Despite claims that machines and robots are now making art on their own, we are actually seeing the number of humans involved with creating a singular artwork go up, not down, with the introduction of machine learning-based tools.
Claims that AI is creating art on its own and that machines are somehow entitled to copyright for this art are simply naive or overblown, and they cloud real concerns about authorship disputes between humans. The introduction of machine learning as an art tool is ironically increasing human involvement, not decreasing it. Specifically, the number of people who can potentially be credited as coauthors of an artwork has skyrocketed. This is because machine learning tools are typically built on a stack of software solutions, each layer having been designed by individual persons or groups of people, all of whom are potential candidates for authorial credit.
This concept of group authorship that machine learning tools introduces is relatively incompatible with the traditional art market, which prefers singular authorship because that model streamlines sales and supports the concept of the individual artistic genius. Add to that the fact that AI art - and more broadly speaking, generative art - are algorithmic in nature (highly repeatable) and frequently open source (highly shareable), and you have a powder keg of potential authorial and copyright disputes.
The most broadly publicized case of this was the Edmond Belamy work that was sold by the French artist collective Obvious through Christie’s last summer for $432k. I have already explored that case ad nauseum (including an in-depth interview with the collective). I cite it here only to point out that there were a large number of humans that were involved in creating a work that was initially publicized as having been created by a machine.
In this article we look in detail at the recent GANbreeder incident (which we outline below) that has received some attention in the mainstream press. This is another case where the complexity of machine learning has driven up, not down, the number of humans involved with the creation of art and led to a great deal of misunderstanding and hurt feelings.
For this article I spoke with several people involved in the incident:
Danielle Baskin, the artist who alleges that Alexander Reben used her and other people’s images from GANbreeder
Alexander Reben, the artist accused of using other people’s GANbreeder images
Joel Simon, the creator of GANbreeder
I was also lucky enough to speak with Jessica Fjeld, an attorney with the Harvard Cyberlaw Clinic, who has written about and researched issues involving AI-generated art relative to copyright and licensing. She is the first lawyer I have spoken with who truly understands the nuances of law, machine learning, and artistic practice.
The GANbreeder Incident
GANbreeder is the brainchild of developer Joel Simon. Simon created a custom interface to Google’s BIGgan so that non-programmers can collaborate on generating surreal images that combine pictorial elements of the user’s choosing to “breed” child images. If you are not sure what GANs (generative adversarial networks) are, you can check out this earlier article we wrote covering the topic.
Let’s look at a super simple GANbreeder example here. I clicked a few buttons in the GANbreeder interface and chose to cross an agaric mushroom with a pug. GANbreeder then outputs 6 images with varying degrees of influence from both the mushroom and the pug. Results below:
You can get more sophisticated and breed many things against each other in combinations, but the tool is dead simple (thanks to Joel Simon’s great design) and literally anyone can use it in seconds without training.
It was Simon’s vision that people would collaborate using GANbreeder and expand the tool through other creative uses. Along those lines and with Simon’s support, conceptual artist Alexander Reben wrote a scraper for GANbreeder that automatically grabbed images and stored them locally to his PC. Once local, Reben applied a custom selection algorithm that would choose images that Reben liked or disliked based on his body signals.
Reben believed the images he scraped from GANbreeder were randomly generated (as he states in this early interview with Engadget). He then sent the images selected via his body signals to a painting service in China where anonymous artists created painted versions on canvas. He called the project amalGAN.
Reben then shared the painted images widely on social media in support of his upcoming gallery shows. This triggered an avalanche of anger and frustration from other GANbreeder users. They began to complain that Reben had stolen images that they had created using the GANbreeder system.
Reben acknowledged that he did not realize the images were being created by humans. It was his understanding that the images were automatically generated at random by the algorithm.
At the time of my interview, Reben could not confirm that his scraper had not included exact images by other artists, but he believes a tiny percentage (3 out of 28) of his images were subtle variations of works that other artists had created.
