The culture of neural networks: Synthetic literature and art in (not only) the Czech and Slovak context
Abstract
The possibilities of generated cultural production have undergone fundamental changes in recent years, leading to a rethinking of existing approaches to the text and the artwork as such. To grasp this process, Zuzana Husárová and Karel Piorecký propose the term “neural network culture,” which captures a wide range of generative practices and reception mechanisms. The Culture of neural networks contextualizes the phenomenon of literary texts and other artifacts generated using the latest technological techniques. The generation of literary texts using neural networks is part of a broader cultural process, to which this publication formulates a position through the lens of literary science, media theory, and art theory. The scholarly debate over this topic has been inconsistent—on the one hand, it underestimates the diachronic connections between generated texts and the tradition of experimental and conceptual literature; on the other hand, it does not sufficiently clarify the new-generation procedures and the contribution of human and technological actors in them. Therefore, Husárová and Piorecký propose the notion of synthetic textual art, which reflects the specific roles of the different actors involved in generative practice and its intermedial nature. In doing so, they approach the topic from both historical and theoretical perspectives, analyzing the current state of generative practice in all three basic literary types and in the intermedial space using selected foreign and Czech-Slovak projects. This state of affairs is often distorted in media discourse and even mythicized in terms of the capabilities of “artificial intelligence”; therefore, a critical analysis of this media discourse is essential. Finally, the authors summarize the implications of this stage in the development of generative practice on creativity theory and literary theory.
Full text
“Disjunctive, vernacular, regional and nuanced, Th e CoNN off ers insightful tendrils into an ecosystem of art practices emerging at the intersection of language and neural-networks. It swift ly and thoroughly covers many items of machine-literature history, unravels the language around AI and proves through an abundance of clearly-organized examples that ‘neural networks are not merely a technological tool or even a new generation of soft ware, but a phenomenon with distinct culture-creating potential’.” — DAVID JHAVE JOHNSTON, author of ReRites and Aesthetic Animism “Piorecký and Husárová expertly leverage the collective knowledge and critical methodologies the fi eld of electronic literature has developed for decades on text generation, bots, multimodality, and their technological and cultural contexts to guide us into a mature understanding of neural networks and their possibilities. Th is book inoculates us from hype and moral panics, showing us ways to productively engage with these technologies.” — LEONARDO FLORES, author and scholar of electronic literature, member of the MLA-CCCC Task Force on AI and Writing “Th e book provides a geographically, disciplinarily, and conceptually broad framework to navigate today’s heated debate about the impact of AI on our civilization. It introduces a number of terms and theories from fi elds such as literary studies, media studies, and the history of technology.” — JANA HORÁKOVÁ, author of Robot jako robot “Th is ground-breaking publication off ers an extensive insight into Central European AI-literature and −art and their cultural implications. A great start for an introduction to the topic!” — JÖRG PIRINGER, author of günstige intelligenz and datenpoesie INSTITUTE OF CZECH LITERATURE KAROLINUM PRESS CZECH LITERATURE STUDIES KAREL PIORECKÝ– ZUZANA HUSÁROVÁ Th e culture of neural networks Synthetic literature and art in (not only) theCzech and Slovak context Th e culture of neural networks KAREL PIORECKÝ– ZUZANA HUSÁROVÁ
The culture of neural networks Synthetic literature and art in (not only) the Czech and Slovak context Karel Piorecký – Zuzana Husárová INSTITUTE OF CZECH LITERATURE is part of the Czech Academy of Sciences Na Florenci 1420/3, 110 00 Prague 1, Czech Republic www.ucl.cas.cz KAROLINUM PRESS is a publishing department of Charles University Ovocný trh 560/5, 116 36 Prague 1, Czech Republic www.karolinum.cz Translation © Isabel Stainsby, 2024 Text © Karel Piorecký, Zuzana Husárová, 2024. The percentual division of the textual content is Karel Piorecký 50% and Zuzana Husárová 50%. Photo © Authors’archive, 2024 © Institute of Czech Literature of the Czech Academy of Sciences, 2024 © Karolinum Press, 2024 Cover and graphic design Designiq Set by Karolinum Press First English edition The publication was created as part of the research project of the Institute of Czech Literature of the Czech Academy of Sciences, v. v. i. (RVO: 68378068). The work utilized bibliographic resources from the research infrastructure Czech Literary Bibliography – https://clb.ucl.cas.cz/ (code ORJ: 90243). This book was published with support from the Czech Academy of Sciences. The text was created with the support of the Academic Premium awarded by the Czech Academy of Sciences to Prof. PhDr. Pavel Janoušek, DSc., and also with the support of the Zdenek Pešat Scholarship of the Institute of Czech Literature of the CAS, awarded to Zuzana Husárová in 2021 and 2022. The book is an output of the grant project VEGA 2/0163/22 “Literature in Bioethics and Bioethics in Literature.” Principal investigator: Mgr. Bogumiła Suwara, PhD. Project duration: 2022–2025. This work is licensed under a Creative Commons Attribution 4.0 International License (CC BY 4.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited. ISBN 978-80-246-5784-4 (pdf, Karolinum) ISBN 978-80-246-5783-7 (Karolinum) ISBN 978-80-7658-100-5 (pdf, Institute of Czech Literature, CAS) ISBN 978-80-7658-099-2 (Institute of Czech Literature, CAS) https://doi.org/10.14712/9788024657844 The original manuscript was reviewed by András Cséfalvay (Digital Arts, Academy of Fine Arts in Bratislava), Jana Horáková (Theory of Interactive Media, Masaryk University, Brno), Josef Šlerka (Institute of Information Studies and Librarianship, Charles University, Prague)
Charles University Karolinum Press www.karolinum.cz [email protected]
Contents 1. The dusk of artificial intelligence 7 2. The technological imagination as asource of the culture of neural networks 19 3. Generative literature and its history 41 4. Artificial neural networks and their functioning principles 53 5. Generating poetic texts 63 6. Prose texts and narrative assistants 91 7. Deploying neural networks in drama and theatre praxis 111 8. Synthetic visual art 133 9. Intermedia and musical synthetic works 153 10. Presentation strategies for synthetic textual media 169 11. Reception mechanisms for synthetic textual media 177 12. The consequences of generative praxis for the theory of creativity 187 13. The vernacularization of synthetic creation 211 14. On the myth of artificial intelligence 219 15. The dawn of the culture of neural networks 229 References 233 List of Illustrations 246 Editorial Note 248 Name Index 249
