In his seminal 1967 essay, Roland Barthes declared “The Death of the Author” to encourage a stronger relationship between a written text and its reader over a relationship between the reader and the written text’s Author. In Barthes’ view, “To give a text an Author is to impose a limit on that text, to furnish it with a final signified, to close the writing”.[1] By interpreting a text through the analysis of its authorship (any biographical information, subjective contexts of the Author’s life including their mental/psychological states at the time of writing) one ultimately perverts the true meaning of the text which should be interpreted through a lens that, at length, divorces authorship from content. The Author-god complex as an analytical framework of a text must pass into oblivion in order for the future of reading to flourish, as Barthes concluded “the birth of the reader must be at the cost of the death of the author”.[2] Certainly, one can apply Barthes’ framework of authorship and identity to a myriad of creative practices where the meaning of both content and its creator seemingly blend into the audience’s interpretation of the work itself; painters and their paintings, musicians and their compositions, actors and their performances or directors and their films. However, technological advancements unique to the 21st century have brought about a new form of creative practice; one which derives its unique aesthetic from the mechanical production of a work rather than from the imagination of the Author himself. The written prose of the Author is transformed into the written code of the software, which in turn produces what can broadly be classified as "Machine Learning" production. Machine Learning artistic practice can be achieved through a variety of technologies such as robotics, Artificial Intelligence, Deep Fakes and GANs (Generative Adversarial Networks), all which produce their own, unique style of creative output ranging from poetry to classical music to renaissance portraits. The latter of these technologies, GANs, will serve as the case study in order to approach the conflicts, authorship politics and aesthetic value of Machine Learning produced “fine art” within Barthes’ post-structuralist framework.
Overview of Generative Adversarial Networks (GANs)
A Generative Adversarial Network is a Machine Learning (also known as “deep learning”) system that is programmed to recognize patterns to mimic for output via the intricate relationship between two entities: a generator and a discriminator. The generator produces (generates) images according to the large database of similar images it is programmed to study and replicate, and the discriminator subsequently rates, judges and then either accepts or rejects the effectiveness of the generator’s output. For instance, if a GAN is told to generate a human face, the two entities, trained on a large database of human face recognition, are in a constant conversation with each other as they attempt to learn the patterns, nuances, familiarities and “rules” of the human face (i.e., typically two eyes, a few inches under the eyes there is a nose and below that a mouth etc.) The generator will attempt to produce a semblance of a human face and the discriminator will either accept its endeavour or inform the generator it has to analyze more data and try again. What generally begins as outputs of white noise and murky renderings gradually progresses to the final output; a completely original, typically realistic human face. It must be stated that GAN generated images are not composite images; they are not a combination of its database nor can they clearly indicate a referent. Rather, GAN generated images are wholly new, original works that have been trained on a database to produce a unique output.
Though the human face serves as a relatively simple example of what a GAN system can produce, as of late, advancements in Machine Learning outputs have taken on more sophisticated forms of creative practice, some of which are beginning to seriously compete within contemporary art markets.
Does GAN Art have an Author?
In 2018, an artwork titled “Portrait of Edmond de Belamy” was sold at the world-famous Christie’s auction house for an astonishing $432,500 USD. According to Christie’s official description of the work, “The portrait in its gilt frame depicts a portly gentleman, possibly French and — to judge by his dark frockcoat and plain white collar — a man of the church. The work appears unfinished: the facial features are somewhat indistinct and there are blank areas of canvas”.[3] Edmond de Belamy is not a portrait of a historical figure, in fact Edmond de Belamy is not a human at all, he is a character created by a Generative Adversarial Network trained on a database 15,000 portraits painted between the 14th – 20th century. Edmond de Belamy is simply an output of a Machine Learning process or, as Christie’s claimed, “the arrival of AI art on the world auction stage”.[4]
Code as Author
It is true that one could argue that the only process, as outlined in Obvious’ manifesto, that constitutes an act of creation would be “building the algorithm”- which entails writing the instructional code the GAN or Machine Learning network learns and builds from. In fact, rather than signing their names or their collective’s name onto the portrait itself, Obvious signed the portrait with part of the code the GAN output was trained on:
Computer code converts letters, symbols and numbers into information which the software then uses as a series of instructions. In the age of information, code is oftentimes romanticized as a language itself and has been described as eloquent, beautiful or poetic (i.e., to say “she writes beautiful code”). Could one, then, think of writing code as a kind of prose- comparable to writing an epic novel, a screenplay or a book of poetry? In so far as that may be true- could one, then, assign authorship to a piece of code? Paradoxically, the Obvious Collective found themselves at the centre of this very debate. Shortly after “Edmond de Belamy” was put up for auction at Christie’s it was revealed that the art collective didn’t actually write the code for the GAN piece themselves, rather, the code was written by 19-year Robbie Barrat; a recent high school graduate and a GAN artist in his own right.[8] Barrat never claimed ownership over the code itself, in fact he publicly published the design on an Open-Source coding website allowing it to be widely disseminated without copyright infringement. However, Barrat did take issue with Obvious’ use of his original code, claiming that their massive profit within institutionalized art markets went against the democratic ethos of the Open-Source movement. Barrat never sued the collective, nor did he try to claim a portion of the $432,500 USD Christie’s sale, however, he publicly voiced his frustration over social media. In so far as code is a form of prose (writing), would Barrat then be the rightful Author of “Edmond de Belamy”? Working within Barthes' post-structuralist framework, since Barrat was not aware his code was being used for the creation of “Portrait of Edmond de Belamy” what authorial intent exists that can potentially taint a reading of the artwork itself?
