In 2020, Carlos Marcial sold an NFT depicting a melting Lamborghini for 0.2 ETH, equivalent to $28 at the time. He has since become a successful digital artist and an outspoken advocate of NFTs’ power to help artists support themselves, but in 2022 he caught the attention of Lamborghini, who argued that his image was a copyright infringement.
Marcial is just one of many digital artists currently facing copyright threats as powerful actors look to enter the NFT market. Artists minting new work are therefore engaged in Russian roulette — risking an infringement allegation both at the moment of creation and for many years in the future.
One popular form of NFT art, produced with generative adversarial networks (GANs), is especially touch-and-go for copyright liability. GAN-generated work has figured prominently in the NFT space from the very beginning, indeed the first ever token ID on the NFT marketplace SuperRare was for a GAN-generated work by Robbie Barrat.
What are GANs?
GAN art forges a collaborative partnership between humans and machines. But how does that partnership work? To create a machine learning algorithm that can produce visual images, computer models are trained to find patterns in large sets of images. Human artists carefully curate the data sets on which these algorithms are trained, but for every gem produced by the early algorithms they also generated large volumes of underwhelming outputs. GANs addressed this issue by joining together two machine learning models: A “generator” that creates new examples based on a data set, and a “discriminator” that judges the generator’s outputs as real or fake. When the discriminator starts routinely mistaking the generator’s outputs for images from the training set, the GAN is working, and artists can then further tweak its parameters until it produces exactly the desired output.
The artworks resulting from this process therefore depend on both the training set and the input of the artist. Some artists create their own algorithms, and others rely on open-source algorithms available to all. For example, the Eponym website allows users to enter any word or phrase into a text box, which then produces an AI-generated artwork that can be minted as an NFT.
Lamborghini and other rights-holders are presently seeking to enforce their rights in a way that gives them control over NFT markets, thereby suppressing this emerging field of creative expression. For our purposes, the crucial point is that it is the data sets used to train GANs that give rise to copyright issues. However, while it’s certainly possible to create GAN works that infringe copyright, the basic process of training a GAN should be considered fair use.
What is fair use?
All artists have a right to copyright protection for their original works, which means that, in a copyright dispute, an artist who is an alleged infringer also has rights. Indeed, not all uses of copyrighted works are automatic infringements: Sometimes a use is actually “fair” under copyright law.¹ Fair use can be hard to understand, because most of the rules are based on court decisions.
In 1994, the US Supreme Court decided a watershed case regarding lowbrow song lyrics. The court held that 2 Live Crew’s song Pretty Woman (1989) didn’t violate Roy Orbison’s copyright in the original song because it was a parody — a kind of transformation.² The court sought to balance, on one side, how extensively the song had been transformed, and on the other side, the fact that 2 Live Crew was making a commercial use of the protected work. 2 Live Crew was clearly profiting from the song, and courts often balk at finding commercial uses fair. However, on this occasion, the court reasoned that because the work’s transformation from sappy oldie to 90s rap track was so extensive, the two works actually served different markets.
The court decided that although Pretty Woman had a commercial purpose, overall it was fair use. In other words, although Roy Orbison had a right to copyright protection of his work, 2 Live Crew had a right to their fair use of it. Nearly 30 years on, courts still adopt the framework laid out in this case, especially when evaluating the use of older cultural material.
Why NFTs are fair use
Whether something is fair use or copyright infringement can have a number of expensive implications for artists. In the US, copyright suits must be argued in federal court and their costs can quickly escalate to tens or hundreds of thousands of dollars. This could stifle an emerging field like GAN art until a court somewhere rules on the fairness of training algorithms on a data sets of images.
While it is clear to these authors that, under the framework set out in the Pretty Woman case, artwork created by GANs trained on copyrighted works is fair use, not everyone agrees. Indeed, a number of NFT artists and platforms have recently received takedown requests alleging that their works are infringements. The fair use analysis of GAN-generated art is two-fold. First, there’s the question of whether the images used to train the GAN have been copied, and second, there is the question of whether the GAN’s ultimate output — sold by the artist — too closely resembles a pre-existing work. In both cases, the fair use analysis will usually turn on whether the use is commercial and whether it is transformative.
Training a GAN is transformative, but often the purpose is non-commercial. Many artists freely trade such algorithms, even posting them online for public use. This practice is now so common that even when a group of artists used a publicly available GAN to produce a work that subsequently sold at Christie’s for $432,500, its programmer Robbie Barrat (also a young artist), only objected mildly on Twitter, choosing not to pursue a lawsuit.