Theduskofartificialintelligence 7 1 Theduskofartificial intelligence This book is not about artificial intelligence. Rather, it is about what is labelled artificial intelligence in public discourse. That is, it is about the shorthand that makes artificial neural network technology an achievable objective of artificial intelligence, though this still remains afuturistic fantasy. However, this book is not about technology either; rather, it is about what artificial neural network technology will achieve in the world of art and literature– and indeed in the inner worlds of the human imagination. This book is about culture, which we understand as aset of “performative language-games” (Barker 2004:45) circulating without clear rules or boundaries around apoint that only recently ceased to be merely an experimental technology and has become adriver of cultural processes and interactions. This point, which on this occasion constitutes the focus of our professional interest, is the action of neural nets (NNs), as we are already used to calling the type of software that has rapidly become part of our daily lives and has thereby earned this familiar label. By “culture of neural networks” (CoNN), aterm that has also infiltrated the title of this book, we mean the processes and interactions in the network of actors formed particularly by the technologies of artificial neural networks themselves; the corporations developing them; the software users; the products (works of art) generated; their recipients– the media disseminating these products; and last but not least the metatexts and paratexts accompanying them. The culture of neural networks is very buoyant and is evolving dynamically; it functions as anew culture, albeit one intrinsically linked to cultural tradition. For this reason, one of the questions we ask ourselves is
8 Theduskofartificialintelligence directed at the relationship between old and new; at the relationship between cultural tradition and how it is evaluated during the generative processes activated by neural networks; at the relationship between new technology and older methods of producing artistic artefacts; at the differences between existing ways of perceiving artworks and the anticipated reception of NN-generated artworks. Are we on the brink of anew phase in the digital revolution, as Anthony Elliott indicates in the subtitle of his book The Culture of AI. Everyday Life and the Digital Revolution (2019)? Or is the time denoted by these grand words yet to come? It is certain, and undeniable, that the culture of neural networks (or rather the language games comprising them) has great demythologizing potential (e.g., in relation to the aura of artworks and their creators, or to the general idea of what creativity is), but may also still create amythology, possibly even by mythologizing itself. One of the myths obscuring this entire field is the concept of artificial intelligence, or to be more precise, the uncritical, unreflective use of artificial intelligence. We definitely would not define the age in which we live as the age of the “rise of artificial intelligence”. In acertain sense of the word, we find ourselves in the opposite situation: the key term “artificial intelligence” is starting to decline, simply because the media falsify it and use it as aname for software systems that, although they complete their tasks ever more perfectly, still have nothing in common with intelligence, if only because they are unaware of the said tasks, they are set up to complete asingle type of activity, they depend on sets of training data, they “hallucinate” unpredictably, and so on. Strictly speaking, the metaphor epitomizing our time is not the dawn of artificial intelligence, but its dusk. As aresult, you will hear nothing about the “culture of artificial intelligence” from us. On the contrary, we will attempt to avoid the term on the following pages– unless, of course, we intend targeted criticism. Naturally, we do not wish to delegitimize in this way the concept that has become the core of Elliott’s approach in the above-mentioned monograph. The issue is rather the fundamental
Theduskofartificialintelligence 9 methodological differences between our work and his, the difference in our initial assumptions and still more in the different objectives we are moving towards. Similarly, our approach and goals differ from other authors whom we might label as technooptimists (or technoutopians), whose ideas we do not plan to debate– precisely because they operate in areas beyond the scope of our expertise in literary and art theory, such as Mustafa Suleyman in his book The Coming Wave, who considers the future of AI primarily with regard to its application in biology (and the subsequent impacts on the quality and length of human life). We also do not wish to venture into the realm of forecasting the future and the time remaining until AI reaches the state of singularity– like Ray Kurzweil does, although some of his ideas are very close to us (e.g., the assumption that human and machine creativity will reach astage of mutual indistinguishability). We mention Elliott’sbook here as an example of ascientific approach (not visionary essayism) to which we want to offer an alternative in the cultural field of art and literature. In our work, we build upon the positions of authors who deal with the topic of AI in the field of computer and cognitive science (such as ErikJ.Larson), in the context of digital literature (such as David Jhave Johnston and Hannes Bajohr), in the context of art studies (such as Lev Manovich and Emanuele Arielli), and others. Anthony Elliott defines the culture of artificial intelligence thus: “the general social process by which everyday life and modern institutions become increasingly influenced and shaped by the digitalized and technical apparatuses of AI” (Elliott 2019: 51). In doing so, he emphasizes that understanding these processes is “crucial for understanding the world today, aworld which is increasingly overlaid by digital networks of communication, AI technical systems, institutionalized automation and advanced robotics” (ibid.). In reaching this understanding, he focuses on the transformation of social relationships instigated by digital technologies and constructs his basic argument from this sociological position: “Acentral argument of this book is that the robotics revolution and AI impact upon not only
