his copyrighted material would be deemed a derivative of his work.
That would definitely not fly, and again your ignorance of what a derivative work is in copyright is really harming your argument.
However, it is extremely unlikely that the generative AI people will be given carte blanche to include any works that they want into their models and then be able to use the output of those models however they want.
I think they have a pretty good chance of winning.