AI Art and the Creativity Space of the Arts
1The key question of this essay is: How creative is the AI art that is starting to emerge today? I will begin with an overview of the “state of AI art.” Then, as a frame of reference to analyze this question, I briefly outline a model of aesthetic modernity in the second part. Thirdly, this model will be used to describe the creativity space of AI art in comparison to “analog” contemporary art.
I. The State of the AI Art
2Ahmed Elgammal, along with his team at Rutgers University, has developed an AI-based “almost autonomous artist” (Mazzone at al. 2019). The program is called AICAN and is based on a 2014 invention by Ian Goodfellow called generative adversarial network, or GAN for short (Elgammal at al. 2017). The use of GANs produces a specific Francis-Bacon-like AI style, also called GANism. Elgammal’s goal was to transform GANs into CANs, that is, to transform a generative adversarial network into a creative adversarial network by formulating an additional criterion. AICAN or “AI-CAN” is just the name for a special CAN, or rather: AICAN is a CAN with a pun.
3AICAN is structured in such a way that, on the one hand, it uses over 80,000 images from all styles and genres from the last 500 years of Western art to learn what distinguishes a painting that is art from a mere image that is not art. But on the other hand, AICAN is trained to ensure that a newly generated image does not resemble any known art style. In the end, the program managed to create remarkable abstract images (Fig. 1). Apparently, AICAN identifies figurative painting itself as a style and, in this respect, tries to avoid any realism.
4AICAN could have ended here without any major impact on the art world as a purely scientific research project. But the team took some additional steps. First, they developed a Turing test that invited human beings to compare the images generated by AICAN with images from Art Basel 2016. The astonishing result was that the test subjects could no longer distinguish whether an image had been created by a human or a machine (Elgammal at al. 2017). In addition, the paintings were shown in exhibitions and sold for high five-figure sums. And finally, Ahmed Elgammal and art historian Marian Mazzone published the essay “Art, Creativity, and the Potential of Artificial Intelligence” (Mazzone at al. 2019). This essay can certainly be read as a challenge to art philosophy, as it attempts to reformulate fundamental aesthetic terms. The authors make three thought-provoking statements concerning AICAN: AI art is art, AI art is conceptual art, and AI art is creative.
5Before I discuss these far-reaching theses, I would like to present a few other examples of AI art, such as the interactive installation of Memo Akten, Learning to see (Fig. 2). In the foreground, the exhibition visitor can move objects such as a rag and a charging cable on a table. A camera records these hand movements, which can be seen directly on the left screen. The same information is simultaneously sent to an artificial neural network that has been trained on ocean images, fire images, cloud images, and flower images. Accordingly, the network interprets the hand movements with the rag and the cable as water, fire, clouds, and flowers, which can be seen on the right screen.
6I consider Learning to see an essential work of AI art because it reflects its basic principles in a very poetic way: The Deep Learning systems that have made today’s emerging AI art possible learn to see the world from the data and images with which they have been trained.
7Helena Sarin took some photos in the department store “Bed, Bath and Beyond” and trained a GAN with them. She then selected some of these images and arranged them into a grid with a high degree of order and contrast, thereby maximizing – as is typical for an ornament – the experience of beauty (Lehmann 2016). The result is an aesthetically very attractive visual work with the title Bed, Bath and Way Beyond (Fig. 3). What I especially like about this work are the ironic titles. Everyday household items from the department store “Bed, Bath and Beyond” become abstract art—something “Way Beyond” ordinary life. And in addition, the whole series is called #latentShelfie. The neural network takes a selfie of itself, and what you see are abstract shelves or a “shelfie”—the essence of the data with which the AI has been trained (Baileyn 2018).
8Chris Peters trained GANs with landscape paintings of Tonalism, a romanticizing painting style that developed in America between 1880 and 1915. He then selected some of these AI images, such as Composition No. 4 (Fig. 4, left), and painted them himself with oil paint on canvas. The result is a symbiotic art form between AI and painter, which is also reflected in the title Symbiosis No. 4 (Fig. 4, right). In the logic of the art system, it would be considered kitsch if Peters were still painting in the style of Tonalism today. For that, one would need at least a good artistic reason, for instance, a strong concept, such as using AI as an image generator (Tensordream 2019).