16 Theduskofartificialintelligence of the main systemic impacts brought to the world of art and literature by artificial neural networks is the further democratization, or more concretely, vernacularization of these creative fields; this is the main theme of the following chapter, The vernacularization of synthetic creation, which however also distinguishes three phases in the process of establishing neural networks in the culture: 1. the verification phase; 2.the artistic-subversive phase; 3.the vernacular phase. The idea of mythologization pervades our entire book, and likewise, principally, the culture of neural networks. This mythologization is, we feel, the main barrier to acritical and realistic view of the risks and benefits associated with so-called artificial intelligence. The chapter On the myth of artificial intelligence summarizes our reflections on this topic; together with the concluding chapter, The dawn of the culture of neural networks, it is an appeal to arational, reflected approach, which is sensitive to ethical questions, to the cultural transformations in which artificial neural networks are involved and which we are only now learning to understand. We are convinced that the path to understanding these processes should start where such transformations are birthed. Phil Turner, in his book Imagination + Technology (2020) assumes that all digital products are products of our imagination (see Turner 2020: 122). Within this technological imagination, he further distinguishes between “Imagining the Possible,” and “Imagining the Improbable”. “Imagining the Possible” refers to the imagination integral to the development of specific technologies, as well as anatural part of using them (often metaphorical in nature: this technology is something like...). We can use this form of technological imagination to “translate” digital technologies into an analogue language, making them easier to understand. (After all, even the use of artificial intelligence technologies is based on this interaction metaphor: software that creates sentences or images like humans “must” be somewhat like ahuman...). However, our focus in this study is on the second type of technological imagination, namely the imagination of the improbable, which is primarily expressed through what Turner calls “design
Theduskofartificialintelligence 17 fictions”– texts, films, and other artefacts that, through their thematic focus and form, may foster interest in the future of technologies. The purpose of design fictions is not to show how things will look in the future but to open up space for discussion. Design fictions are based on provocation, asking questions, and exploring possible innovations (Turner 2020: 125). Turner also emphasizes that discussions about the future of technologies should stem more from the social and cultural sphere than from the technical realm. This perspective is hard to disagree with– designing the future is not aspontaneous shift in technology, but is based on social and cultural realities or needs. These desires and needs leave their mark on history in the form of artefacts, because art is amedium sensitive enough to capture and reify this kind of imagination. In this book, then, we will analyse the traces of this imagination in the texts constituting fertile ground for the growth of the culture of neural networks. We also fully identify with the opinion of Marc Coeckelbergh, who states that the fears and hopes associated with artificial intelligence “have clear links to fictional narratives in human culture and history” and therefore any research on them must be incorporated into research on AI, so that we can better understand “why certain narratives prevail, who is creating them and who benefits from them” (Coeckelbergh 2020: 16–17). But first, let us take alook at the inception of the theoretical discussion on artificial intelligence, which transformed the previously mentioned imaginative line into aform of scientific discourse.
Thetechnologicalimagination 19 2 Thetechnological imaginationas asourceoftheculture ofneuralnetworks In the mid-20th century, pioneers in computer science, notably Alan Turing and John von Neumann, discussed the analogies between the human brain and computers, suggesting that human intelligence (mainly reduced to following concrete tasks, in their understanding) could be replicated by computers. The term artificial intelligence can be traced back to the legendary two-month summer seminar at Dartmouth College in 1956, organized by John McCarthy, Claude E.Shannon, Marvin L. Minsky, and Nathaniel Rochester, attended by ten young leading computer scientists; the term was coined by John McCarthy to distinguish it from the field of cybernetics. As he stated in his book Defending AI Research, the focus of cybernetics on “analogue feedback seemed misguided” (McCarthy 1996: 73). The idea behind the concept of AI as aresearch discipline was the “conjecture that every aspect of learning or any other feature of intelligence can be in principle so precisely described that amachine can be made to simulate it” (Dick 2019). As McCarthy later admitted, nobody liked this name because the goal was genuine intelligence, not artificial intelligence (see Mitchell 2019: 18). In the field of artificial intelligence, we can distinguish between two paradigms: the symbolic and the subsymbolic (or connectionist) (see Mitchell 2019: 21). While the symbolic paradigm was inspired
20 Thetechnologicalimagination by mathematical logic and conscious thought processes, and can be considered transparent because it follows rules and processes set by humans, the subsymbolic paradigm lacks such transparency– the term “black box” is ajustified poetic descriptor here– it learns from prepared data and performs certain tasks based on that data. The symbolic paradigm (especially in the form of expert systems) defined the first thirty years of AI research after the Dartmouth workshop. Its proponents no longer claimed that AI could be created by copying human thought processes, but argued that general intelligence could emerge through the right symbol-processing programs. The subsymbolic paradigm drew inspiration from neuroscience and from attempts to capture even unconscious thought processes (fast perception), such as facial recognition or speech identification. At its core, its approach to symbol-processing emphasizes neural architectures that provide afoundation for learning character recognition. Although it flourished only with the rise of deep learning, an early example of this paradigm was the perceptron program developed by psychologist Frank Rosenblatt in the late 1950s, inspired by neural information processing. He even proposed that perceptron networks could be capable of recognizing faces or objects and designed the perceptron-learning algorithm (see Mitchell 2019: 24–26). However, the field of artificial intelligence did not see much promise in the subsymbolic paradigm (Minsky and Papert even labelled the multilayered composition of perceptrons as a“sterile” path in their book Perceptrons, published in 1969), and for along period, it promoted the symbolic paradigm, which became the foundation for establishing AI research centres at American universities in the 1960s and ameans of obtaining government funding. The predictions of artificial intelligence pioneers from the 1950s and 1960s, who linked their proposals to the advancements in computer science during the 1950s, did not come true as expected, and research in machine translation stagnated. AI research went through periods known as AI Springs and AI Winters, during which the initial ecstatic enthusiasm and high hopes for the emergence of artificial
