DISCLAIMER: The information in this interview is provided for informational purposes only and should not be construed as legal advice on any subject matter.
Brady Walker: Let’s talk about AI art. Stable Diffusion and Midjourney are. Is DALL-E also under the gun?
Kayvan Ghaffari: There are two pending lawsuits relating to artificial intelligence art engines. DALL-E is not part of the pending lawsuit against Stable Diffusion, Midjourney, and DeviantArt, which alleges the three AI engines used billions of copyrighted images used to train AI art generators to reproduce unique styles without compensating artists or asking for consent. However, OpenAI, the creators of DALL-E, are involved in a separate lawsuit concerning the creation of GitHub’s Copilot program, an OpenAI-powered coding assistant. So, while DALL-E is not directly implicated in either case, the two cases do put artificial intelligence under the copyright microscope.
BW: What are the different forms of copyright infringement?
KG: In order to understand copyright infringement, it’s important first to understand what copyright is. Copyright is a form of intellectual property for “works of authorship.” U.S. copyright law applies to an expression that is created and fixed in a tangible form. It does not protect facts, ideas, procedures, or an artist’s style, no matter how distinct. For example, copyright does not protect Piet Mondrian’s style of grid-based abstract art. Instead, copyright is limited to the actual expression contained in a medium. With Piet Mondrian, copyright would protect the precise grid patterns of a particular art piece. The copyright laws grant a set of exclusive rights – also referred to as a “bundle of rights” – to the creator and protect the creator regarding issues of reproduction, distribution, adaptation, and derivative works.
Here’s a basic idea: I am sitting at a coffee shop, drinking my cold brew, and drawing some art on a napkin. That drawing is eligible for copyright protection, even if my artistic capabilities leave much to be desired…
Now, once I have that drawing, I can grant or restrict other people from using that art. That’s where we get into licensing and infringement. We’ve talked about copyright licensing before, which can be found here.
The use of copyrighted materials can come in various flavors. It’s frankly what makes this space so interesting and intellectually stimulating – it’s why I started my legal career as a copyright attorney. Some uses of copyrighted materials will be legally allowed, while others will not be. The latter category is known as “copyright infringement” – i.e., I am using the copyrighted material in a way that is not allowed or authorized by the creator or by law. For example, I copy a picture or a sound recording and use it without authorization. And this act of “copying” takes on many legal contours, making copyright litigation its own form of art.
The legal issues in the two pending cases involve whether the use of allegedly copyrighted material to train the artificial intelligence constitutes copyright infringement. Training an artificial intelligence engine involves a process in which the machine learning program learns its behavior from a set of material and/or data referred to as training data.
This issue reminds me of a case I litigated – a first of its kind – involving copyright infringement and AI. I represented legal research startup ROSS Intelligence in a bet-the-company case against Thomson Reuters/Westlaw. Thomson Reuters alleged ROSS infringed its purported copyrighted materials, including U.S. judicial case law when ROSS trained a legal research AI search engine. Similar to ROSS, OpenAI, Stable Diffusion, Midjourney, and DeviantArt all created AI engines that are trained with certain content – a necessary step in training any technology.
In the case of Stable Diffusion, Midjourney, and DeviantArt (“Defendants”), which is the focus today since it involves the art space, the Plaintiffs’ legal issues are whether the Defendants used allegedly copyrighted images for training the AI algorithms and whether that use constitutes copyright infringement. The Court will have to determine if the algorithm is using the content in a way that infringes on these rights.
Assuming the Plaintiffs can actually identify any allegedly copyrighted materials in the training data, of critical importance for the Court will be to determine how the Defendants used the materials. Did Defendants verbatim copy the expression in the allegedly copyrighted material? Did the Defendants store any of the allegedly copyrighted materials? Did the Defendants abstract themes from the images rather than a verbatim copy? If there is copying, is the Defendants’ conduct considered “fair use” under copyright law? Overall, it is a complex legal issue that will require careful examination and analysis.