AI-ness as Author
Perhaps in this case, Barrat does not have the “Author-god” complex Barthes speaks of, but rather it is the paratext surrounding the event which can influence the interpretation of the artwork, presenting conflicts in an objective reading. One could argue that both the Obvious/Barrat controversy and the sheer novelty of “Portrait of Edmond de Belamy” is what gives the piece of an identity that is separate than solely its form, aesthetic and execution. Barthes writes “Writing is that neutral, composite, oblique space where our subject slips away, the negative where all identity is lost, starting with the very identity of the body of writing”.[9] “Portrait of Edmond de Belamy” was sold for nearly 40x the estimated price not because it was a beautiful piece of art, but rather because it was unconventional. If one were to separate the paratext (circumstances of production) from the text (the piece itself), the aesthetic information of “Portrait of Edmond de Belamy” would read quite differently. However, one could argue that it is the novelty attraction of auctioning a piece of art made by a Machine Learning network that solidified the art’s $432,500 USD value. So, although one cannot ascribe the signifiers of particular authorial or biographical data to the portrait, certainly the circumstances of its production still imposes a value judgement on the work. Is the fact a work like “Portrait of Edmond de Belamy” is seen as ground-breaking, innovative or ultramodern clouding rational, objective judgement of its aesthetic value? Scholarship surround GAN art is often written through a lens of the technological singularity; the theory that accelerated technological development will lead to drastic changes in human civilization. Within the discourse surrounding Machine Learning art (whether the output is visual art, written work etc.) questions are constantly raised pertaining to the future of the role of the artist or the sentience of the machine; Are artists now obsolete? Will all art be produced my machines? Is artificial intelligence covertly trying to deliver a message to humanity through the production of artworks? Though many marvel at the aspect of intelligent machines creating artistic content, there is very little scholarship or critique on the actual aesthetic value of a Machine Learning work, divorced from its unconventional production. As Christie’s definition of “Portrait of Edmond de Belamy” noted, the piece appears “unfinished”.[10] Would a different, contemporary artist be able to showcase and profit more than $400,000 USD off of an unfinished piece of art? Perhaps not, unless said artist was considered to be “one of the greats” or infamous in their own right (i.e., a Picasso, Rembrandt or DaVinci). Once again, this bias ties into the notion that the value and meaning of a work is directly tied to its authorship. As Barthes writes “To give a text an Author is to impose a limit on that text, to furnish it with a final signified, to close the writing”.[11] Similarly, to ascribe meaning/value to the mode of production over the content itself is to impose a limit on the work itself, for at the end of the day it can be seen as nothing more than a promethean step towards the singularity.