9In Mario Klingemann’s installation Memories of Passersby I, the GANs were trained with portraits of men and women from the seventeenth to nineteenth centuries. Visitors are encouraged for a moment to take a seat in a chair in front of two screens and a wooden box that constantly creates new, unique portraits in GANism style (Fig. 5).
10While Memo Akten’s work demonstrates the basic principle of AI art—the algorithm sees what it has learned—Klingemann’s work makes visible the telos of all AI art, namely the infinite stream of artificially generated images. In Hegel’s words, the artwork shows a “bad infinity” unleashed by AI in the arts: “This infinity is the bad or negative infinity, in that it is nothing but the negation of the finite. The finite, however, arises again in the same way and is just as much not sublated. In other words: This infinity expresses only the intention of the sublation of the finite” (Hegel 1830 : § 94, Trans HL). In this respect, I consider Memories of Passersby I to be as much a key work of AI art as Learning to see by Memo Akten.
11If we look at the previous examples, it becomes clear that it is not the AI images on their own that are considered AI art, but that these AI images are first valorized into art. So far, we have observed at least four such strategies of art valorization: the Turing test, the use of irony, manual craft, and media-specific self-reflection.
12My last example of AI art is almost completely absorbed in such valorization strategies. It is also the example with the biggest media echo that AI art has triggered so far. The AI painting Edmond De Belamy (Fig. 6) was created by the French three-person artist collective Obvious and sold by Christie’s auction house for nearly half a million dollars. However, this tremendous success in the art market was met with some reluctance in the AI art community. Mario Klingemann said, for instance: “Pretty much everyone who is working seriously in this field [of AI art] is shaking their head in disbelief about the lack of judgment when it comes to featuring Obvious, and of course Christie’s decision to auction them out of all artists who work with neural networks” (Schneider at al. 2018). In general, the criticism was that the artist collective had only used a freely available GAN from the AI artist Robbie Barrat without truly programming and training it themselves. It was also said that the aesthetic quality of the AI image was rather poor.
13Indeed, the Paris-based artist collective was less interested in technical and aesthetic issues than in artistic valorization strategies, starting with the title. In French, the name Belamy sounds like “belle amie” and means “the lovely friend,” which in turn is a reference to the name of Goodfellow, the inventor of GANs (no 1). A second valorization is the signature, which consists of a mathematical formula underlying the AI algorithm used to produce the image (no. 2). In addition, the printing of the image is valorized by the gilt frame as a classical painting (no. 3). The work Edmond De Belamy is also part of a fictional family tree of eleven portraits. The claim below the illustration of that family tree: “The Belamy family tree, displaying centuries of art history, as seen from the standpoint of an artificial intelligence” (Obvious 2020 : 4)—is, of course, an ironic hoax (no. 4). Another valorization (no. 5) is this artistic “manifesto” in the spirit of the historical avant-garde (Obvious 2020 : 3). A press release (no. 6) with the provocative slogan “Creativity is not only for humans” is also part of this kind of art valorization with theoretical statements (Vincent 2018). And last but not least, the artist collective created an abstract logo referring to Leonardo da Vinci’s Vitruvian Man (no. 7). As in Renaissance times, a new unity of art and science is symbolized here, in the light of AI.
14The artist collective has virtually overloaded the work Edmond De Belamy with valorization strategies, revealing an internal conflict in the AI art community. The pioneers of AI art are, for the most part, professional programmers or software engineers and are rarely educated at art schools or socialized in art scenes. However, the artist collective from Paris seems to be quite familiar with art history, art theory, and the mechanisms of the art market. The subtle irony of the story is that technologically advanced AI art has been hacked by old-fashioned analog art—to use the jargon of digital natives.
II. Model of Aesthetic Modernity
15This brief overview of AI art highlights a contradiction: On the one hand, there is a view that Deep Learning systems themselves are creative and produce art (Elgammal at al. 2017, Mazzone at al. 2019). But on the other hand, AI artists also work with a variety of valorization strategies. Therefore, one could think that only the creative human intervention transforms an AI image into a work of art. In the following, I would like to briefly outline a model of aesthetic modernity based on which this contradiction can be discussed and analyzed. A detailed description and historical contextualization of this model can be found under the entry ‘Gehalt’ in the Handbook Art Philosophy (Sigmund 2022).
16The core idea of this model is that in the twentieth century, an internal differentiation process took place in the art system and that the major caesuras of modern art history can thus be reconstructed. This process consisted in a decoupling of the three essential components of art communication: the work, the medium, and the concept.