In GAN art, the end goal is to create a new expression that builds on the training set. Any reproduction of that data set is incidental. In one case, the courts held that the images used to train an algorithm are raw material designed to further a “distinct creative… objective.”³ While in another, the creation of a database for purposes other than simply reproducing an underlying work was regarded as a transformative use.⁴
GAN-generated artworks are fair use because they are transformative. So just as 2 Live Crew reimagined a well-known song in the burgeoning domain of hip-hop, contemporary AI artists use existing works in the production of crypto art. Lest we forget, even where a use is commercial, it can still be fair. Indeed, the stronger the argument that a use is transformative, the less commerciality matters.
The transformative use argument for GAN-generated NFTs is particularly strong because the goal of using data is to create something new. Moreover, the resulting works are unique, original creations and typically have a “different aesthetic” than the copyrighted works on which the algorithm was originally trained.⁵
While it is clear that NFTs should be considered fair use, a court will likely need to address the issue in the near future. This means that someone will have to make the fair use argument in front of a court. Once that court weighs in, all artists using GANs will be impacted by its decision. The stakes are high and this type of litigation can be very expensive, particularly when the opposing side is a large, well-funded institution. Under the current NFT market structure, this burden will most likely fall on one or more unfortunate independent artists.
Practically speaking, in order to sell an NFT, artists must list them for sale on a marketplace (with few exceptions). Marketplaces like OpenSea, SuperRare, and Nifty Gateway require artists to agree to their Terms of Service, which state that the artist is responsible for ensuring their NFTs are not infringing and to indemnify the marketplace from any alleged copyright infringement should it arise. This means that if an NFT listed on a marketplace is alleged to be a copyright infringement, the cost of arguing that it is fair use will always fall solely on the artist, not the marketplace. While marketplaces take a percentage of sale proceeds from these artists’ NFTs, they are currently forcing artists to bear the legal burden all on their own.
Samantha Altschuler is a member of the Harvard Law School class of 2022 and co-president of the Harvard Law School Blockchain and FinTech Initiative. She is also an advanced student at Harvard’s Cyberlaw Clinic. In addition to her role as an advisor to the Blockchain for Social Impact Coalition (BSIC), she has served as a research assistant for the Berkman Klein Center for Internet & Society. Her work has been published in the Stanford Journal of Blockchain Law & Policy. She will join Wachtell, Lipton, Rosen & Katz as an associate in the Fall of 2022.
Jessica Fjeld is a Lecturer on Law at Harvard Law School and the Assistant Director of the Cyberlaw Clinic at the Berkman Klein Center for Internet & Society. She is a member of the board of the Global Network Initiative, which works to protect free expression and privacy of internet users. Fjeld is also a poet, the author of Redwork (2018), and the recipient of awards from the Poetry Society of America and the 92nd Street Y/Boston Review. She holds a JD from Columbia Law School, where she was a Hamilton Fellow, James Kent Scholar, and Managing Editor of the Journal of Law and the Arts. She also has an MFA in Poetry from the University of Massachusetts and a BA from Columbia University.
Maude Wilson is a member of the Harvard Law School class of 2022. She is co-president of the Harvard Law School Blockchain and FinTech Initiative and an advanced student at the Cyberlaw Clinic at the Berkman Klein Center for Internet & Society. She serves as an advisor to the Blockchain for Social Impact Coalition and as research assistant to Christian Catalini at MIT. She will join Davis Polk & Wardell as an associate in the Fall of 2022.
¹ Copyright Act of 1976, 17 U.S.C. §§ 101-1332, 2012.
² Campbell v. Acuff-Rose Music, Inc., 510 U.S. 569, 114 S. Ct. 1164, 127 L. Ed. 2d 500, 1994.
³ Blanch v. Koons, 467 F.3d 244, 253, 2d. Cir., 2006. In the case of Blanch, the Second Circuit court held that the artist had engaged in a transformative use when he scanned a copyrighted fashion photograph, adapted its color, background, medium, and size, then incorporated it as part of a collage painting which differed in purpose and meaning from the original photograph.
⁴ Authors Guild, Inc. v. HathiTrust, 755 F.3d 87, 97, 2d Cir. 2014.
⁵ Cariou v. Prince, 714 F.3d 694, 706, 2d Cir. 2013.