Thetechnologicalimagination 21 intelligence turned into resignation due to the inability to meet expectations and predictions, leading to financial cutbacks from government institutions. Asignificant shift in artificial intelligence research occurred around 2006, when multilayer neural networks (an extension of Rosenblatt’s perceptrons) yielded remarkable results. Since that time, there has been talk of another AI Spring, associated with the advent of deep learning and machine learning. THE CONNECTIONIST PARADIGM IN THE CULTURE OF NEURAL NETWORKS The functionality of neural networks is based on recognizing sequences in data and attempting to replicate or mimic these sequences. Neural networks do not operate at the level of alphabetical characters, musical notation, or visual representations; instead, they recognize these symbols via numerical relationships. As early as 1999, N.Katherine Hayles, in the prologue to her book How we became posthuman: virtual bodies in cybernetics drew attention to the technological principle in which “the erasure of embodiment is performed so that ‘intelligence’ becomes aproperty of the formal manipulation of symbols rather than enaction in the human lifeworld” (1999: xi). The Turing test described the disappearance of the body from the definition of ahuman being as amagic trick, which allowed the “formal generation and manipulation of informational patterns” (ibid.) to stand as asufficient definition for both human and machine intelligence. It is also important to note that artificial neural networks are extremely abstract versions of brain neural networks. Artificial neural networks transmit simple numerical signals, whereas biological ones transmit aseries of pulses. Brains operate based on parallelism, while artificial networks, though significantly faster than their biological counterparts, can only perform computations serially, making them less efficient overall. Another essential aspect is the principle of fragility: since each artificial neuron acts as an independent processor,
22 Thetechnologicalimagination an error in one significantly affects the functionality of the entire model1. Our brains are accustomed to neuron death and can adapt to new circumstances. If we shift from biological terminology to the technological aspect introduced at the beginning of this study, specifically to the symbolic and subsymbolic (connectionist) paradigms, we can find the connectionist paradigm related to neural networks in Hannes Bajohr’s study specifically dedicated to the literature of neural networks. Bajohr (2022) distinguishes between “sequential” and “connectionist” paradigms in digital literature. The “sequential” paradigm pertains to linear algorithms, which means digital literature created using readable code, while the “connectionist” paradigm concerns digital literature generated using neural networks. The essential nature of the generated literary works lies in the code created by the authors or technical collaborators, accessible to recipients for reading and even critical reflection through tools of “critical code studies” (as advocated by Marc Marino, see, for example, Marino 2020). This involves perceiving the artwork not only through interface presentation but also through mutual interaction with its code background. The connectionist paradigm draws on the terminological discourse of AI and, in contrast to the sequential paradigm, emphasizes the nature of neural networks as a“black box,” implying the inability to see beyond what is presented. Unlike explicit programming, this case involves implicit learning: “There is no code to inspect in this case; instead, there is only alist of numbers representing the structure of the network and its weighted connections, but such alist is extremely difficult to interpret.”2 (Bajohr 2021b: 483). 1 Further future research into ‘sparse networks’ might reveal that many connections are redundant and can be removed without degrading performance. 2 Es gibt dabei keinen Code, der zu inspizieren wäre, sondern nur eine Liste von Zahlen, die die Struktur des Netzes und ihre gewichteten Verbindungen darstellen; einesolche Liste ist jedoch ausgesprochen schwer zu interpretieren.
Thetechnologicalimagination 23 The premise of the culture of neural networks is adiscourse in which various degrees of the technological revolution’s practical achievements intersect with specific creative practices in the fields of digital art, literature, music, and various intermedia and transmedia projects. Simultaneously, it is influenced to an equal extent by human imagination in the form of literary texts and artworks, as well as technological visions within the scientific community. Artistic imagination has on many occasions predetermined technological directions, as evidenced by numerous examples from science fiction and art, with technicians drawing inspiration from it in their practice. The archaeology of (synthetic) media can help us find answers to questions about how unique contemporary synthetic artistic practice is and how it reinforces old likenesses between the human and the divine. THE BEGINNINGS OF ARTISTIC IMAGINATION OF (TECHNOLOGICAL) PROGRESS There, intent, Pygmalion stood before an altar, when his offering had been made; and although he feared the result, he prayed: “If it is true, OGods, that you can give all things, Ipray to have as my wife—” but, he did not dare to add “my ivory statue-maid,” and said, “One like my ivory—.” Golden Venus heard, for she was present at her festival, and she knew clearly what the prayer had meant. She gave asign that her Divinity favoured his plea: three times the flame leaped high and brightly in the air. (Ovid 1922)
24 Thetechnologicalimagination Pygmalion’s desire for apure and beautiful being, in contrast to the “shame” and “faults” of women in his city, led him to pray to Venus to bring his own artistic creation to life. Galatea, astatue with aperfect female body, carved from ivory by the sculptor Pygmalion, becomes human, marries him and gives birth to adaughter named Pafos. This story has various versions: according to one of them, Pygmalion was aCypriot king who fell in love with astatue of the goddess Aphrodite, while according to others, Aphrodite came to see Galatea and was delighted that she was sculpted to copy the goddess’s body. According to the Greek version, the goddess was Aphrodite, while her Roman counterpart was Venus. Although the circumstances may vary somewhat, the foundation remains the same: aman’s unsatisfied desire for a perfect woman results in a creative appeal to a deity. Such astereotypical approach has been confirmed in other later imaginations, which we will explore further, or has become the target of artistic satire, as evidenced by