17In the fine arts, these three components were tightly coupled; today, there is art where these components are entirely decoupled. The driving force behind this history of differentiation is the ability of the modern arts to negate all conventional notions of art. Art only gained this ability for self-negation in the nineteenth century, when an autonomous art system began to develop that became independent of the directives of religion, politics, and law. Without such heteronomous restrictions, all art communication processes now run via a specific code with art-specific programs, leading to the art system’s operative closure (Luhmann 1997).
18Until the nineteenth century, the fine arts medium intensified the expression and experience of beauty. At the beginning of the twentieth century, the negation of the medium leads to the art of Classical Modernism with its isms. Here, specific concepts—such as the concept of Cubism or Expressionism—determine how the aesthetic material is to be arranged in an artwork (cf. theory of aesthetic eigenvalues (Lehmann 2016 : 33 – 79)).
19In the next step, the negation of the artwork constitutes the art of the historical avant-garde, whose most radical manifestation is conceptual art. Here, art is reduced to the perceptual representation of a concept. Strictly speaking, conceptual art can be defined by an isomorphism principle (Siegmund 2022): Conceptual art forms a one-to-one relationship between a concept and a percept, i.e., a simple perceptual impression (C ↔ P). Robert Rauschenberg’s Erased de Kooning drawing is exemplary for this isomorphism: the concept of the work is given by the title, namely, to erase a drawing by de Kooning, and what one can perceive is exactly such an erased drawing, and vice versa.
20In the context of AI art, the picture Edmond de Belamy is as much a prime example of conceptual art as Erased de Kooning Drawing is in the context of classical painting—especially since both works are marked as art by a gold frame. As with any conceptual art, Edmond de Belamy is extremely limited both in terms of aesthetic and content. The picture does not use all the aesthetic possibilities that AI art already possesses today, and it fails to articulate any world-related content that can be unlocked through interpretation. All of the many valorization strategies—the title, the frame, the signature, the family tree, the manifesto, and the logo—support the statement that this AI-generated image is art.
21Basically, the work comes down to the statement: AI art is art, making it an “analytical proposition” as described by conceptual artist Joseph Kosuth: “Works of art are analytic propositions. That is, viewed within their context—as art—they provide no information whatsoever about any matter of fact. A work of art is a tautology in that it is a presentation of the artist’s intention, that is, he is saying that that particular work of art is art, which means, is a definition of art.” Kosuth, of course, has not presented a general definition of art here, but a definition for so-called analytical conceptual art. The fact that the artist collective Obvious includes AI art “as art” in the art system according to the program of analytical conceptual art is, of course, brilliant. Such a spectacular crossing of boundaries can only be achieved once in the art system which in turn justifies the high auction price that the work has achieved at Sotheby’s (Kosuth 1969 : 134 – 137).
22Postmodern art overcomes the avant-garde by negating a strong taboo of the avant-garde, namely to work in the classical medium of fine art with its preference for the aesthetic value of beauty, central perspective, and figurative representation. We are dealing here with a double negation, a negation of the negation of the medium of the visual arts. In postmodern art, for example, figurative painting returns. But this return is now mediated through the concepts of irony and quotation. Roy Lichtenstein quotes comics, for example. This kind of conceptual mediatedness is represented in the model with a gap between the two segments.
23And in the meantime, contemporary art has also emancipated itself from postmodern art, with the help of a negation of the negation of the artwork. Thus, not only the use of the media but also the artwork itself is conceptually mediated now. The three gaps between the three segments in the circle symbolize not only this new kind of conceptual mediatedness, but they also signify new degrees of freedom in the production and reception of art gained in the art system of the 20th century.
24Two aspects, in particular, are important for our further considerations of AI art. First, with these newly gained freedoms of contemporary art, the creativity space of art also expands. And second, with this model, we can comprehend a paradigm shift that took place in the twentieth century.
25In Classical Modernism, newness was generated through the creation of new aesthetic material, which led to genuinely new aesthetic experiences for the recipient. Art is created here under the paradigm of material aesthetic. However, this idea of novelty was quickly exhausted. Sooner or later, in all art genres, people have noticed that it is no longer possible to create new isms and styles.