The first person to call Reben out on this on Twitter was artist and serial entrepreneur Danielle Baskin. Baskin is a GANbreeder power user who often stayed up until 5:00 a.m. breeding images. She even started a service called GANvas where people could select images on GANbreeder and she would print them on canvas and ship them to customers around the world.
When I spoke with Baskin about her experiences with GANbreeder, she was careful to state that she felt she was “discovering” images on GANbreeder vs. “creating” them.
I feel like I am discovering them, not creating them. They all exist; you’re finding them. That is why I view the image as having an intelligent force behind it. It’s like I am discovering someone’s works.
Then why get so upset with Reben for “discovering” a similar image? Baskin explained the source of her frustrations with Reben’s work.
I thought that the whole project was so awful. Like, it was just so bad that it couldn’t have been real, but that it was a statement. Then I learned that it was real and I was like, “F*ck this project.”
Not that it is a competition or something. But he sort of took all the things I had in progress and had been thinking about for a long time and was immediately able to get a gallery show and sell work and stuff. And he didn’t present a clear story as to what he was doing. So that upset me. All these things were on my mind because I was so obsessed with GANbreeder.
It’s like you are writing a history book and you have been researching your subject matter for a year, and someone publishes a history book on the same subject matter, but they barely researched it and were able to sell tons of their books on Amazon. Someone took your content and got all this credit for it, but it wasn’t even good.
It was clear to me that Baskin was not a fan of Reben’s work. But I wondered if she thought he had done anything malicious or with bad intentions. I also wondered if she felt he had resolved the issue. Project aside, what did she think of him as a person?
When I met him in person, I realized Alex has built an incredible community of artists that use technology and he is a great person. It’s funny because I hate his art, but I like him - but I don’t like him as an artist.
In giving him the benefit of the doubt and in talking with him, I think he genuinely didn’t know how it [GANbreeder] worked. He thought when he refreshed the home page it was totally random images from latent space; he had no idea that other people created the images. He knew the creator of GANbreeder, so maybe he thought that Joel would have explained that to him if it were the case that it was created by other people.
I told Reben that it looked to Baskin and others like he was trying to take shortcuts, or was at least trying to remove himself from the work in some aspects. He partially objected and explained:
There was still a lot of work with me training the data sets on the art that I like and I didn’t like. The real idea was that all of the work was done before the art was made. And the actual art making process was just two simple steps back and forth. Everything involved with that is complicated, involving servers and building computers and learning algorithms and all that sort of stuff.
The interesting thing is that a lot of effort and knowledge came in to the code making. A lot of the creativity was compressed into that code, whereas now that the code is made, it is now a tool for me to… Like I said in one of the reports, I can now lay in a hammock and look at a screen and be able to just use this system to produce output.
I asked him specifically what the amalGAN project was about.
The project was, to me, about human/machine collaboration and how many steps and layers of abstraction I could add. On my website I have like seven steps of human to machine, human to machine, back and forth. I had the idea that the final step is basically the machine giving a human somewhere else the activity of using their brain to upscale the image, using their brain to interpret how to turn pixels into paint. It is basically like the machine using human knowledge to execute rather than being printed out on a printer. To me, that is conceptually interesting because it has to do with that human/machine collaboration.
I then asked Reben how he felt about the issue with Baskin and others in the GANbreeder community and what, if anything, he had done to reconcile it.
I’m sorry this happened out of mostly my ignorance of the system. I probably should have done a bit more research. When I learned there was an issue, I changed my system so it would never happen again. I’m sorry people feel this way. I think I did as much as I could at the time to get permission from Joel and to address as many concerns as I could by inviting people over to discuss. I do have the disclaimer on my website and again in my talks that some images may have come from the GANbreeder community. I have no way to verify that because there are no records of who made what.
I think most reasonable people at this point, including Baskin, acknowledge that it was done unknowingly. However, it could have become more serious - Baskin shared with me that she had considered sending Reben a cease and desist letter.
