Generative AI Has a Visual Plagiarism Problem | Hacker News Reader

Generative AI Has a Visual Plagiarism Problem

spectrum.ieee.org

If you have trained a person to draw, or paint, and they're really good at it, and if they have general knowledge of pop culture, you could ask them to draw the Simpsons, or Mario and Luigi, and they could. It is not literally copying so much as re-synthesizing. The artist's mind contains an archetype of Mario, and knows how to draw figurines in general. And if you commissioned an artist to draw that, and published it, it would be copyright infringement. But the artist does not infringe merely by having that potential.

I'm of the interpretation that the models do not infringe; the person running inference infringes when they intend to create copyrighted imagery using the model.

Yeah but one of the points the article makes is that people don't necessarily know if they're infringing or not with Midjourney, especially when text is generated. Only rarely if at all does this happen with humans (the subconscious is a mysterious place) but it is easy with LLMs and harder to detect because the person doing the generation is not as involved in the creation of the text. Of course LLMs could be patched to detect and alert but... how? and what incentive to do so?

Of course the models don't infringe but perhaps the people building them did by basing their IP on other people's IP. for the courts to decide...

> people don't necessarily know if they're infringing or not with Midjourney

it is the copyright owner's responsibility to sue for infringing works.

When I was quite young and visiting my grandmother's home, I made a painting or drawing of Garfield. I was definitely not the artistic type and I had no training or guidance on drawing or coloring, but on that day I felt inspired, and pretty much reproduced that ugly fat cat without reference to an official image.

This must be the common case for practically any kid creating art these days; they are going to imitate commercial and copyrighted content that they are most exposed to. It just seems a natural thing to do, the unnatural course being that they come up with something beyond the Threshold of Originality.

Yeah the infringement (in this old school model) would be if you then used that image as -- say -- the cover of a book or something. Then, it would be analyzed to see if it's "fair use" (satire, etc) or infringement.

I guess the question with MJ is -- if I make such a Garfield with MJ, show it to my parents etc., put it on the fridge, is that a problem? prob not. i'm not a lawyer but -- that doesn't seem like I'm infringing. but... did MJ infringe by selling me a subscription that gave me that image for pay? (pay that otherwise Garfield's IP owners might've received)

Correct me if I'm wrong, but my understanding is that the creation of the image is still infringement.

No one is going to sue a child over a drawing on a fridge, though. The PR would be a nightmare, there's no earnings to seize, and it's not cannibalizing the market for the real thing in any way.

Midjourney has issues on a couple of fronts

One is the direct infringement, ie who is the creator? Is it the AI for creating it or the human who entered the prompt? Whoever created it is infringing, but again, probably not worth it to sue Joe Schmoe over it.

The second is that Midjourney hosts all of the images they generate, some of which are infringing. I'm guessing they're trying to use Safe Harbor protections, but I'm a little dubious that they can actually call the output of their own AI "user content". Users are promoting the AI, but Midjourney is still the one generating the content.

Imagine you take all sorts of copyrighted pictures, compress them lossily, encrypt them separately using a unique passphrase that can be guessed, and place them into a big tarball to be distributed.

Are you infringing copyright?

I think this is a useful mental model here, as it is difficult to argue that guessing a passphrase is copyright infringement.

That is a bit different as it suggests strict comparmentalization of the reproductions. There is a lot of "concept bleeding" that happens in the models I've used.

Perhaps if the compression directly leveraged overlap among the images, via keywords.

I imagine OpenAI gets to court on a copyright issue and hides behind the DMCA: If you did not like an output, you could have sent a DMCA takedown notice explaining what content was plagiarized, and we would have censored every page that included that content and also put in place a block to prevent that output from being shown again.

What little understanding I have of US copyright law is that the DMCA is merely a convenience to content hosters - you can optionally get a chance to right the wrong to potentially avoid a lawsuit.

Notice all the weasel words in there? Yeah... They don't have to file a DMCA to sue you, nobody does, they can just sue you. (Mandatory IANAL)

This quote from the Midjourney CEO is hilarious: "There’s no way to find a picture on the Internet, and then automatically trace it to an owner and then have any way of doing anything to authenticate it".

It should be "There's no way to stop us from using every picture on the Internet, and not be tracked while doing so".

I expect this to be a something-burger but not even close to a Napster-level extinction event for AI. With gen AI, the use is not a surrogate for the copyrighted work (a replicated image of Toy Story is not like you're getting the whole film for free).

However, there is money involved so... I predict a small bump in the road of sorts

I think "plagiarism" and "copyright" are two related, but different problem.

Plagiarism is about taking credit when it shouldn't.

They complain both about potential reproduction and complain about efforts to block it. (A red team without consent is just a hacker!) Talk about speaking out of both sides of your mouth.

Reply on news.ycombinator.com