Questions of “ownership” in art can be a matter of law, of social norms, or of art theory. New art forms and new methods of producing art can fall foul of existing answers to these questions or creatively re-open them. Often they do both. “AI Art” produced using contemporary “Artificial Intelligence” artificial neural network software is a good example of this. “Rare Art” produced using blockchain token software is another, which we will consider below in relation to one particularly notorious example of AI Art.
“Portrait of Edmond Belamy” was produced by the artist group “Obvious” using an existing artificial neural network model trained on a corpus of images of classical paintings. Obvious did not credit the author of that network model, or any of the artists whose paintings were included in the corpus. At this point there are already three different layers of questions around ownership.
Firstly, the assembly of an image corpus. Accurate reproductions of paintings that are no longer in copyright should not attract copyright, and in the US at least this is quite rightly the case. A collection of such reproductions may attract copyright on the collection itself, but this should not affect individual works within the collection. If the images were of paintings that are still under copyright, copying each image might infringe that copyright. I say “might” because doing so might fall under fair use/fair dealing (hereafter just “fair use”) exceptions to copyright. These exceptions are popular both with artists who work with appropriation and with large Internet companies who work with search and advertising. Both groups, and others such as Digital Humanities scholars, might wish to assemble such corpora of images so it is difficult to generalise about motives and outcomes regarding them. In the case of art, however, fair use for artists is a key defence of artistic creativity in an age where the visual environment is dominated by corporate media.
Beyond this legal view is the ethical and art theoretic view. Is it right to treat individual artworks in a corpus as tokens of a type or as just part of a set, as fodder or as raw material for an industrial process? With apologies to Clement Greenberg, does discarding the tactile elements of painting still meaningfully capture it, and does discarding visual detail and differences in scale discard more for processes of derivation than processes of study?
Secondly, the training and use of the artificial neural network model on that image corpus. The model is trained by processing images in the corpus, by copying and reading their data. The model will contain representations of parts of the images from its training corpus, and its output will also resemble parts of the images they are trained on. Mechanical copying and creating derivatives of images are covered by copyright. Cutting up images and juxtaposing them with the work of other artists is covered by the moral rights that accompany copyright. Again, copyright does not apply to works that are out of copyright (moral rights vary by country…), and artists should have a claim to fair use of such materials. The degree to which artistic use of source materials transforms them should be a factor in establishing that such use is indeed fair use, and the output of artificial neural networks certainly transforms the images that they are trained on. Style is not copyrightable (let’s not talk about “trade dress” here), but forgery and “passing off” can be legal matters, and the application an artist’s signature style to a work that they have not made but is sold under there name is the same whether performed by human hand or algorithm.
Again, the ethical and art theoretic view raises more questions. Signature styles are a matter of pride as well as profit for artists, and while this can be critiqued within art theory it is a strongly established norm that simple imitation of style, without a critical framework for doing so, is a breach of artistic norms. Artificial neural network models need not operationalise an individual artist’s signature style in order to devalue the concept of signature styles in general.
Thirdly, Obvious’s use of existing neural network software to generate an image has caused widespread debate. “Signing” the image with the algorithm used by the artificial neural network software to produce it functions as a double-bluff whatever the intention behind doing so. We know that Obvious produced and sold the image, their attribution is not threatened by this. But it erases the work of both Robbi Barret in producing the model of art that the image is simply a product of and of the artists that the neural network’s model already erases the authorship of, both in terms of attribution and in terms of control of their work (even if from beyond the grave in the case of the corpus artists). Software authors should not be able to control uses of the tools that they produce – Microsoft should not be able to censor your writing using Word or claim joint authorship of everything you write using it. But a trained artificial neural network model is a more complex thing than a text editor from this point of view – it is as much content as it is tool. Microsoft should not be able to tell you how you change the contents of an empty text document, but changing a novel or a painting whether represented physically or digitally may infringe on the copyright and moral rights that it may have. Again fair use should be strongly considered for artistically transformative use of artificial neural network models.
Art theroretically, such direct and uncredited use of existing materials, even materials created by another artist, may count as appropriation art, which is an established category within the arts. Appropriation art is deliberately transgressive, often for critical effect. Appropriating non-art or low-art materials is very different from appropriating canonical art or the art of leading contemporaries but both can be critical moves. Artistic labour can be appropriated directly, in the case of contemporary artists who use studio assistants to produce art under their own signature such as Jeff Koons or Damien Hirst. The signature that the products of this labour are exhibited and sold under is a key part of its erasure. And where software used to make art is free software(/open source), attribution may or may not be a strong social norm but past a certain point that attribution is useful information to have for artistic, critical, and art historical engagement with the work.
Prior to this, Obvious had already encountered the question of ownership and found an answer that led directly to “Portrait of Edmond Belamy” being sold at auction. That answer was based on AI’s twin in contemporary technological hype, the blockchain.
Christie’s discovered Obvious via their work on Superrare, a blockchain-based “Rare Art” platform. Rare Art is named after the “Rare Pepe” project that developed the techniques of using cryptocurrency and blockchain token technology to record limited edition certificates for digital images. This produces “artificial scarcity” and allows a form of ownership for pieces of digital art art that would otherwise be infinitely reproducible. This use of certificates as ownership proxies for art was pioneered by conceptual art. Compared to a flammable piece of paper with a handwritten signature on, the authenticity of a blockchain transaction secured by a not inconsiderable fraction of the world’s computing power each day only increases over time. It may not be entirely clear what the authenticity is of, but the terms and conditions of Rare Art platforms and the community norms of their users and consumers do produce a vivid image of a novel and very strong concept of ownership.
“True digital ownership” on a blockchain secured by cryptographic keys is seen by its proponents as stronger, more trustworthy, and more absolute than previous conceptions of property. This makes AI art a natural fit for Rare Art because each has needs that the other fulfills: ownership in the case of the products of AI art, strongly perceptible uniqueness but also recognisability as art in the case of Rare Art. Sale at auction also provides this kind of closure for the financial value of art, but new art and in particular digital art faces a bootstrapping problem in which it must establish its value in order to be sold at auction but cannot be sold at auction without first establishing its value. Christie’s saw art by Obvious selling on Superrare and could react to that market signal more quickly and with lower risk than with signals from gallery or online sales of physical goods.
It is a truism of International Art English that art questions things. There are many questions in play in both AI Art and Rare Art. They involve the concept of ownership considered in terms of the law, of social norms, and of art theory. The answers to these questions from within each of these realms individually may be obvious and simple to their practitioners, but between them they may be more at odds than each realises. Negotiating this without closing the door to cultural creativity or opening it to corporate exploitation is a task that is of interest far beyond the artworld.
(I am not a lawyer, etc.)