AI Lawsuits: They’re just DUMB! MidJourney, Stable Diffusion, and Dall-E haven’t infringed on anyone’s copyright, and it’s easy to prove that.
In the past few days we’ve seen a pair of lawsuits drop from two completely separate groups accusing AI generated art like MidJourney and Dall-E of copyright infringement.
The two cases were brought by radically different entities and are likely to have equally different outcomes, but they’re both inherently flawed and alleging things that frankly couldn’t have happened.
The first is a class action suit and honestly seems like a cash grab from some shady lawyers who are taking advantage of the fact that AI art isn’t terribly well understood.
Read about it here: https://www.theverge.com/2023/1/16/23557098/generative-ai-art-copyright-legal-lawsuit-stable-diffusion-midjourney-deviantart
The second suit was brought by Getty Images against Stable AI, makers of Stable Diffusion. It’s worth noting that Getty is suing Stable AI in the UK vs the US.
More info: https://www.theverge.com/2023/1/17/23558516/ai-art-copyright-stable-diffusion-getty-images-lawsuit
Neither of these suits seem likely to win on merit, though to be fair that might not be their goal. There really isn’t any evidence that these tools commit copyright violations, even though yes, you can see weird distorted Getty watermarks in images on Stable Diffusion such as the one below. I’ll explain why that isn’t copyright violation a bit later on.
So… let’s break down the issues.
First we’ll explain how AI art actually works. A lot of people seem to think that it looks at existing art and somehow changes it up a bit, and then publishes a new image based on specific work or two that were input into its dataset that fit the description of the text prompt.
That’s not at all what happens in AI art generators.
AI generates art via machine learning processes that use what’s called a neural network, which is a computer system that is designed to mimic the human brain.
Neural Nets (both biological and computerized) are made up of a large number of simple processing units, called neurons, which are connected to each other and work together to solve a problem.
Each neuron receives input from other neurons, processes that input, and then sends output to other neurons. By doing this, the neural network can learn to recognize patterns and make predictions based on the input it receives.
Imagine a group of people trying to guess the weight of an object where each person has a different approach. Some may use size, some color, some shape, etc. but each person handles it their own way.
In this metaphor each person is a “neuron” and may or may not reach the correct answer, but if you look at all their data in aggregate they can guess the weight more accurately than one person could ever do alone.
This type of technology is often used in image recognition, language understanding, and many other fields. It’s a powerful tool that can help computers learn and make decisions like humans do, but on a much larger scale and faster. It’s also done without a human having to “program” behaviors step by step the way they do in traditional software applications.
It’s important to note that no human can articulate exactly why a neural net makes the decisions it does, any more than we can account for an individual person’s taste in art. It’s just how that network worked with that particular set of data.
Apps like MidJourney and Dall-E work by training a neural network on a dataset of images. To do that, they “scrape” images from the web and load them into the neural net for analysis.
In practice, the neural net doesn’t so much scan or look at the art as it does find patterns in the art and save the data around those patterns. For example if I gave a neural net 50,000 images of chairs and then asked it to draw me a chair, it would find the elements in common from those chairs and generate something entirely new from those elements.
So… let’s ask MidJourney to give me a few pics of a chair. That’s nice and and non-controversial.
When I entered the prompt for a chair, MidJourney checked its database from however many thousands of images identified as a chair and created something entirely new that isn’t derived from any particular work of art, but instead from the patterns it found common to all of them.
Now let’s boldly go into the frontier of copyrighted works: I’ll ask it to generate a chair from the bridge of the USS Enterprise on Star Trek the Next Generation.
As a person who has consumed every piece of Star Trek media produced in the last 50 years, I can state with confidence that while these images absolutely look like they belong in a Star Trek episode, none of them actually have been.
Why? Because instead of looking at the actual chairs from the show, like these:
It looked at EVERY chair it knows about, and then EVERY picture of Star Trek furniture it knows about, and then created something entirely new from them based on its understanding of the language and the common patterns it found in the relevant source material.
Where’s the copyright violation there? There isn’t one. It’s new art inspired by previous art, just like all art.
That means that if you’ve got a bunch of watermarked Getty Images in your training data, the AI isn’t smart enough to recognize that the watermark isn’t a legit part of the picture. It isn’t copying any specific copyrighted image, it just thinks that watermark belongs there because it shows up so many times in the dataset. (Note also that it doesn’t faithfully reproduce the watermark, it approximates it because it doesn’t actually understand what it is. Again… no copyright issue.)