Isaac Asimov’s feminist science fiction short story Galatea from 1987, in which the female protagonist, Elderberry, is the scientific experimenter. The male statue Hank, brought to life by Galatea’s uncle George using thepower of the imp Azazel, fails to meet her expectations because the trait that was given to her as defining, namely softness, also describes the statue’s male genitals: “When Isaid Iwanted Hank soft, Ididn’t mean soft all over, permanently” (Asimov 1987). The earliest pioneers and inventors from 9th-century Baghdad were three Persian brothers with the family name Banū Mūsā ibn Shākir: Muhammad, Ahmad and al-Hasan. The sons of the famous astronomer and astrologist Mūsā ibn Shākir were scholars and devoted their lives to geometry, astronomy, mechanics and music, and were key for the translation of ancient Greek manuscripts. They wrote the manuscripts Kitāb al-Ḥiyāl (The Book of Ingenious Devices, approx.850) and Kitāb al-urghanun (The Book of the Organ, approx.850), in which they described around 100 innovations, such as remote control and an automatic handle, valves, automatic fountains, water dispensers, various lamps and ahydro-powered organ. Their extraordinary
Thetechnologicalimagination 25 treatise al-Āla allatī tuzammir bi-nafsihā (The Instrument that Plays by Itself) describes aplan and design for the first programmable machine, amechanical flautist. The tool is amechanical hydraulic organ with air to drive anine-hole flute. “The holes are opened and closed by eight levers, the end of which make contact with the fixed raised pins arranged on the lateral surface of arevolving cylinder so as to produce awell-known melody” (Sanjakdar Chaarani 2021). The flute produces amelody in line with the melody programmed on its rotating cylinder. As the Banū Mūsā brothers proposed in their manuscript: “If we want to create the humanoid flautist, we simply have to incorporate the whole device in the body of the statue, fix the flute in its mouth and disguise the levers as fingers and adapt it to his arms” (Sanjakdar Chaarani 2021); in this way, they designed the first automatic musical humanoid. Another Muslim inventor, Ismail al-Jazari (1136–1206), originally from Jazira, whose visual designs and projects made their way from the Near East to Europe, was apolymath, engineer, artist, mathematician, astronomer, designer and inventor. Al-Jazari, often referred toasthe father of robotics, described programmable humanoid automatons in his publication al-Jāmiʿ bain al-ʿilm wa al-ʿamal al-nāfiʿ fī ṣināʿat al-ḥiyal (The Book of Knowledge of Ingenious Mechanical Devices) of 1206. He enriched his text with specific illustrations, assembly instructions, and design methods. His automatons included arobotic girl serving drinks, afountain with apeacock, elephant clocks, automatic gates, various automatic machines, and musical automatons. Automatons like the robotic girl serving tea or water, wooden figurines in aboat playing musical instruments, water clocks with drummers, and other automatons that he invented can be considered as remarkable examples of apractical use of humanoids. It is said that even da Vinci was inspired by al-Jazari’s approach when creating his own automaton in 1495, which took the form of ametal knight. The 1565 automaton, aFranciscan monk made of wood and iron, who walks and kisses the rosary, was probably constructed by the Italian-Spanish clockmaker, engineer, and mathematician Juanelo
32 Thetechnologicalimagination A film that significantly advanced expressionist aesthetics on screen while connecting technological revolution with asocialist vs. capitalist charge was Fritz Lang’s Metropolis (1927). The division of the world into two opposing poles– the heavenly skyscrapers with the Tower of Babel, home to the city’s chief architect and wealthy people, contrasted with the gloomy, dirty underworld, where the poor labour to keep the machines running– provides the backdrop for the stereotypical spark of love between Freder, the mayor’s son, and Maria, apoor worker and revolutionary. The mad scientist Rotwang creates ahumanoid replica of Maria called Futura to prevent the “upper” and “lower” classes from mixing, thus maintaining the status quo. The story slightly resonates with the narrative of L’Éve Future, except that Freder was supposed to be the victim of deception resulting from the collaboration between his father and the scientist, rather than awilling participant in his own assignment. The metaphor of human as machine also found its way into theatre in the 1920s, but in acontext different from Čapek’s. This was specifically evident in the work of Vsevolod Meyerhold and his theatrical biomechanics, which consisted of 16 exercises that actors had to master to control their bodies like machines. The director played the role of the constructor of commands that the actor’s body had to execute. Meyerhold developed this principle in his Moscow production of The Magnanimous Cuckold (1922), where the entire ensemble of actors was perceived as acollective machine on stage, constructed from ladders, stairs, platforms, wheels, and beams. The comparison of actor to machine was also utilized by the German choreographer, designer, sculptor, and painter Oskar Schlemmer. In his Bauhaus productions Triadic Ballet (1922) and Figurine Dance (1926–1927), Schlemmer used costume to transform the actor’s body into amechanical figure. Russian theatre artist Nikolai Foregger is credited with the dance technique known as “tafiatrenage” (choreography presented as Mechanical Dances in 1923). Unlike Meyerhold’s biomechanics, Foregger’s approach was not just atraining method; it was also an artistic form that directly represented technological progress, mechanization
Thetechnologicalimagination 33 in production, and the automatism of operations. “Foregger’s mechanical dance represents apure mechanistic artistic form focused on the machine-like qualities evoked by the movements of the dancers” (Horáková 2010: 29; for more information on the portrayal of machines in 1920s theatre, see Horáková 2010). LITERARY EXAMPLES OF METAMORPHOSES BETWEEN THE HUMAN AND THE TECHNOLOGICAL IN THE 20TH CENTURY Iask myself, to no purpose, what is likely to happen to him? Can he possibly die? Anything that dies has had some kind of aim in life, some kind of activity, which has worn out; but that does not apply to Odradek. Am Ito suppose, then, that he will always be rolling down the stairs, with ends of thread trailing after him, right before the feet of my children, and my children’s children? He does no harm to anyone that one can see; but the idea that he is likely to survive me Ifind almost painful. (Kafka 1971)5 Odradek, from Kafka’s short story The Cares of aFamily Man, narrated by ahomodiegetic narrator, the head of the family, is abeing whose ontological status has been the subject of much contemplation. As the narrator perceives it, Odradek takes the shape of apurposeless mechanical star with threads, yet it laughs with alaughter that can be produced even without lungs, and its dwelling is ever-changing. The narrator suggests that Odradek may once have been part of acomplex form, so its purposelessness is aresult of this loss, but he admits that he cannot say anything more about it because it “is extraordinarily nimble and can never be laid hold of” (Kafka 1971). Odradek, as an automaton that at other points in the story appears as wooden when silent, stretches our interpretative possibilities with 5 Translated by Willa and Edwin Muir. https://mediationsjournal.org/articles /worries-of-a-family-man Accessed 19.02.2024