26But newness can also be generated by new aesthetic content, whereby I prefer to speak of Gehalt here. “Gehalt” is a non-translatable German term in the philosophical tradition of Hegel and Adorno (Lehmann 2022). For our purposes, the following short definition of Gehalt will be sufficient: Gehalt is content created through the interpretation of an artwork.
27Conceptual art thus appears as a turning point in modern arts. It marks a zero point in art history, where a paradigm shift from material aesthetic to gehalt aesthetic took place. This gehalt aesthetic turn in the twentieth century is essential for assessing the creativity space of AI art today.
III. Creativity Space of the Arts
28In this last part, I will try to place AI art in this theoretical framework, whereby I understand ‘AI art’ to mean any art created with the help of AI algorithms. What is crucial here is that this model also allows us to understand how the creativity space of the arts expanded in the twentieth century.
29The most striking side effect of Ahmed Elgammal’s visual Turing test is that Art Basel and AICAN arrive at the same kind of universal abstract modernism. And the explanation for this is that the AI program and the art dealers solve the same optimization problem. The question that has troubled and haunted art fairs since the end of Classical Modernism is: How can the attractiveness of images be optimized when all possibilities of radical aesthetic innovation have been exhausted, and new styles and isms can no longer be invented?
30At that point, uncompromising artists like Marcel Duchamp abandoned the exhausted genre of painting and created conceptual art instead. However, the galleries that exhibit at important art fairs like Art Basel cannot simply follow this logic of the art system because they sell high-priced art to a clientele for whom the works also have a subtle “use value.” Art serves as a status symbol with which villas, yachts, or large apartments are furnished. For this purpose, the art must be clearly recognizable as art and aesthetically pleasing, which is not the case with conceptual art.
31Under these additional restrictions, the art market will, on the one hand, stick to the genre of painting because a painting can be clearly identified as art. On the other hand, it will also try to avoid any recognizable style because any known style or ism looks old-fashioned – not contemporary or new – to art collectors and buyers. And it is precisely this style avoidance that is also encoded in AICAN: “The machine is trained between two opposing forces—one that urges the machine to follow the aesthetics of the art it is shown (minimizing deviation from art distribution), while the other force penalizes the machine if it emulates an already established style (maximizing style ambiguity). These two opposing forces ensure that the art generated will be novel but at the same time will not depart too much from acceptable aesthetic standards” (Mazzone at al. 2019). The “acceptable aesthetic standards” toward which AICAN was programmed are thus by no means universal standards for artistic creativity, but rather the aesthetic standards of a particular social milieu in which art is part of the lifestyle.
32AICAN’s successfully passing the Turing test will not remain without consequences for the art market. As soon as it becomes clear that this “almost autonomous artist” (Ibid) is creative in the same way as those painters from Art Basel, this will lead to a demystification of this style of painting on the high-priced market. Presumably, it will focus more on figurative painting, since the semantic content of such images cannot be generated with any AI available in the foreseeable future.
33On the other hand, abstract modernism will be democratized because of its technical producibility and will turn into individualized design—a design that manifests in unique pieces, like art, and can be bought for low sums of money. In liberal democracies, this development could certainly be understood as a gain towards greater possibilities of individualization.
34At the same time, an AI art that passes a Turing test also sharpens our concept of creativity. As can be seen, the creativity space of AICAN is strongly limited because it is a form of creativity that arises from infinite variation and does not produce anything qualitatively new such as a new genre, a new style, a new ism or a new gehalt. As I mentioned earlier, this form of novelty can be characterized as a kind of “bad infinity.” The new work emerges from a negation of all combinations of abstract forms and colors ever created; in this sense, each new AICAN painting is unique and, accordingly, a creative achievement that has never existed in this guise before. At the same time, it is also an expectable manifestation of the new, which remains within the limits of abstract modernism.
35This deficient mode of infinity, which occurred sooner or later in all genres of Classical Modernism, was finally overcome with the help of conceptualizations. In the context of art, “concepts” are linguistic determinants that channel the reception of art—this happened implicitly in the fine arts and explicitly in the reflexive arts. After the differentiation of medium, work, and concept and the decoupling of concepts from both medium and work, concepts can either define the use of media, or limit the interpretive scope (the gehalt) of artworks in advance, or both.
36This kind of contemporary art is not conceptual art in the strict sense of the isomorphism principle; it is concept-based art. It also includes the works of other AI artists, which we discussed in the first part.
