One of the key principles quoted is "questioning whether artists can substantiate copyright infringement in the absence of identical material created by the AI tools". Or to paraphrase "how can it be copyright infringement if the output is different?".

I'm not here to make or dispute that argument, but it is interesting to ponder the notion that, should a future AI become able to similarly digest and remix existing content in order to create viable new content - say, absorb Game of Thrones and Star Trek and etc-etc, and spin out a whole new engrossing series where everything is at-once very familiar and slightly-changed - and in the process "launder" (to borrow a concept from finance) the copyright, then that could indeed vastly change the future outlook for some existing areas of art and entertainment industries.

I'm not sure what you mean by "laundering" the copyright. Even human authors regularly take inspiration from other works. Often very substantial inspiration, and copyright does not prohibit this. Take Warcraft and Warhammer for example. Those two settings are much more similar to each other than a Star Trek & GoT spinoff would be to either Star Trek or GoT.

As long as the work produced by an LLM drawing inspiration from Star Trek and GoT was sufficiently distinct from either setting to infringe on copyright, I don't see why it'd be different from a human author producing such a work.

I agree, and also reference the (western) music world, where sequences of the same (12) note frequencies are borrowed and interbred with one-another in an endless self-referencing repertoire of "unique content".

Perhaps one difference - of arguable validity - between human and AI content creation is the way a digital system can in theory be step-by-step debugged to forensically inspect at the lowest level exactly which parts of what prior content have been referenced, or altered, and how, at each step of the program iteration. (much more difficult to do, with we analog humans! - we must mostly rely upon inspection of the outputs, there)

The corpus of text used to train the AI is the contextual knowledge about the world. When you give it a word like "phasor" it's going to associate that word with the Star Trek universe (and probably angular representations of complex numbers from EE textbooks too). When you give it words like "Godswood" or "Andals" it's going to make associations with those words.

Regardless of whether we think generative AI is empty regurgitation, the copyright claims are going to have difficulty sticking unless models are regularly reproducing copyright texts. Even when they are it's usually in short snippets, e.g. "It was the best of times, it was the worst of times..." not regurgitating a whole Dickens novel.

> Aren’t LLM’s just predicting the next token with zero contextual knowledge of the world they inhabit

Only partly true, especially with Q* learning coming onto the scene as it is much more dynamic, able to self-learn.

Don't human authors often glean things from other copywritten works?

> This means that plaintiffs across most cases will have to present evidence of infringing works produced by AI tools that are identical to their copyrighted material. This potentially presents a major issue because they have conceded in some instances that none of the outputs are likely to be a close match to material used in the training data. Under copyright law, a test of substantial similarity is used to assess the degree of similarity to determine whether infringement has occurred.

This has always seemed obvious. If it's making Mickey Mouse you're doing something wrong.

What is the minimum copywritable unit?

It would vary across media, but a copyeme (like morpheme, grapheme, phoneme) is very hard to define.

Will it end up being some weird tensor, or something human recognizable?

I think the recent cases on music show it's too variable to encode this way. Sometimes something as common as a common chord progression is enough for courts to decide it's a copy. Sometimes not. Consistency would mean the Amen break is worthy of copyright, and that would cause a lot of people a lot of trouble.