34 Thetechnologicalimagination Kafka’s imagination and the narrative gaps: it is amechanical being that appears and disappears, it is amachine and at the same time perhaps acomponent of another machine, it answers some questions but mostly remains silent. Similar themes were also explored by two other cult writers in their short stories, namely Herman Melville in The Bell-Tower (1855), where the story focuses on the posthuman character of the bell-tower machine, whose basis was the architect’s blood, and Edgar Allan Poe in The Man That Was Used Up (1839), about the prostheses of ageneral’s body6. However, what remains in any interpretation of Odradek are human reflections on our own mortality and on the machine that will outlive humans (even literally our children’s children), and indeed the feelings of recognizing the mechanical as purposeful and the human as an existence that was not created for aspecific purpose. Odradek, which is purposeless, with its possible purposeful past unknown to us, is closer to humanoid notions due to its naivety and childlike perspective than other machines in artistic imaginations because it has its own attitude, own intention, and was not created with the idea that its final form would have acurrent function. The sorrow of human mortality contrasted with mechanical permanence, expressed in the last paragraph of the story (and quoted at the beginning of the discussion of this story), leads us to the concept of “Promethean shame,” as articulated by the German philosopher Günter Anders. “Promethean shame,” aconcept introduced by Anders in his philosophical book Die Antiquiertheit des Menschen (1956, The Obsolescence of Man), stands in sharp contrast to artistic notions which, since the Pygmalion myth, place non-human entities made of clay, metal, chemistry, and components in asubordinate position to humans, who create these beings either for their own pleasure and delight or for protection. Anders’ thesis, on the other hand, portrays humanity as incapable of competing with machines, and describes an anthropological crisis caused by technological development in the 6 For further exploration of other posthumanist works, see I. Lacko‘s book ABeautiful New Post-World (Lacko 2021: 28–30).
Thetechnologicalimagination 35 second half of the 20th century, in which machines appear to be more efficient and complete than modern humans. Referring to feelings of powerlessness and emotional exhaustion in the face of artificial forces that have no doubts and never malfunction (especially after the experience of world wars and atomic explosions) and the idea that humans will become obsolete compared to their technological “descendants,” Anders proposes “anew conception of human finitude based around our inability to see or comprehend the artificial powers we blindly place our hope in.” (Müller 2016: 11). Anders argues that he has recognized anew human feeling: “Believe Ihave found the signs of an entirely new pudendum this morning; aform of shame that did not exist in the past. Iwill provisionally call it ‘Promethean shame’ for myself. Iunderstand this to mean the ‘shame when confronted by the “humiliatingly” high quality of fabricated things (selbstgemachten Dinge).’” (Anders 1956: 23). Anders observed this shame when visiting atechnical museum with his friend T. and described it as the difference between the physical clumsiness and imprecision of humans compared to the perfection of machines. He contrasts Promethean shame with the typical self-made man of the 19th century, who viewed everything, including himself, as apersonal achievement. Anders perceives the mirror that scientific and technological progress presents to human beings as apsychological stance towards the value, self-confidence, and self-love of humanity. He calls his approach “the philosophy of discrepancy,” which involves analysing the differences between what we are capable of producing and our conception of it. CYBORGS AS MYTHOLOGICAL PERSONIFICATIONS OF RESPONSES TO HISTORICAL CHALLENGES Examples of artistic imagination illustrate, on the one hand, the fear of the unknown, Freudian uncanny feelings that are associated with
36 Thetechnologicalimagination automatic beings because they are not “ours”. On the other hand, they inject an element of romantic adventure that is evoked and provoked precisely by the “unknown”. However, imagination materialized in narratives has predetermined philosophical contemplation and mechanical constructs. Asimilar tendency can be traced in algorithmic thinking and its implementations, from its beginnings in the 9thcentury to realizations in the 1950s. Since the 1960s, the sci-fi boom has moved from the realm of geeks to pop culture: robots, cyborgs, androids, and various forms of artificial intelligence presented from Star Wars to Asimov’s stories and Dick’s novels, such as in Do Androids Dream of Electric Sheep?, as well as in cult roles like Johnny5 or Terminator, have filled the cultural space with the notion that AI is close, alongside the fear that it will manipulate humanity. However, Viennese cyberneticist and literary experimenter Oswald Wiener introduced the idea of abio-adapter resembling some kind of shell or spacesuit that would save Central Europe and humanity as awhole. He described it as follows: “viewed from <outside>, the adapter places itself between the unsatisfying cosmos and the unsatisfied human being. it hermetically seals off the latter from the traditional environment and in the first stage of adaptation only falls back on its own information, which it has stored for this purpose, or on that which the human being contains.” (Wiener 1965–1966: 6).7 In “appendixA” of his experimental work die verbesserung von mitteleuropa, Oswald Wiener outlines his bio-adapter as ameans of connecting the human organism and acybernetic device. Given the transhumanist approach, this connection can be termed acyborg (aportmanteau of “cybernetic” and “organism”), with the bio-adapter also playing arole in preserving human consciousness after death. The main function of this special interface is to adapt the human to the constantly 7 der adapter legt sich– von <<aussen>> betrachtet– zwischen dem ungenügenden kosmos und den unbefriedigten menschen. er schliesst diesen hermetisch von der herkömmlichen umwelt ab und greift nur in den ersten stadien der adaptation auf zu diesem zweck gespeicherte eigene informationen und auf solche seines inhalts zurück.