Database as Author
In lieu of the “Author-god”, Barthes suggests there should be a reorientation from Author as meaning maker to, more simply, “scriptor” as writer. Barthes writes that no text produces a “single ‘theological’ meaning (the message of the author-god) but a multi-dimensional space in which a variety of writings, none of them original, blend and crash”.[12] The suggestion here is that no Author can claim ownership over a work with such authority because no work can truly be considered wholly original. Further to this point, Barthes writes “Succeeding the Author, the scriptor no longer bears within him passions, humours, feelings, impressions, but rather this immense dictionary from which he draws a writing that can know no halt”.[13] To acknowledge what the act of writing truly is; a blend of pre-existing influences, ideas, theories etc., is to strip the Author of his god-like complex and the power he presides over the reader’s interpretation of a text. “The writer can only imitate a gesture that is always anterior, never original” Barthes writes, “His only power is to mix writing”.[14] In so far as that may be true, GAN art presents interesting conflicts to Barthes’ “The Death of the Author” for it is precisely the act of mixing the “immense” dictionary that serves as the main, visible function of the GAN’s processes of creation. Whereas Barthes’ text is shedding light on the invisible reality that the Author is truly a collage maker (i.e., using pre-existing ideas and subsequently claiming those ideas as wholly original concepts in his writing), GAN art is entirely predicated on a collage-maker ethos. So much so, in fact, that the “immense dictionary” itself perhaps takes on new authority- one that might even rival Barthes’ notion of authorial supremacy. To reiterate- in GAN art production the network itself is contingent upon a vast database of images that closely resemble the desired output. The two entities of the network, the generator and discriminator, are working in tandem to produce the closest, possible mimesis of the database by intently analyzing, studying and learning the aesthetic rules of the vast number of images. Though GANs do produce a new work (not a composite of various images) it is, in its very essence, drawn from pre-existing media. Much like the Author Barthes refers to, the database itself can have an insidious influence over the meaning and/or perceived value of the artwork it produces. NYU Professor Kate Crawford and artist Trevor Paglan have conducted extensive research into the social and political consequences of image categorization systems used for Generative Adversarial Networks and other Machine Learning networks. “Training sets are the foundation on which contemporary machine learning systems are built” Paglan and Crawford write in their seminal work Excavating AI, “They are central to how AI systems recognize and interpret the world”.[15] For instance, in the case of “Portrait of Edmond de Belamy”, the final output was trained on a database of around 15,000 portraits. Though this is an impressive number of training sets, a central question remains- how was the database of 15,000 portraits determined? In terms of contemporary GAN art, the most popular tool, what Paglan and Crawford describe as the “canonical training set”[16] is the database hosting website ImageNet. ImageNet is an online dataset that consists of hundreds of thousands of images, from human faces, to animals, artwork, items and beyond. The project was undertaken by Stanford Professor Fei-Fei Li in 2009, who claimed the idea behind the database was to “map out the entire world of objects”.[17] Though the immense undertaking of mapping out, categorizing and classifying hundreds of thousands of images may appear to be like “sorcery” in Crawford and Paglan’s terms,[18] one must not forget that ImageNet’s endeavour is a firstly product is human labour to subsequently be used for Machine Learning. Li and her research team hired labourers through the “Amazon Mechanical Turk”; an outsourced, virtual labour platform.[19] “While technology continues to improve, there are still many things that human beings can do much more effectively than computers” claims the platform’s official description, “Amazon Mechanical Turk is a crowdsourcing marketplace that makes it easier for individuals and businesses to outsource their processes and jobs to a distributed workforce who can perform these tasks virtually”.[20] Contractors of Amazon’s Mechanical Turk are, on average, paid less than minimum wage and there are little to no barriers of entry into its virtual workforce.[21] ImageNet is a prominent example of an online system that is, as of right now, dependent on human labour rather than computer power because of the nuances and idiosyncrasies surrounding object or human image categorization. For example, rather than categorize a photo of a human face as simply “person” – ImageNet classifies images according to synset (natural object), category (person), subcategory (human face) and subclass (adult female face). Though this is a rather straightforward example, Crawford and Paglan note that ImageNet contains 2,833 subcategories under the top-level category “person”.[22] It is within this space of intricate, precise categorization where a database such as ImageNet begins to show more political or social consequences. Crawford and Paglan write “As we go further into the depths of ImageNet’s Person categories, the classifications of humans within it take a sharp and dark turn. There are categories for Bad Person, Call Girl, Drug Addict, Closet Queen, Convict, Crazy, Failure, Flop, Fucker, Hypocrite” and their list continues on.