Now that we’ve gone through how the neural network works and interacts with the training dataset, let’s get into a subject a bit touchier and perhaps a bit more subjective: What if I ask MidJourney to create a chair from Star Trek in the style of Banksy’s street art?
OK… that’s not 100% what I had envisioned when I input the prompt, but it’s interesting, unique, and absolutely not covered under copyright. It’s still completely new and you can’t match it to any existing picture.
So now that we’ve done that, let’s get into imagery that IS clearly derivative of existing art. How about we ask MidJourney to do “Mr. Spock combined with Marilyn Monroe in the style of Banksy protege Mr. Brainwash?”
OK… finally we have something that looks like it has elements from copyrighted images. That’s clearly Leonard Nimoy and the hair, background color, and makeup from what looks like Andy Warhol’s Marilyn portrait.
We’re now in territory where it appears that we’ve got an art piece that was directly created by combining pre-existing artwork.
I’m guessing that’s a press photo of Mr. Spock from one of the movies starring the original cast. It looks like it could be this one flipped horizontally with the uniform trimmed a bit.
Then let’s look at the Andy Warhol Marilyn portrait.
So yeah, those two images were probably directly combined to create the third one above.
Uh oh… have I backed myself into a corner here with my previous statements about the frivolity of these lawsuits? Not even the tiniest bit. Let me explain.
The first thing I should bring up is that the above picture of Spock combined with Marilyn was not AI generated, it’s an actual Mr. Brainwash piece that some guy in Dubai is selling for $4,500 on Ebay.
Yes, really: Link
Here’s a sample of what I actually got out of MidJourney with repeated attempts using different prompts to try and get it to rip off Mr. Brainwash.
The top left is my favorite because it looks like Dexter Morgan with a blonde Elizabeth Taylor as Edith Keeler in the classic Trek episode “The City on the Edge of Forever,” the first episode broadcast on NBC, then below that is Sean Connery with Nurse Chapel. Then we’ve got blue haired Jerry Lewis with Joan Rivers, and finally a hung over John Mulaney with actual Marilyn.
It created new and completely original art clearly not directly inspired by existing works that fit the prompt it was given. This isn’t really looking great for the lawsuits.
In addition to generating images from scratch, another thing MidJourney lets you do is combine existing images. Let’s try that. I’ll upload the two images that Mr. Brainwash combined to create the pic for sale on Ebay and tell it to output in his style.
Again… new art, totally original. Definitely not a copyright violation any way you look at it.
Even having identified the source images and attempting to use them specifically to generate a copyright violation, I couldn’t get MidJourney to do it. Believe me, I tried.
Why? Violating copyright isn’t what this machine was designed to do, nor is it what it actually does. It creates new art based on its understanding of language after being trained on a large set of images. The result is frankly far less derivative than a lot of art created by humans.
If it were the ripoff machine these suits claim it is, we’d have gotten something that even vaguely resembled the actual Mr. Brainwash’s human made art. (Instead we got a Paul McCartneyish pic, either Rick Wright from Pink Floyd or Che Guevara below him, a woman who looks more like Emma Stone than Marilyn, and what appears to be Kiefer Sutherland wearing Gene Kelly’s hat from Singin’ In the Rain.)
So… now let’s get into the legality of all this, as in what the law actually says and the arguments these lawyers might make.
Are these lawyers going to argue that while it’s fine for a flesh and blood artist like Mr. Brainwash to combine two pieces of easily identifiable art into a new work that said artist owns the copyright on and directly profits from, it’s somehow not okay for a computer program to generate completely new artwork without even selling it? (Don’t forget what the human artist did isn’t even close to what the computer program actually does.)
When you put it like that, it really doesn’t seem like the class action suit in particular has much of a case. Then when you follow it up with questions like “What AI images specifically are you claiming violate copyright law?” or “Demonstrate how to use this tool to create content that violates copyright” the argument falls apart real fast. That’s why I don’t think it’ll make it to trial.
So what about the “charge” that they scraped the web and included all this art into the machine learning models without permission?
UPDATE: It’s been brought to my attention that the next paragraph isn’t as cut and dry as I believed.
In the time since that second ruling by the Ninth Circuit the lower court reversed its decision and the case ended up settling out of court.
The case is a bit confusing but it appears the ruling was reversed because HiQ flat out lied about what it was doing.
It seems like that leaves the legal status of actual scraping fairly secure as there is other caselaw on this (The Google News ruling comes to mind) but it’s much more of an open question than I believed when I wrote this.