Thetechnologicalimagination 37 changing external environment. The bio-adapter is supposed to provide an extension of consciousness and the senses, to correct any health complications, and to enhance humans overall. This description of the gradual merging of humans and acybernetic interface resonated in Austrian literature in 1969, and it was reintroduced to an English-speaking audience precisely half acentury later by Beate Geissler in the book Oswald Wiener: The Bio-Adapter (2019) with aforeword by the renowned German media theorist Siegfried Zielinski. As we have shown through examples of human imagination dating back to antiquity, where men dressed in Prometheus’s skin to satisfy their romantic or protective needs, often ending tragically, especially for those who desired or constructed robots, humanity has been artistically projecting its own technological dystopia for along time. Various forms of science fiction stories since the 1960s have built upon myths, romantic sci-fi novels, and expressionist theatre or film productions. From the 1950s onwards, we can also speak of robotic involvement in artistic endeavours, specifically the first cybernetic sculpture, CYSP1, created in 1956 by Nicolas Schöffer, originally from Hungary. UNFULFILLED VISIONS OF INTELLIGENT MACHINES From as early as the late 1920s, the technological visions of computer scientists represent adeviation from the theme of creating an artificial robot-person (that is, amachine body) towards the development of intelligent systems, in which their resemblance to humans is not measured by their anthropomorphic appearance but rather their “intelligent” behaviour. Two of the pioneers of artificial intelligence, the computer scientists Herbert Simon and Allen Newell, stated in their lecture of 1957 that “there are now in the world machines that think, that learn and that create” (Simon– Newell 1958: 8). The lecture was published in the journal Operations Research and in it, the authors
38 Thetechnologicalimagination predict ahuge shift in computer technology over the next ten years: adigital computer is to defeat ahuman being in the World Chess Championships, discover and prove anew mathematical theorem, write aesthetically high-quality music and theories in psychology are to resemble computer programs or qualitative statements (Simon– Newell 1958: 7–8). They contextualize their proposals with the results and speed of research by artificial intelligence in the 1950s, especially with research into understanding natural language. None of these assumptions came true in the 1960s, however, and research into machine translation stagnated, although the American government supported this type of research at several American universities to the tune of millions of dollars. Jerry Fodor described the disappointment of the research teams thus: “[they have] walked into agame of 3-dimensional chess, thinking it was atic-tactoe” (Dreyfus– Dreyfus 1988: 21). The significant shift in AI research was brought about by noteworthy results in neural networks resembling deep learning programs in around 2006. This shift consists of the development of hardware (especially graphics cards and specific processors) that enabled abroader use of deep learning. Academic research found applications in commercial companies, whose main business was big data, or in imageor text-recognition companies. Machine learning, the term used to describe the functioning of neural networks, denotes “computational treatment of induction– acquiring knowledge from experience” (Larson 2021: 133). Machine learning is an application of artificial intelligence that enables systems to learn and improve based on experience, without being explicitly programmed to do so. This term more accurately describes neural network learning processes than the general term ‘artificial intelligence’, whose imaginary breakthrough was supposed to be the Turing test, in which Turing reduced intelligence itself to problem-solving. Instead of actually measuring the intelligence (or description of intelligence) of machines, however, the Turing test leaves ahuman jury to decide who actually wins in the human “imitation game” during the test. For the program to succeed, it needs to convince athird of the
Thetechnologicalimagination 39 jury that it is talking to ahuman being, not amachine, in afive-minute text conversation. Eugene Goostman’s successful chatbot model of 2014 simulated athirteen-year-old Ukrainian boy who spoke English, and proved that it is the concept of mimicry that is productive and convincing rather than the intelligence of the machine itself. Neural networks function on the basis of recognizing sequences in the data and attempting to replicate or imitate these sequences. Neural networks do not operate at the level of alphabetical characters or visual representations; for them, these symbols are always represented by numerical relationships. In our book, we will use the general term synthetic media to designate works created using neural networks; that is, media created in the generation process by computer algorithms (today neural networks are used for this generation process), where the results are hard to distinguish from human creations. Although this term is most commonly associated with visual media (culture) such as deepfakes or creating photos of non-existent people, which frequently constitutes deliberately manipulating the public, musical and literary works created using neural networks (in terms of their formal aspects) can also be designated synthetic media. For this reason, we also use the term synthetic media in our book and, in the following chapters, we use different variants thereof, where the term synthetic textual medium (and any derivatives such as synthetic art, synthetic literature, synthetic poetry, etc.) may be regarded as fundamental. Because we agree with Lev Manovich’s statement that “we must remember that these methods are neither the first nor the last in the long history and future of simulating human art abilities or assisting humans in media creation” (2023), we also introduce in the next chapter the principal works in the history of generative literature that substantially influenced the modern context of synthetic textual media.