[23] Though there is obvious value in attempting to create a world-wide database of images for research, artificial intelligence and Machine Learning networks, there are inherent conflicts and consequences of attempting to categorize humans as mere objects. Devaluation and degradation are rampant in a database like ImageNet. The reductionist approach of trying to categorize humans into specific subsets based on visual information harkens back to the dehumanization of persons based on race, gender, eye colour or even skull shape (the practice of phrenology). To add to Crawford and Paglan’s study, similarly, one could argue a database like ImageNet presents further problems in attempting to not only categorize humans as objects but art as objects as well, as the taxonomy of a given artwork cannot be reduced to one, single visual style. For instance, what exactly constitutes a work of avant-garde art if it stripped away from its political and social implications? As outlined by Clement Greenberg in his 1939 essay “Avant-Garde and Kitsch”, broadly, the ethos of avant-garde art is “art for art’s sake”, or, as Greenberg claims “the avant-garde poet or artist sought to maintain the high level of his art by both narrowing and raising it to the expression of an absolute in which all relativities and contradictions would be either resolved or beside the point.”[24] In its abstractive ethos and experimentalism, avant-garde art can take form in painting, collage, poetry, dance and various other mediums. There is no singular, defining visual style that encompasses the avant-garde, for it is was the ideology that defined the movement- a rebellion against Alexandrianism and classical forms of production that had so far dominated the landscape. An art movement like the avant-garde evades the possibility of narrow classification or taxonomy when divorced from its ontological roots; it is, in a way, dependant on the very norms (Alexandrianism) which it attempts to differentiate itself from. How can one first qualify and then classify the works of art, such as the avant-garde, which are bound to ideology rather than a singular aesthetic? Further to this point, must there also be a level of knowledge or education necessary to classify a work of art as avant-garde (as opposed to a work of renaissance, kitsch or neo-romanticism)? Certainly, not all the labourers that are contracted by Amazon’s Mechanical Turk services can be well-versed in art history, and so, as Crawford and Paglan explain “Images are laden with potential meanings, irresolvable questions, and contradictions. In trying to resolve these ambiguities, ImageNet’s labels often compress and simplify images into deadpan banalities.”[25] In this case of “Portrait of Edmond de Belamy” what were the exact perimeters that defined and classified the 15,000-portrait database that served as the GAN’s training set? The Obvious collective never explicitly explained how the database was chosen, whether it was through outsourced labour or their own classification system, but an interesting question is raised as to what power the database has over the final output itself and whether the database, fraught with the political and social nuances and complexities mentioned above, can claim some sort of authorship over the final, produced work- regardless of the fact the it may not constitute a wholly artistically or historically accurate training set.
Conclusion
Pertaining to the production of a given work, it may appear as though the presence of a creative, intelligent machine like a Generative Adversarial Network constitutes the absence of a singular Author, and thus, is perfectly compatible with Barthes’ “The Death of the Author”. However, paradoxically, although it is nearly impossible to attribute the output to one, single Author or entity, Machine Learning is still fraught with the same conflicts of conflating the meaning of content with its authoritative creator. Perhaps the artificial intelligence art sector cannot simply kill the Author, in the fashion Barthes wished for, but conceivably, with time, it might kill the paratext or circumstances that has so far been tainting GAN art’s reception and reading by its audience. If this style of artistic technological production becomes more common place, for instance, maybe the novelty of AI-ness as Author will cease to have such a great influence over interpretation. Likewise, as databases (like ImageNet) and training sets become more sophisticated, perhaps they will cease to be as controversial and wholly unstable as they currently appear. In short, once more art is produced by Machine Learning networks, leading to familiarity, perhaps it is only then that one can say with confidence these new forms of artistic creation are truly compatible with Barthes’ “The Death of the Author”. Until then- it is the dissonances, conflicts and incongruities that exist within the medium’s production that will take on the role of the Author, leading to an equally contaminated interpretation of what should be an otherwise objective evaluation of the outputs/artworks.
Footnotes
[1] Roland Barthes and Stephen Heath, “The Death of the Author,” in Image-Music-Text: Roland Barthes (Glasgow: Collins, 1977), pp. 147, https://doi.org/http://artsites.ucsc.edu/faculty/Gustafson/FILM%20162.W10/readings/barthes.death.pdf.
[2] Roland Barthes and Stephen Heath, “The Death of the Author,” in Image-Music-Text: Roland Barthes (Glasgow: Collins, 1977), pp. 148, https://doi.org/http://artsites.ucsc.edu/faculty/Gustafson/FILM%20162.W10/readings/barthes.death.pdf.
[3] “The First Piece of AI-Generated Art to Come to Auction,” Christie's, December 12, 2018, https://www.christies.com/features/A-collaboration-between-two-artists-one-human-one-a-machine-9332-1.aspx.
[4] “The First Piece of AI-Generated Art to Come to Auction,” Christie's, December 12, 2018, https://www.christies.com/features/A-collaboration-between-two-artists-one-human-one-a-machine-9332-1.aspx.
[5] “Obvious Art: Manifesto,” Obvious Art, http://obvious-art.com/wp-content/uploads/2020/04/MANIFESTO-V2.pdf.
[6] Roland Barthes and Stephen Heath, “The Death of the Author,” in Image-Music-Text: Roland Barthes (Glasgow: Collins, 1977), pp. 147, https://doi.org/http://artsites.ucsc.edu/faculty/Gustafson/FILM%20162.W10/readings/barthes.death.pdf.
