>the problem that people cannot think for themselves any longer (as seen by the downvoting in this submission).

Quite an interesting take to assume that everyone who disagrees with you cannot think for themselves.

There was also the lawsuit against google for the Google Scholar project, which is not only very similar to how AI use ingest copyright material, but even more than AI actually reproduced word for word (intentionally so) snippets of those works. Google Scholar is also fair use.

The fair use test (in US copyright law) is a 4 part test under which impact on the market for the original work is one of 4 parts. Notably, just because a use has massively detrimental harms to a work's market does not in and of itself constitute a copyright violation. And it couldn't be any other way. Imagine if you could be sued for copyright infringement for using a work to criticize that work or the author of that work if the author could prove that your criticism hurt their sales. Imagine if you could be sued for copyright infringement because you wrote a better song or book on the same themes as a previous creator after seeing their work and deciding you could do it better.

Perhaps famously, emulators very clearly and objectively impact the market for a game consoles and computers and yet they are also considered fair use under US copyright law.

No one part of the 4 part test is more important than the others. And so far in the US, training and using an LLM has been ruled by the courts to be fair use so long as the materials used in the training were obtained legally.

It seems like you're responding to a question about training by talking about inference. If you train an LLM because you want to use it to do sentiment analysis to flag social media posts for human review, or Facebook trains one and publishes it and others use it for something like that, how is that doing anything to the market for the original work? For that matter, if you trained an LLM and then ran out of money without ever using it for anything, how would that? It should be pretty obvious that the training isn't the part that's doing anything there.

And then for inference, wouldn't it depend on what you're actually using it for? If you're doing sentiment analysis, that's very different than if you're creating an unlicensed Harry Potter sequel that you expect to run in theaters and sell tickets. But conversely, just because it can produce a character from Harry Potter doesn't mean that couldn't be fair use either. What if it's being used for criticism or parody or any of the other typical instances of fair use?

The trouble is there's no automated way to make a fair use determination, and it really depends on what the user is doing with it, but the media companies are looking for some hook to go after the AI companies who are providing a general purpose tool instead of the subset of their "can't get blood from a stone" customers who are using that tool for some infringing purpose.

This is largely because search engines are a concentrated market and AI training is getting done by everybody with a GPU.

If Google, Bing, Baidu and Yandex each come by and index your website, they each want to visit every page, but there aren't that many such companies. Also, they've been running their indexes for years so most of the pages are already in them and then a refresh is usually 304 Not Modified instead of them downloading the content again.

But now there are suddenly a thousand AI companies and every one of them wants a full copy of your site going back to the beginning of time while starting off with zero of them already cached.

Ironically copyright is actually making this worse, because otherwise someone could put "index of the whole web as of some date in 2023" out there as a torrent and then publish diffs against it each month and they could all go download it from each other instead of each trying to get it directly from you. Which would also make it easier to start a new search engine.

It's not a stretch because I'm not claiming they're the same thing, I'm incrementally walking the tech stack to try and find where we would want to draw the line. If things something has to be generative in order to be a violation, that (for all but the most insane definitions of generative) clears `wc`, but what about publishing the DVD or BluRay encryption keys? Most of the "hacker" communities pretty clearly believe that isn't a violation of copyright. But is it a violation of copyright to distribute that key and also software that can use that key to make a copy of a DVD? If not, why? Is it because the user has to combine the key, with the software and specifically direct that software to make a copy of which the copy is a violation of copyright but not the software and key combination?

If that's the combination of the decryption key and the software that can use that key to make a copy of a DVD is not a violation of copyright, does that imply that distributing a model and a piece of software separately that can use that model is also not a copyright violation? If it is a violation, what makes it different from the key + copy software combo?

If we decide that generative is a necessary component, is the line just whenever the generative model becomes useful? That seems arbitrary and unnecessarily restrictive. Google Scholar is an instructive example here, a search database that scanned many thousands of copyright materials, digitized them and then made that material searchable to anyone and even (intentionally) displayed verbatim copies (or even images) of parts of the work in question. This is unquestionably useful for people, and also very clearly producing portions of copyrighted works. Should the court cases be revisited and Google Scholar shut down for being useful?

If market effect is the key thing, how do we square that with the fact that a number of unquestionably market impacting things are also considered fair use. Emulators are the classic example here, and certainly modern retro gaming OSes like Recalbox or Retropie have measurable impacts on the market for things like nostalgia bait mini SNES and Atari consoles. And yet, the emulators and their OS's remain fair use. Or again, lets go back to the combination of the DVD encryption keys and something like handbrake. Everyone knows exactly what sort of copyright infringement most people do with those things. And there are whole businesses dedicated to making a profit off of people doing just that (just try and tell anyone with a straight face that Plex servers are only being used to connect to legitimate streaming services and stream people's digitized home movies).

My point is that AI models touch on all of these sorts of areas that we have previously carved out as fair use, and AI models are useful tools that don't (despite claims to the contrary) clearly fall afoul of copyright law. So any argument that they do needs to think about where we draw the lines and what are the factors that make up that decision. So far the courts have found training an AI model with legally obtained materials and distributing that model to be fair use, and they've explained how they got to that conclusion. So an argument to the contrary needs to draw and different line and explain why the line belongs there.

I am pretty sure companies keep one copy of each item.