37The concept of Chris Peters specifies the medium of painting as “AI tonalism” without evoking a reality-based interpretation of these images—hence the formal title Symbiosis 4. The concept of Helena Sarin specifies both the medium of her work as a medium for the formation of bed-bath-and-beyond images and the gehalt with the title Bed, Bath and Way Beyond. Part of the concept of Memo Akten’s work, on the one hand, is that one knows that his AI has been trained (among other things) with cloud images and that the visual medium has been prepared accordingly for the formation of cloud forms. On the other hand, there is also a narrowing of the interpretation through the title Learning to see so that one reads the work as a self-reflection of Deep Learning systems that underlie AI art today. The concept of Mario Klingemann’s work uses the title to establish a particular reading of the infinite sequence of unique GAN-generated portraits, namely that these are “Memories of Passersby.” From a philosophical perspective, this title makes the work weaker than it actually is because it is a great allegory for that bad infinity that becomes visible all at once with AI art in the art system.
38In comparison with the output of AICAN, such concept-based AI art has a much broader creativity space because it is essentially about a novel combination of AI images with concepts. However, I would guess that with the anticipated development of an Artificial General Intelligence (AGI for short), even such concept-based art will face competition from AI. Valorization strategies and art concepts, in general, can be found in thousands and thousands of works in art history. Art concepts are not always invented for the first time but are usually copied and modified. Once such repetitions occur, AGI systems will also be able to recognize these patterns and operate with them. Chris Peters’ concept can be copied and applied to other styles and genres, not only by other human artists but most likely by an AGI. Thus, I would expect AGI systems in the future to generate a variety of concepts that define the framework in which AI images are to be perceived and interpreted.
39What does that mean? Could the development of AGI eventually lead to the human artist being supplanted by an autonomous AI artist? Or is there such a thing as a human creativity space that is inaccessible to machines? To be able to discuss this question, we need to look at the entire creativity space that has opened up for the arts in the twentieth century.
40Let us take an example that can stand for an immanent distinction between medium, work, and concept, such as Maurizio Cattelan’s La Nona Ora (Fig. 8). The sculpture portrays Pope John Paul II being struck by a meteorite and crying out as Jesus once did on the cross: “My God, my God, why have you forsaken me?” The title The Ninth Hour hints indirectly at this Bible quotation: “And about the ninth hour Jesus cried with a loud voice, saying, Eli, Eli, lama sabachthani? that is to say, My God, my God, why hast thou forsaken me?” (Matthew 27: 46).
41Cattelan’s work is usually understood as blasphemy, but actually, the work also proves the existence of God. The probability is so extremely low by earthly standards that a meteorite would strike the pope that divine intervention would be needed for this. The artwork thus simultaneously proves the existence and non-existence of God, which in turn can be seen as a metaphor for the status of religion in liberal democracies: No one is forced to believe in God and everybody has to tolerate the faith of others. In an interview about La Nona Ora, Cattelan said: “I want religion and blasphemy to collide, as they do in our daily life.” He certainly succeeded—probably in an even more comprehensive way than he had intended (Cattelan 2001).
42This once again raises the crucial question: Is it conceivable that, at some point in the future, artificial general intelligence could create such artwork? The work has only five elements: the pope, the meteorite, the broken glass, the red carpet, and the Bible reference. What requirements would an AI have to fulfill to be able to assemble an artwork from five such building blocks? One reason why the development of artificial general intelligence has failed so far is that computers have no common sense: “Many people believe that until AI systems have common sense as humans do, we won’t be able to trust them to be fully autonomous in complex real-world situations” (Mitchell 2019 : 104). This refers to the infinite universe of background knowledge that humans have and gives them an idea of how the world of physical objects and social relations does work in principle. Common sense is also required in the reception of La Nona Ora, for example, to understand that this stone is a meteorite. The artwork does not directly provide you with that information. However, the broken glass on the red carpet activates the viewers’ common sense and suggests a meteorite impact.
43The real problem for AI art, however, is not that such works presuppose common sense knowledge but rather that they create uncommon sense. Artworks like this are the real “singularity” on earth. As a compact synthesis of meaning, such artwork is incomparable to anything that exists in reality. I suspect even artificial general intelligence could learn nothing from thousands of such examples. The synthesis of meaning that a human artist accomplishes in such works does not follow any general criteria; it is highly idiosyncratic.