That’s entirely legal in the US with lots of caselaw behind it including two rulings from the Ninth Circuit. https://techcrunch.com/2022/04/18/web-scraping-legal-court/
Plus… at least in the case of StableAI they didn’t even scrape the data themselves. They got it from a third party non-profit. If that were illegal (which it emphatically is not) the suit would need to be directed at that group, which is notably not US based.
Violating a site’s TOS is also very much not criminal, though you can still pursue civil remedies (I’ll talk about how that’s not going to help the class action suit a bit later). https://www.eff.org/deeplinks/2018/01/ninth-circuit-doubles-down-violating-websites-terms-service-not-crime
Short of a Supreme Court case or new legislation this status quo won’t change.
I’ll admit this particular incarnation of the Supreme Court doesn’t seem terribly interested in precedent and <cough> a bit weighted towards corporate interests so I guess we can’t rule that out, but in terms of a civil suit in today’s world it’s as hard to argue that the AI firms intentionally violated the law as things stand.
Then there’s Fair Use. The reason Mr. Brainwash can “Remix” Andy Warhol’s Marilyn Monroe and Spock is because of the Fair Use Provision of Copyright Law.
Among other things, that provision states that if a work is sufficiently “Transformative” that it can be considered an entirely new work with the copyright assigned to the creator even if the source content was already copyrighted. (I’m sure the choice to flip Mr. Spock and change the background to Purple was done with the that specific Fair Use provision in mind.)
So with all that, here’s what we know is true:
Scraping web content is legal*.
Website Terms of Service aren’t legally binding.
The process of generating AI imagery does not involve “ripping off” any particular artist or work of art, nor are individual artists’ works even directly stored or referenced. It’s also very difficult in practice if not impossible to get AI art generators to directly emulate existing work without transforming it significantly.
So what’s left to argue about? The suits seem pretty cut and dry from a legal standpoint. I’m not sure what exactly these lawyers intend to claim, but on the surface these suits seem silly if not downright stupid, at least under US law.
The class action suit is most likely to end pretty abruptly in the defeat of the plaintiffs if not an outright dismissal if it gets to trial. There’s just no way for them to demonstrate copyright violations have occurred, and it’s very easy to demonstrate that they did not, as I did above.
Tech issues aside, the legality of the Mr. Brainwash piece hasn’t been questioned since it was produced nearly a decade ago so it seems pretty difficult to insist now that the same thing is infringement just because a machine did it, especially when the machine didn’t even do anything like that.
All that said, the lawyers involved are fairly scary and most of the defendants are young companies without a lot of litigation experience, so they may end up settling regardless of the merits of the case. We’ll just have to see.
We don’t have the full details of the Getty suit yet, but it seems like they’re also suing for a terms of service breach regarding scraping the Getty website vs just copyright violation. I honestly don’t know the particulars of UK copyright laws nor how serious the UK is about enforcing TOS on the web, but the fact that it was brought there vs. in the US coupled with the fact that the UK does have some mechanisms for punishing companies that bring frivolous suits, there may be some teeth to this suit.
I don’t mean teeth in the sense that Getty is likely to win the suit based on a strong legal case, but since Getty is among the most litigious companies in the copyright space with very deep pockets and do seem to understand the ramifications of AI image generation to their business, they’ll likely end up with what they came for: Royalties paid by AI firms to “ethically” scrape their images.
That’s not speculation on my part, it’s actually their CEO’s stated goal vs suing for damages or trying to stop AI art as a concept and I guess that’s… good? It’s better than the labels’ response to Napster, anyway. Still, Getty isn’t the entity I’d want setting policy for copyright on the internet.
So now that we’ve established how flawed these lawsuits are, let’s talk a bit about the actual things people are upset over, the very real ethical and practical questions surrounding some of them, and what might make sense as a way forward.
People are upset that machines are creating art from scratch that’s as good or better than the art humans can create.
It touches on a spiritual point. We like to think of ourselves as “above” the rest of nature if not created with some kind of divine spark, and generally we look at creative pursuits like art as Exhibit A for proving that.
Machines created by humans are now able to do effectively the same thing, but faster, cheaper, and in many cases objectively better. From a purely emotional standpoint that makes a lot of people uncomfortable because suddenly humans don’t seem as special, and it’s easy to understand why we have that reaction.
If that weren’t enough, this new technology has the potential to utterly decimate an entire sector of the labor force. Why would a business owner pay an artist when they can just use a text prompt to get what they need?
Artists now join assembly line workers, music publishers, cab drivers, and so many other types of worker who have had their income stream disrupted or destroyed by technology in the past few decades.