[7] Roland Barthes and Stephen Heath, “The Death of the Author,” in Image-Music-Text: Roland Barthes (Glasgow: Collins, 1977), pp. 143, https://doi.org/http://artsites.ucsc.edu/faculty/Gustafson/FILM%20162.W10/readings/barthes.death.pdf.
[8] James Vincent, “How Three French Students Used Borrowed Code to Put the First AI Portrait in Christie's,” The Verge (October 23, 2018), https://www.theverge.com/2018/10/23/18013190/ai-art-portrait-auction-christies-belamy-obvious-robbie-barrat-gans.
[9] Roland Barthes and Stephen Heath, “The Death of the Author,” in Image-Music-Text: Roland Barthes (Glasgow: Collins, 1977), pp. 142, https://doi.org/http://artsites.ucsc.edu/faculty/Gustafson/FILM%20162.W10/readings/barthes.death.pdf.
[10] “The First Piece of AI-Generated Art to Come to Auction,” Christie's, December 12, 2018, https://www.christies.com/features/A-collaboration-between-two-artists-one-human-one-a-machine-9332-1.aspx.
[11] Roland Barthes and Stephen Heath, “The Death of the Author,” in Image-Music-Text: Roland Barthes (Glasgow: Collins, 1977), pp. 147, https://doi.org/http://artsites.ucsc.edu/faculty/Gustafson/FILM%20162.W10/readings/barthes.death.pdf.
[12] Roland Barthes and Stephen Heath, “The Death of the Author,” in Image-Music-Text: Roland Barthes (Glasgow: Collins, 1977), pp. 146, https://doi.org/http://artsites.ucsc.edu/faculty/Gustafson/FILM%20162.W10/readings/barthes.death.pdf.
[13] Roland Barthes and Stephen Heath, “The Death of the Author,” in Image-Music-Text: Roland Barthes (Glasgow: Collins, 1977), pp. 147, https://doi.org/http://artsites.ucsc.edu/faculty/Gustafson/FILM%20162.W10/readings/barthes.death.pdf.
[14] Roland Barthes and Stephen Heath, “The Death of the Author,” in Image-Music-Text: Roland Barthes (Glasgow: Collins, 1977), pp. 146, https://doi.org/http://artsites.ucsc.edu/faculty/Gustafson/FILM%20162.W10/readings/barthes.death.pdf.
[15] Kate Crawford and Trevor Paglan, “Excavating AI: The Politics of Training Sets for Machine Learning” (The AI Now Institute, NYU, September 19, 2019), https://www.excavating.ai/.
[16] Kate Crawford and Trevor Paglan, “Excavating AI: The Politics of Training Sets for Machine Learning” (The AI Now Institute, NYU, September 19, 2019), https://www.excavating.ai/.
[17] Kate Crawford and Trevor Paglan, “Excavating AI: The Politics of Training Sets for Machine Learning” (The AI Now Institute, NYU, September 19, 2019), https://www.excavating.ai/.
[18] Kate Crawford and Trevor Paglan, “Excavating AI: The Politics of Training Sets for Machine Learning” (The AI Now Institute, NYU, September 19, 2019), https://www.excavating.ai/.
[19] Kate Crawford and Trevor Paglan, “Excavating AI: The Politics of Training Sets for Machine Learning” (The AI Now Institute, NYU, September 19, 2019), https://www.excavating.ai/.
[20] “Amazon Mechanical Turk”, accessed December 7, 2020, https://www.mturk.com/.
[21] Alana Semuels, “The Internet Is Enabling a New Kind of Poorly Paid Hell,” The Atlantic (Atlantic Media Company, January 23, 2018), https://www.theatlantic.com/business/archive/2018/01/amazon-mechanical-turk/551192/.
[22] Kate Crawford and Trevor Paglan, “Excavating AI: The Politics of Training Sets for Machine Learning” (The AI Now Institute, NYU, September 19, 2019), https://www.excavating.ai/.
[23] Kate Crawford and Trevor Paglan, “Excavating AI: The Politics of Training Sets for Machine Learning” (The AI Now Institute, NYU, September 19, 2019), https://www.excavating.ai/.
[24] Clement Greenberg, "Avant-Garde and Kitsch." Partisan Review 6, no. 5 (1939): pp.2
[25] Kate Crawford and Trevor Paglan, “Excavating AI: The Politics of Training Sets for Machine Learning” (The AI Now Institute, NYU, September 19, 2019), https://www.excavating.ai/.
Comments