My best guess is the courts will (eventually) rule that content creators must be paid for their work.

There's hardly any basis for it. The most likely outcome would be that downloading data and training a model with that data does Not infringe upon copyright at all. You are not reproducing and distributing Nor are you publicly performing a work by training a model.

On the other hand: when using the trained model to generate images based on it - the resulting image might be infringing. This Infri

When AI must pay for content they have 2 choices, train LLMs with Public Domain media for free or pay for a limited set of up-to-date content

Isn't that a false dichotomy that misses (rather badly) that copyright is automatic? IF they use say appropriately licensed creative commons works,. those are still copyrighted worjs if created where copyright is automatic, for instance.

Maybe ... we REALLY need to stop using "copyrighted" as a synonym for "needs to pay" or "off limits" - when the issue is more lice

"train LLMs with Public Domain media for free"

So we could get the level of literacy of Jane Austen and the Brontë sisters? James Fenimore Cooper? Daniel Defoe? Jack London? Edgar Allen Poe? H.G. Wells?

I can think of ever so many worse outcomes.

Of course when Chandler's and Christie's works hit public domain the LLM may get just a little paranoid.

> "Generative AI" as it is implemented right now does not create anything. It merely copies.

Point one, IMO, seems hardly substantiated if one wants to get really pedantic in that a specific combination of elements not existing before exists now - seems in that basic context it'd be created whether man or machine does it.

Second one IMO DEFINITELY needs a citation. My understanding of LLMs and AI is rudimentary at best, but it seems like this is not how they are supposed to work, not even "as they are currently implemented. IMO if that is correct (that this notion is incorrect)... can we not?

> It's one thing to be inspired by something, it's another to copy and paste their work into yours. "Generative AI" as it is implemented right now does not create anything. It merely copies.

No, not really. It is more like a series of probabilities. If you train on a book that includes the line "The quick brown fox jumped over the lazy dog," it might know that there's a 1% chance that "the" is followed by quick or lazy, a 1% chance that fox is preceded by quick and brown, a 1% chance that lazy is followed by dog, etc.

The original data is not present in the training set, except in very limited circumstances where there are not very many items in the training set that match against the query we

"Merely copies" reflects a misunderstanding of how artificial neural net models work. They copy and mix information from vast input datasets, representing abstracted commonalities discovered in the whole of the input. Then they output a different mix of information, depending on large parts of the intermixed and abstracted information model, and depending on the particular prompt requesting some output. Note that it's called a prompt (a prompt to take the action of touring

associatively relevant part

Which would mean you could sue for infringement. They aren't suing for infringement. They're challenging the learning.

Technically to view a movie on the internet I need a computer to copy it as well. Its up to the courts to decide if its fair use or not.

Of course taking the information from the movie into our human biological neural net is copying the information (and its copyrightable patterns).

That's not different in kind to what the computer is doing copying it to be represented as many additions/substractions to many weights between the nodes of an artificial neural network.

That's the crux of the issue.

We're going to need some new understanding, in the legal system, of what is really going on with AI learning and output generation/synthesis.

I can't be sued for writing a book in the style of, say, Michael Crichton. I can be sued for creating output that is substantially similar enough to be considered a violation (I take Jurassic Park, rename it Jurassic Meadow and rename the characters, but keep the story). Regardless, that's suing someone based on the output they create, which isn't what's happening here. They're challenging the learning.

In addition, no retroactive extensions would be a good fix as well. Authors who copyrighted their stuff pre-CTEA knew what they were signing up for. There was no reason to extend their copyrights.