This exchange of course opens up all kinds of legal questions, and it is here that I believe things actually become interesting. For example:
Does Reben have the legal right to use an image that is either similar to or the same as the one that Baskin created in GANbreeder?
Would Reben’s work meet the legal definition of a “derivative work”?
How much would Reben need to change the image for it to be considered fair use? Is turning it into a painting enough?
What if it was the same image, but he used it as a conceptual component instead of as an aesthetic component?
Does it matter that Joel Simon’s intention for GANbreeder was for artists to build on each other’s works?
As the developer of the interface/tool, does Simon deserve some ownership over the works?
What about the folks who created BIGgan or the folks who designed the graphics cards? Do they deserve credit?
To help navigate all of this, I spoke with Jessica Fjeld, the assistant director of the Cyberlaw Clinic at Harvard Law School. I share the majority of our interview because I believe Fjeld does an excellent job of shedding light on an incredibly murky topic. It is my hope that sharing her explanations might help other AI artists from entering into sticky situations around copyright and authorship moving forward.
The Legal Implications - Jessica Fjeld
Fjeld patiently walked me through several concepts that helped me to better understand how law interacts with the new AI-generated works. Like me, Fjeld believes that all the talk about whether machines deserve copyright is overblown and distracts from real issues surrounding increased complexity of human attribution. Unlike me, she can explain the reasons why and the implications within our legal system. Fjeld explains:
Mostly the question that gets asked is, “Will AIs get smart enough that they can own their own copyright?” To me that is not that interesting because I think AGI (artificial general intelligence) is a ways out. If we get AGIs and we decide to give them legal personhood the way we give to humans and corporations, then yeah, they can have copyright, and if we decide not to do that, then no, they can’t, end of question.
In the meantime, what we really have are sophisticated, interesting tools that raise a bunch of questions because of the humans involved in collaboration making stuff with them. So we get these complicated little knots. But they are not complicated on a grand philosophical level, like, “Can this piece of software own copyright?” They are just complicated on the level of which of these people involved do [own copyright], and what parts of it.
I asked Jessica what the legal implications were in the GANbreeder incident. Disclaimer: Alexander is a past client of Jessica’s, but she is not currently representing him in relation to the GANbreeder incident.
It is a fascinating question. I have tooled around a little bit with GANbreeder myself, so I can understand it. One thing that is important to note is that copyright protects original expressions that are fixed. So “original,” “fixed,” and “expression” are the key terms here.
Something has to be new, and obviously, much of what is on GANbreeder is. Part of what makes it an exciting website is you get some of these really unfamiliar feelings - sometimes eerie, sometimes funny.
Then the next word we learned about is “expression.” Copyright does not protect ideas; it only protects particular expressions of those ideas. So if someone said I had the idea the put into GANbreeder “dog, mountains, and shell,” and I got an image that was similar to the one that someone else is now using, that is not protectable. The exact image, maybe; but a very similar one, no. And something that is very interesting about GANbreeder, as I was tinkering with it, if you have it create a child on the scale from similar to different, if you say to make it very similar, a lot of the children images that come out are very, very similar. There may be individual pixels or a slight shift in the orientation, but at a casual glimpse, you wouldn’t even necessarily see [the difference].
It’s interesting especially because of the timing of when Alex took these images off, when all the works on GANbreeder were unsigned because there were no accounts. It’s a little hard to say. If you were thinking about pursuing an infringement case, you would really have to prove the exact image had been copied rather than a similar idea where, say, one is orange and one is red.
I asked Fjeld how different a work had to be to be considered original.
In GANbreeder, if you keep making tiny changes, eventually you are going to get something that does have what we would call “originality” in copyright. But it is really hard to say when that happens. And in a lawsuit, it will just be a fact-specific inquiry: Is this the same or is it not? And we have this concept of derivative works for works that are very similar. It can be an infringement to make something that is extraordinarily similar, but not just a mere reproduction.