BW: What about source images? With Midjourney, you can directly use source images in generation, which allows for easy copying of styles. Just copy an image link and add it to a prompt, then adjust the weight of the image to determine how closely the output resembles the source image. There are limitations to the direct control one has over the results, but it’s remarkably easy to steal a style this way.
KG: That’s a different issue, similar to Napster or Grokster. In those cases, technology was used to infringe in a different way. To differentiate, let’s separate these issues into two parts. One is effectively a “pre-launch phase,” where a company is directly training its algorithm with certain data. This applies to ChatGPT and other machine learning models like Google. This “direct” training by the company would be considered “direct copyright infringement” if there actually was proven copyright infringement.
After the technology goes live, any future user input may continuously train the system. Let’s call this “post-launch phase,” and it is where user uploads are at issue. Voluntary user uploads have the potential to create a separate form of legal liability for the company because the creators of the artificial intelligence engines are not actively doing anything to infringe. In this case, the theory of secondary liability may exist – i.e., where a company is alleged to have induced or contributed another to infringe.
Take Napster. Napster created a peer-to-peer file-sharing technology that was used predominantly to infringe copyrighted materials. Napster did not directly infringe because it was not directly ingesting and training any alleged copyright material. Instead, the Court held Napster would likely be held contributorily liable because it had specific knowledge that its technology was used pretty much solely for infringing purposes.
BW: Do you think that’s something that Stable Diffusion or Midjourney will be on the hook for later down the road?
KG: I personally don’t think so. There are a few reasons The Supreme Court set a precedent in a case involving Sony Betamax regarding secondary copyright infringement. There, Sony was sued for allegedly allowing users to record TV shows and sports programs.
However, the Supreme Court held that Sony was not liable for copyright infringement because the technology had significant non-infringing uses – i.e., the Betamax was used for families to record family memories, not for illegal uses. This established a sort of “primary use” test that looks at the nature of the use of technology and whether there is a significant non-infringing use. I suspect there is significant non-infringing use to these technologies that would help Stable Diffusion and Midjourney from a similar fate as Napster.
It’s important to remember that this technology is in its infancy, and with most things in copyright law, the law doesn’t ordinarily match the pace of innovation in technology. So, it’s too early to say anything conclusively, and the courts may disagree with me.
BW: Do you foresee any other legal ripple effects from these lawsuits or just from the existence of these technologies?
KG: From a legal perspective, it will depend on the outcome of the two cases. If the court rules that Stable Diffusion, Midjourney, and DeviantArt did infringe, it could have a ripple effect. Those companies may have to pay artists, which would likely result in increased fees to use the platforms. This could be costly for users and break their business model.
BW: But in that case, it’s sort of like the cat’s out of the bag, like publishing a file so anybody can 3D-print their own gun. It’s just this circulating thing. I have Stable Diffusion on my own personal computer, but that’s impossible for anybody to remove.
KG: Yep. I don’t disagree.
BW: What other ways might any of these engines be legally vulnerable?
KG: There are a few issues with respect to using artificial intelligence search engines to create art. One issue is affecting the engines and companies that created them, which we’ve talked about. Another big overarching issue is the question of rights and ownership of the output (the actual art). If something is created using Midjourney, is that creation eligible copyright protection? If it is, who owns the copyrights, Midjourney or the individual that entered the words and generated an image?
The copyright law only protects an expression created by a human, not an idea or a style. Patents are used to protect innovations. If a sketch is drawn on a napkin, it’s protected by copyright as long as it’s original and human-made. However, it’s uncertain if a machine-made output will be considered original and human-made enough for copyright protection.
There was a case from now over a decade ago that put this issue in the spotlight. This case involves a “Monkey Selfie” – that’s right, a six-year-old Celebes crested macaque took a selfie using a camera owned by British photographer David Slater.
