@draeath @nihkeys @DJGummikuh @GossiTheDog If it was an accident, or incompetence, then it would be rapidly corrected.

If it's not rapidly corrected, then it is the purpose.

@Em0nM4stodon I cannot express how much I've been dying for glasses that are also screens since I read Snow Crash.

And I cannot further express how angry I am that the companies finally making that a reality are fucking it up by making them inherently surveillance devices.

@christianschwaegerl @JeffGrigg @jamie @fsinn @pluralistic Buddy, I'm not trying to "normalize" anything, especially not LLMs. I'm telling you how the law works. I never said the law was good.

@christianschwaegerl @JeffGrigg @jamie @fsinn @pluralistic You're making a bunch of different arguments now. The topic at hand was, "Is it copyright infringement to make and have an AI model trained on millions of books?" The answer is no. This is wholly legal.

Storing copyrighted work is legal.

Modifying copyrighted work is legal.

Storing modified copyrighted work is legal.

It doesn't matter if they have a model that is literally just plain text of every book, or if the model is a series of mathematical weights that go into an algorithm. It's already legal to have and modify copyrighted works.

What becomes illegal is reproducing and distributing copyrighted material.

No, whether it was for "commercial" or "non-commercial" purposes doesn't matter when determining if something is infringing.

No, whether it was "sold" or "distributed for free" doesn't matter when determining if something is infringing.

"What about Fair Use?" Fair use is an affirmative defense. That means that you acknowledge you are infringing, but it's an allowed type of infringement. It's still an infringement, you just don't get punished for it.

But, as already stated, nothing is infringement until there's a distribution. Without a distribution, no further analysis is needed. When a distribution occurs, it is the distribution that is analyzed to determine if it is infringing, and, if so, if there is a fair use defense. Everything that happens prior to the distribution is irrelevant when determining if an infringement has occurred, as long as the accused infringer acknowledges they have the copyrighted work (which AI companies always acknowledge).

There is one further step, because it is illegal to make a tool that is for copyright infringement. The barrier to prove this is so high, though. As long as a tool has any non-infringing uses--and we must acknowledge AI can generate non-infringing responses--then it won't be nailed with being a "tool for copyright infringement". This has to be, like, "Hey, I made a cracker for DRM, it can only be used to crack DRM. It literally can't do anything legal."

Even video game emulators haven't been hit with being "tools for copyright infringement" because there are legitimate uses for them (personal backup, archival, etc.), even though everyone knows they're 99% for infringement.

@JeffGrigg @christianschwaegerl @jamie @fsinn I think this is missing the point and the law (at least, US copyright law).

I buy a book. I then own that book. I can cut that book into individual pages. I can scan all those pages into my computer. I can have an image-to-text algorithm convert the text in the images into an ebook. I can do this to a billion books. I can run whatever algorithms I want on the text of those books. I can store the resulting text of my algorithms on my computer, in any format.

This is all legal, for both me and for any company. Copyright does not prevent use of a work after it has been sold, "use" meaning just about anything--short of distributing the work.

Because what copyright protects against is the reproduction and distribution of copyrighted works. For AI companies, that "distribution" doesn't happen until somebody puts a prompt into the AI, and receives back a result. That result is the distribution. To sue an AI company for copyright infringement, you would have to have a result that infringes on your copyright, and you would have to prove that the AI company was more than just a tool that the prompter used to infringe your copyright.

For the Disney example, if somebody prompted, "Darth Vader in a lightsaber duel with Mickey Mouse," it would be an uphill battle to prove the AI company is responsible for that instead of just the prompter. The argument that the AI company would make is that the prompter clearly used the AI as a tool to make infringing work, but just like you can't sue Adobe if someone used Photoshop to make the same image, you can't sue the AI company because someone used it as a tool to infringe copyright.

Now, I don't find that a wholly persuasive argument because of the, frankly, complicity in the creation that AI has that Photoshop doesn't, but that's definitely the argument they would make, and judges have seemed receptive to that and similar (and even worse) arguments.

As far as I'm concerned, the original point of this thread proves that the AI company should be mostly-to-wholly responsible, even if the prompter was deliberately asking for infringing works. After all, AI-generated work is not copyrightable because it is not human created, it is computer created.

If it's not human created, how can the human be responsible for the infringement?

If it is computer created, then isn't the computer's owner responsible for the infringement?

After all, if I ask a digital artist to create me "Darth Vader in a lightsaber duel with Mickey Mouse," and they do, the digital artist is on the hook for that infringement. They reproduced the work, and they distributed it. There is a "prompter" and a "creator" in both scenarios; it seems illogical that if the "creator" is a human, they're responsible, but if the "creator" is a computer, they aren't responsible.

This is, per @pluralistic, "It's not a crime, I did it with an app!" Why we let apps get away with crimes we'd never tolerate from people, I don't know. But that's where we are.

@katrinatransfem @fsinn @jamie If the material is acquired legally, they don't need a specific "license" to use it as training material. Copyright holders don't get to determine how their work is used after it's acquired, except to prevent its distribution.

Now, for the even larger than normal scumbags like Anthropic and Meta that torrented millions of books, that's certainly a problem. But Google, for instance, actually bought all the books they scanned.

@fsinn @jamie My understanding was that training an AI model on copyrighted work was fair use, because the actual "distribution"--when the AI generates something from a prompt--uses a diminimus amount of copyrighted content from an individual work, except if the user explicitly prompted something like, "Give me Homer Simpson surfing a space orca," at which point the AI company would throw the user all the way under the bus.