Both of the above issues are very real and likely do need deep societal changes to resolve, but neither of them will be fixed by ill-conceived copyright lawsuits. Not only that, the discussion and misinformation brought on by the lawsuits themselves takes attention away from talking about the actual problems or actual solutions.
With all of the above in mind, we can get to scenarios where these tools might actually be used to violate ethics and possibly even break the law, and why even in those cases these lawsuits are silly and the companies who make the software can’t be liable if it happens.
Let’s start with the case of Greg Rutkowski. He’s a well known artist in the fantasy genre who’s done work for Dungeons and Dragons, various video game and book covers, and plenty of others.
His name turned up as one of the most popular in AI art prompts, and he’s not thrilled about it. His specific concern is that his actual work will be eclipsed by AI generated art indistinguishable from his own style that he had no hand in. It’s a very valid concern, but it’s not a scenario the current iteration of the law is equipped to deal with.
Then we’ve got Kris Kashtanova. She made a graphic novel entirely using AI art and was granted a copyright by the US Copyright Office. UPDATE: That’s actually now in question, with a proceeding underway to determine whether the copyright applies mainly due to the “monkey” type issue I reference below.
The thing is, the main character looks just like Zendaya who had nothing to do with the project. There’s also the question of who actually “created” this artwork. If the author didn’t, they can’t be eligible for copyright, like that time a monkey took selfies and the photographer who owned the camera couldn’t claim copyright on them. (Yes, that really happened and actual lawyers spent time arguing about it in court: Link)
So… again we’ve got a scenario that our current law has absolutely no method of dealing with. They just flat out didn’t think of it when the law was written.
Let’s say Zendaya decides to sue over “Likeness rights” or something. Let’s say Greg Rutkowski wants to go after someone selling T-shirts that look like his work but aren’t, and the site owner SEO’ed Greg’s name on the web store.
In those cases, any attempt to sue the companies that make the AI tools would be thrown out over a provision in the Communications Decency Act that’s been in the news quite a bit lately, Section 230.
That piece of law shields a service provider from being held liable for the actions of their users. Whether it’s an AI art doppelganger or someone making false claims about a restaurant on Yelp, the service provider in question has an almost impenetrable shield from a legal standpoint, at least in the US.
It would be like the US government suing Adobe because someone used Photoshop to make counterfeit money. The criminal act and/or damaging action was done by the user, not the software maker.
(There actually is caselaw that makes exceptions to Section 230 in the case of software that is specifically designed to circumvent copyright, but I think we’ve demonstrated the AI platforms don’t fall into that category.)
So what’s the way forward? Unfortunately, the people who will be writing these laws are largely the same ones who utterly failed to ask meaningful questions of Mark Zuckerberg when they had him in a hearing, plus now George Santos who at least founded Google and sat on the Supreme Court during the Carter Administration.
Both the Democrats and Republicans have indicated a desire to “rein in” big tech in this legislative session, and there are serious problems with both of their approaches largely due to misunderstanding the issues. Once the lobbyists get at it it’ll be even worse.
It would be great if anyone could even articulate a platform around this, but so far that’s not been forthcoming.
Ideas like “Training models should be opt-in vs opt-out,” “AI Art must be clearly labeled and segregated from Human Art,” or “Living artist names can’t be used in prompts without their permission” seem like a good place to start the discussion, but I don’t have a lot of hope that our government will go that route.
The only immediate remedy is if the AI companies themselves decide they want to act on artists’ behalf. OpenAI and StableAI seem to want to, at least for now.
The thing is, at this moment we’re at an early phase and the companies doing this are largely tech led vs profit driven. Once that changes expect much louder voices and much sillier lawsuits.
Microsoft just announced a deal with OpenAI yesterday, and I wouldn’t be shocked to see players like MidJourney gobbled up by behemoths like Adobe or even Getty Images itself. Once companies have to show completely unrealistic quarterly revenue growth or get brutalized by Wall Street, the high minded ideals tend to go out the window.
To sum up, while there’s a lot of important stuff to figure out about the ethics and law surrounding AI art, I hope at this point it’s clear that these particular lawsuits aren’t in any way viable from a legal standpoint nor will they do anyone but the lawyers any good if they’re somehow successful.
Like all other disruptive tech our society has a “Shoot first ask questions later” approach. That method does have benefits as technological progress has shot forward for the past 200 years, but we do end up with messes like this.
On top of that, science fiction is filled with cautionary tales of what happens when the tech behaves in unexpected ways. Maybe I’ll have ChatGPT write me some new ones.