Slater traveled to Indonesia to take pictures of the endangered Celebes crested macaque. He sort of befriended a group of macaques and set up a camera in the middle of the jungle. I don’t remember all the facts, but essentially this macaque sees the camera, goes over to it, and clicks a button that happened to take a picture of him smiling. The picture is actually awesome. It’s the first-ever animal selfie. Wikimedia Foundation displayed the image on its website, claiming Slater didn’t have a copyright because he didn’t take the picture.
There was a bunch of back-and-forths between Slater and several third parties, including the Wikimedia Foundation. PETA also got involved, at one point suing Slater and alleging the macaque was the rightful copyright owner of the image. Ultimately, the dispute ended with a few learnings. First, the U.S. Copyright Office states that works created by a non-human, such as a photograph taken by a monkey, are not copyrightable.
Second, a federal judge in San Francisco stated during a hearing that copyright law does not extend its protections to animals (though this was for a separate issue relating to whether an animal can sue someone for copyright infringement). Ultimately, the monkey selfie will go down in history as an awesome picture and for really kickstarting the legal precedent that copyright protection only extends to human-made content.
For artificial intelligence, I suspect a similar set of issues will arise – i.e., who actually “creates” the output. Is it technology or the individual inputting the keywords? It’s not clear-cut at this time. My personal feeling is the scope of copyright protection will be fact-driven as to how much the ultimate end product was dictated by human intervention. In other words, the more the creator edits and/or manipulates the output, the more likely the end product will receive copyright protection.
BW: It sounds like the photographer set up the precise conditions by which a monkey might take a selfie, and it happened. And Midjourney didn’t set up exactly the conditions that would allow somebody to generate a specific piece of art. So, to me, on the surface, AI engines seem more like the invention of the camera than a photographer setting up the scene for a monkey selfie. Except that, up to a point, you have no idea what the output is going to be.
KG: I think that’s right. With a camera, the photographer sets and views the scene through a lens, then captures that scene with a click of a button. With AI, you don’t have that clear vision of what output will be captured. Of course, you can take a guess as to the output based on the prompt, but you’d be really hard-pressed to have someone tell me for certain that every day, that each output from Midjourney is exactly what they anticipated each and every time. And that’s where I think the two technologies differ.
BW: How would you respond to someone who spends 40-50 hours a week generating AI imagery? They use 1/1000 of what they generate, and they have developed a set of prompts and techniques for getting what they want, but they don’t edit the output. One might say that this person has developed a skill using AI. You’re saying they still might not be eligible for copyright?
KG: Good question. Very interesting, plausible scenario. I personally don’t think it would be copyrighted. That’s my personal opinion, that’s not MakersPlace’s opinion. And the reason why I say that is because what he’s doing is a creating process, it’s an idea. That may be patentable, but the process is not an expression that’s entitled to copyright protection. But reasonable minds can differ. I can see a court saying that’s creative enough and that the outcome is created by the key terms and nothing else, but I still think it’s not yet copyright protected.
BW: Okay. Cool. Is there anything that I haven’t asked that you think would be important to include in an article about considerations with AI art?
KG: Yeah. I think it’s important to be mindful of what art you share with these companies. Let’s say I take a Picasso piece I uploaded onto Midjourney. I may not have a license to do that with a Picasso. Right? So you could yourself be infringing without a license. If you use that output without a license from the Picasso estate, that output would be an infringing derivative right (assuming, of course, the output from the AI engine is copyrightable).
I am a big supporter of people using whatever creative tool they have to help create beautiful works of art. AI is just one tool in the creative process. But it will be important for creators to look at the terms of service of Midjourney, Stable Diffusion, DALL-E, and DeviantArt to understand what rights you’re giving up and what rights you’re getting with respect to inputs and outputs.
Take the time to understand what rights you may or may not have with whatever you’re trying to create. And be comfortable with not having any intellectual property rights if you, in fact, do not. It doesn’t mean you can’t sell it. Duchamp famously made a urinal into an art piece. Does he have a copyright? No. But is that urinal still worth millions of dollars? Yes.