44The decision that these five elements form a context of meaning that might be relevant to another person or to society is based on the unique self-organization of the artist’s human brain. And here, early childhood imprints, conversations, unhappy loves, books, mental disorders, and all sorts of things can play a role. An AI that could create works like The Ninth Hour would have to be an AI that can subjectify itself. Such AI subjectification cannot simply be equated with AI embodiment or AI consciousness, which is often cited as a prerequisite for the development of AGI. Rather, such subjectified AI would require self-awareness and social interaction with other subjectified AIs or humans. Actually, such a hypothetical AI subject would have to live a life like a human being.
45The AI artworks by Obvious, Peters, Sarin, Akten, and Klingemann discussed here may ultimately also be based on concepts that can no longer be found in the sphere of common sense. Precisely because AI art—for the first time—can exploit its own new medium, all these works are unique and brilliant in their own way. However, the line between the common and the uncommon sense in AI art is not nearly as clear-cut as in Cattelan’s installation, which is why his work is better used as a benchmark for artistic creativity.
46The creativity space of art is a very broad field. There is the creativity of infinite variation, of which AI art algorithms such as AICAN are already capable today. There is concept-based creativity in art, which would require—at least—a general artificial intelligence with common sense knowledge. And there is the creativity of the uncommon sense, in relation to which the notion “artificial intelligence” must be subjectified and the word “artificial” loses all its meaning. One would be dealing with artificial intelligence that must draw on the resources of natural human intelligence in order to be able to generate the unusual meaning of reflective art.
47My hypothesis would be that AI art can unfold and prosper under the exhausted paradigm of material aesthetics. However, the idea that an autonomous AI artist could also create full-blown gehalt aesthetic artwork seems to be a fallacy. For the humanities, however, this fallacy would be great news.
References
- Bailey, Jason (2018): Helena Sarin: Why Bigger Isn’t Always Better With GANs And AI Art. In: Artnome, November 26. https://www.artnome.com/news/2018/11/14/helena-sarin-why-bigger-isnt-always-better-with-gans-and-ai-art .
- Cattelan, Maurizio (2001): Lot Essay. https://www.christies.com/en/lot/lot-2051684.
- Elgammal, Ahmed, Bingchen Liu, Mohamed Elhoseiny and Marian Mazzone (2017): CAN: Creative Adversarial Networks, Generating ‘Art’ by Learning About Styles and Deviating from Style Norms. Conference Presentation, ICCC, Atlanta, GA. https://www.researchgate.net/publication/317823071_CAN_Creative_Adversarial_Networks_Generating_Art_by_Learning_About_Styles_and_Deviating_from_Style_Norms .
- Hegel, G. W. F. (1989 [1830]): Enzyklopädie der philosophischen Wissenschaften im Grundrisse. In: Werke 8, hg. v. E. Moldenhauer u. K. M. Michel. Frankfurt am Main: Suhrkamp.
- Kosuth, Joseph (1969): Art after Philosophy, Part I. In: Studio International, October: 134–137.
- Lehmann, Harry (2016): Gehaltsästhetik. Eine Kunstphilosophie. Paderborn: W. Fink.
- Luhmann, Niklas (1997): Die Kunst der Gesellschaft. Suhrkamp: Frankfurt am Main.
- Mazzone, Marian and Ahmed Elgammal (2019): Art, Creativity, and the Potential of Artificial Intelligence. In: Arts 8. Doi: 10.3390/arts8010026.
- Obvious (2020): Artificial Intelligence for Art, p. 4. http://obvious-art.com/wp-content/uploads/2020/04/MANIFESTO-V2.pdf.
- Schneider, Tim and Rea Naomi (2018): Has Artificial Intelligence Given Us the Next Great Art Movement? Experts Say Slow Down, the ‘Field Is in Its Infancy’. In: Artnet, September 25. https://news.artnet.com/art-world/ai-art-comes-to-market-is-it-worth-the-hype-1352011 .
- Siegmund, Judith (2022): Gehalt (Harry Lehmann). In: Handbuch Kunstphilosophie, edited by Judith Siegmund. Tübingen: UTB, 215–225.
- Tensordream (2019): https://tensor.chrispeters.com/
- Vincent, James (2018): How Three French Students Used Borrowed Code To Put the First AI Portrait in Christie’s. In: The Verge, October 23. https://www.theverge.com/2018/10/23/18013190/ai-art-portrait-auction-christies-belamy-obvious-robbie-barrat-gans .