Comment by juunpp

Reproduced to what fidelity? 100%?

If OpenAI's output reproduces a copyrighted image with one pixel changed, is that valid in your view? Where does the line end?

Copyrighted material should never be used for nonacademic language models. "Garbage in, garbage out." All results are tainted.

"But being forced to use non-copyrighted works will only slow things down!"

Maybe that's a good thing, too. Copyright is something every industry has to accept and deal with -- LLMs don't get a "cool tech, do whatever" get-out-of-jail free card.

>Now, if the original copyrighted work can be extracted or reproduced from the model, that’s obviously copyright infringement.

I think there's an important distinction to be made here - "can" be reproduced isn't infringement, only actual reproduction is (and degrees thereof not consisting of sufficiently transformative or fair use).

Trivially a typewriter can reproduce a copyrighted book. Less trivially Google books, with iirc stores the full text of copywrited works has been judged to be legal.

If the AI were a human and that human made an image that copied substantial elements of another human's creative work after a careful review of the original creator's work, even if it was not an original copy and no archival copy was stored somewhere in the second creator's creative space, I would be concerned about copyright infringement exposure if I were the second (copying) creator.

I'm open to the idea that copyright law might need to change, but it doesn't seem controversial to note that scraping actual creative works to extract elements for an algorithm to generate new works crosses a number of worrying lines.

Have you seen the examples of midjourney reproducing exact frames of Dune, Star Wars etc? With vague prompting not asking for the media property specifically. It's pretty close to querying a database, except if you're asking for something that's not there it's able to render an interpolated result on the fly. Ask it for something that is there however and the model will dutifully pull it up.

> You can’t open up the database and find a given duplicate of a work. Instead you have some abstraction of what it takes to get to a given work.

So distributing a zip file of a copyrighted work subverts the copyright?

Well, fine, but you'll have to claim that copyright is unjust and that you are breaking the law as an act of civil disobedience. The AI corps do not want to take this stance as they also have intellectual property to protect. Classic "have their cake and eat it too" scenario.

Imagine if every book or advertisement or public conversation you overheard led to future claims that you had unethically learned from public information. It’s such a weird worldview.

(BTW I forbid you from using my comment here in your future reasoning)

> If you post your images for the world to see, and someone uses that image, you are not harmed.

Let me define a few cases of 'uses that image' and see where your line in the sand drops

* If someone used that image as part of an advertising campaign for their product, they are profiting off your work. Are you not harmed?

* If someone used that image and pretended they created it. Are you not harmed?

* If someone used that image and sold it directly. Are you not harmed?

>I'm not convinced that most copyright infringements are immoral regardless of their legal status.

You're right and wrong. You're right because most infringement is from people who can do minimal damage and in fact so more help by giving awareness to your works by sharing. But this is only becsuse copyright it working (most of the time) against corporate entities who don't want to leave any room for legalities to come in.

If copyright ended I'd bet my bottom dollar Disney and all the other billionaires companies would be spamming any and everything that gets moderately popular. And Disney can put advertise the original artist easily.

My statement above is that their activities are illegal and wrong, not that one implies the other. They are illegal because of the copyright violation, and wrong because regardless of what the law says, using the images for training despite the artists' every appeal to the contrary (being vocal about the issue, putting a robots.txt file to avoid scraping, and now using adversarial techniques to protect their work from being stolen) is just moronic. It's like shitting on their front yard when they've asked you a million times not to shit in the front yard, put a sign that says "Please don't shit on my front yard", and sprayed insecticide all over the grass to try to deter you from shitting on the front yard. And yet you still shit on their front yard and even have the balls to argue that there's nothing wrong or illegal about it. This is absolutely insane.

That GP is utterly confused about copyright law is not an opinion.