I asked Fjeld if it mattered that Joel Simon’s intention for GANbreeder users was to build upon each other’s existing works. Wasn’t Reben simply using the tool as intended? It turns out there is a thing called an “implied license.” Fjeld explains:
The other piece around how GANbreeder encourages folks to draw on other people’s work I think brings up another interesting question, which is that it’s largely settled law, particularly in the Ninth Circuit in the U.S., that you can grant a non-exclusive license to use your work in an implied way, so it doesn’t have to be explicit.
U.S. copyright law does require that if you are going to dispose of your right to the work - so either going to give an exclusive license to someone else, or if you are going to sell your copyright - you have to have a writing. But for implied license, a non-implicit license, you don’t have to have a writing. And at least some courts upheld that it can just be implied - you don’t even have to have a conversation about it.
And when I look at GANbreeder, because of the way it’s set up, because of the way the whole system is architected, it gives you an image created by someone else and encourages you to iterate on it. It certainly looks to me like there is an implied license to do that within the context of the site. Anyone who is creating work there understands that other people are going to use it as a basis to make their own work.
Now, when courts look for implied licenses, it is again a fact-specific inquiry. I think with regard to what Alex did, the question is, did people understand that part of the implied license they were given, not just that you can monkey around with it in the context of the GANbreeder app or you can also integrate it into this other system and have it painted by anonymous painters in China and show it in a gallery. They might not have anticipated that, and that’s probably where the issue comes in.
There was an implied license to do something, but the scope of that implied license wasn’t totally clear. Then that is complicated because it is a site that is architected with a thousand models and images in it, so you are essentially navigating the points in a multi-dimensional space created by that number of models and can have any combination of those thousand images. But it creates a lot of very similar images.
So the combination of the fact that the scope of the implied license wasn’t very clear and the fact that people may have an attachment to their ideas or individual expressions and then may see a very similar one… it is my understanding that Alex’s project shouldn’t have directly just reproduced anyone else’s; it would have started with someone else’s, and then he tweaked it based on his body signals.
I wondered why Reben’s work would not be considered derivative and asked Fjeld if she thought it could legally be considered so.
I would say that yes, there is an argument that Alex’s works could be considered derivative of existing works on the GANbreeder website. There remains the question of the implied license because the derivative work is a copyright infringement, but if the use is licensed, then there is no infringement.
There is also a question of what the damages would actually be, because in copyright, you can get statutory damages if you register your work in a narrow window around its creation or before the infringement happens. If you don’t do that - and to my knowledge, none of the GANbreeder images have been registered - then what you get is actual damages. And it’s not totally clear what the damages would be for folks that anonymously created images on a website and then later found that someone had them painted and displayed them in a gallery.
*I also don’t know if there have been any sales. There is the image that Alex used and whether there is a derivative work in that process, and then he takes this further step and has them painted into oil paintings, which, again, I think is another tweak. So there is a series of manipulations of the underlying content.
*Note: There have not been any sales.
I asked Jessica if she thought these “manipulations” by Reben pointed towards “fair use” (a term I had heard in the past but did not fully understand).
Yes, they do steer me more to think about fair use. As I have heard Alex presenting on this work, he really emphasized that for him, it really isn’t about the outputs; they are not the artwork at all. For him, the artwork is the process by which he had trained this series of systems to produce the artwork, test them against his own preferences, to title them, etc. For him, the interesting thing is the process by which he tried to design a bunch of algorithms to take himself as far as possible out of the creation process. The expression of them is that he ends up putting his name on painted images in a gallery. But even putting his name on them is a little complicated in regards of what he was thinking about in regards to the artwork.
When we think about fair use, one of the main factors that courts consider is how transformative the use is. And I do think there is a strong argument here that because the underlying theme of the work, we could think about it as fair use because we want to incentivize this kind of exploration of the space. The way that Alex talks about it, there is an argument that, ethically speaking, it should be clearer that despite the paintings being up at a show with his name on them, he doesn’t really think of himself as the author of them in a certain way. The use is transformative because it is making this point of how far can we push toward algorithmic authorship.
You could think about the Richard Prince vs. Patrick Cariou case. They are both fine art photographers, but Prince is a conceptual artist, an “appropriation artist,” he calls himself, and Cariou is a more traditional fine art photographer.
Cariou had gone to Jamaica and taken this book of photographs and done a gallery show all of Rastas. He spent all of this time investing in these relationships to produce these images. Prince used them in a gallery show in which he manipulated them a little bit. One of the classic images that gets shown a lot is a full image of a guy hanging out in a jungle setting, and Prince very roughly cut out an electric guitar and pasted it in on top of him. A lot of the original image was still there with this crude-looking addition on top. And Prince won that case as fair use because the argument he made is that he had transformed the content. Yes, Cariou was also a photographer that had a gallery show, but Prince was using it in this conceptual, imaginary space. I think you could think of Alex’s work in a similar way.
Fjeld shared my belief that the driving force behind some of the confusion in AI-generated art was that more people, not fewer, are typically involved. We talked a bit about the developer behind GANbreeder, Joel Simon, and what rights he had, if any, to the works.
In GANbreeder you can click a button and it’s possible to get the coolest thing that GANbreeder ever produced. And how much do we want to think it is in line with the goals of copyright if someone is just clicking a button and the software is producing it… how much do we want the person to click the button to be the person to get the rights? Do we think that Joel, who set up the system, gets some rights? Do we think the people who put in the work to create these models that took thousands and thousands of hours of computing power should get some rights?
There used to be a doctrine in copyright called “sweat of the brow” where courts had an instinct that they wanted to protect people’s investment of time, and that has been rejected. So the notion that people who spent time to create the model should earn rights in the outcomes isn’t the state of copyright in the U.S. right now. But there is something in there that ethically feels to us like if you just click a button once, you are involved in that creation, but maybe you shouldn’t be the person who gets all the rights.
I found Fjeld’s explanations both fascinating and much needed in this space. It was a welcome reprieve to hear a lawyer talk about these issues that we keep seeing coming up in the AI art space without over-focusing on the red herring of whether the machine deserves copyright.
Regardless of what the law says, we all answer to the court of public opinion, and it hasn’t been particularly kind to Alex Reben over the GANbreeder incident. I think the animosity towards Reben stems from folks not liking that he appears on the surface to be doing less work than other artists, yet getting more attention. A common complaint waged against conceptual artists. But more importantly, I think people can see with their own eyes that at least one of his works looks the exact same as an image created by Danielle Baskin, and a few others are similar to images made by other members of the GANbreeder community.
I like Alex and consider him a friend. I also like Danielle and plan on following her work moving forward. So I thought back to what I learned from Jessica Fjeld about it being important that Alex’s work not be the exact same as Danielle’s. This seemed like a pretty easy thing to figure out, so I compared the two images using James Cryer‘s excellent tool called Resemble.js, which can compare two images and highlight the differences.
Other than a little bit of aliasing (I took a lower-resolution screenshot of Baskin’s image), they look the exact same to me. I shared my new findings with Alex and asked if he would consider removing the image from his website in light of the new evidence. He did one better and called Baskin to discuss the best way to move forward. Reben then crafted the following statement, which he first ran by Baskin for approval.
I spoke to Danielle by phone to work out what she thought would be fair for me to do to move past this issue, given all the information we have at this time. We landed on giving her a credit under the artwork on my website as "Original GANnbreeder image sourced from Danielle Baskin" and to make the credit for GANbreeder more obvious on the page. If any other images happen to arise with a similar issue, I'll have to deal with them on a case-by-case basis. But since the images from the website at that time have no authorship information and may be randomly generated, there may be no other issues apart from the few which were already identified. I'm also only concerned with images which are basically the same, not images which are similar and "bred" from a like set of "seed words," as this use aligns with the spirit of the website. Of primary concern to both of us was to put this issue to rest so that GANbreeder can continue to be used as a creative tool and grow from what was learned.
Score one for the court of public opinion.
As always, if you have questions or ideas you can reach me at email@example.com.