Generativeliteratureanditshistory 41 3 Generativeliterature anditshistory Literature written by neural networks can be considered arecent example of literature produced and presented by technology, which is broadly captured by the encompassing term “electronic literature”. Theoretical considerations of electronic literature offer several ways of contextualizing its historical background: either it is seen as acontinuation of experimental tendencies in (print) literature, or its technical and multi-modal nature is emphasized and it is viewed as adistinct genre of digital (media) art with its own history (as Chris Funkhouser writes: “Poetry is poetry and computer poetry– although related to poetry– is digital poetry”; Funkhouser 2007: 80). Some theorists have approached this issue by creating their own terminology: the Norwegian theorist Espen Aarseth coined the term “cybertext” to include digital games and various media projects in addition to computer literature, thus drawing attention to the connections between these digital projects. Aarseth stressed the complexity of decision-making within the reception process compared to the traditional literary text. He coined the term “ergodic literature” for the historical background of cybertexts and used it to refer to literature in which, when it is read, both the reader’s/user’s possible choice of approach to the texts is emphasized, as are the increased demands on the reader/user. Aarseth used this term to refer to an array of different texts, ranging from inscriptions spread out over multiple walls of Egyptian temples, the IChing (or Book of Changes), calligrams, and unbound literature, e.g. Marc Saporta’s Composition no.1, Roman, (1962; Eng. trans. 1963), B. S. Johnson’s The Unfortunates (1969), and Raymond Queneau’s Cent mille milliards de poemes (1961;
48 Generativeliteratureanditshistory parapoetry”), that is to say, texts that are comparable to “natural”/ non-machine poetry, but whose structure bears no trace of any authorial personality or intentionality, which they considered necessary in “real” poetry: The master, who will be futilely anointed, does not curse the lower cloud/ And today he does not complicate / He does not manually understand/ How the father of an African drunken whim creates aromantic desperate melon (Pala– Sus 1967: 42–45; Transl. ChatGPT).11 Oleg Sus revisited the topic of computer-generated texts one more time, and this time as the sole author. In the politically heated atmosphere of 1968, Sus published agenerated text in the same journal, Host do domu. The text was generated using words taken from aspeech by Jiří Hendrych, amember of the Communist Party of Czechoslovakia’s Central Committee and an opponent of the ongoing reform efforts in society and culture. Sus did not hide his neo-dadaist intention of mocking the Communist bigwig’s platitudes, but at the same time presented the parodical nature of the generated texts as ideologically unmarked, the objective result of computer activity merely creating new combinations from the linguistic material entered: Infected Marx stands here out of grief. Bearer achieves democratic criminality. With risk, the quiet limited world humanism bleeds. For us, acitizen flies there. Quiet Marx progressively silent for money. In the battle, a Švejk-style political figure quietly lures assets for victory. (...) 11 Mistr, kterého budou mazat marně, neproklíná spodní obláček / Adnes nekomplikuje/ Ručně netuší. / Jak vytváří romantický zoufalý meloun tatínek jednoho afrického zpitého rozmaru (Pala– Sus 1967: 42–45).
Generativeliteratureanditshistory 49 The alligator will reach the establishment of democratic technique with courage. Under the government serves adegenerated worker to humanity’s history. Limited obscene imperialism just isn’t eating. Dynamically, with courage, it remains silent.12 (Sus 1968: 50–51; Transl. ChatGPT) Slovak computer-generated poetry began to appear only in the 1980s, although as early as 1965, Slovak literary scholar Klement Šimončič published in the journal Slovenské pohľady an article that had originally been delivered as alecture at Columbia University, New York, at the Second Congress of the Czechoslovak Society of Arts and Sciences in 1964. In his paper, “Poetika surrealistov abásnická kompozícia zmatematických strojov” [Surrealist poetics and poetic composition by mathematical machines], the author discussed computer poetry, emphasizing its historical predecessor in the surrealist technique of automatic writing. “Schematically speaking, the surrealists carved the path to electronic poetry compositions by starting to emphasize the meanings of individual words at the expense of the overall meaning of the sentence. In this way the sentence lost its semantic coherence or ‘logic’. Individual words have become semantically independent.” (1965: 30).13 The series called Obrazobasne [Imagepoems] by the visual 12 Infikovaný Marx zžalu stojí tady. / Nositel dosahuje zločinnosti demokratické. / Srizikem krvácí tichý omezený světový humanismus. / Pro nás letí tam občan. / Tichý Marx pokrokově mlčí pro peníze. / Vboji vábí aktiv pro vítězství švejkovský polický tichý funkcionář. / (...) / Aligátor na zřízení dojde techniky demokratické sodvahou. / Za vlády slouží člověku historie degenerovaný dělník. /Omezený obscénní imperialismus právě nejí. / Dynamicky sodvahou mlčí. 13 In the 1960s, Ivan Kupec used his “verse machine KLOMP965”– which was ahat– to create two poems as apoetic response to Klement Šimončič’s study devoted to computer poetry and surrealism. Kupec states: “The genius of the super-electric-machine, discovered in the USA, could even replace ashepherd’s hat in acottage in Liptov: hurrah!”. The media artist Milan Adamčiak also devoted himself to conceptual, visual and phonic poetry evolving from the generative principle, but without acomputer. For more information, see the study “Syntetická
50 Generativeliteratureanditshistory artist Daniel Fischer was generated on aCDC3000 computer with subsequent plotting, which causes the written quotation gradually to “disintegrate” into acubist visual form. This was not an authorial text; Fischer used an appropriation reminiscent of conceptual work with text. Poems that can without hesitation be deemed generative poems were published by Rudolf Legel in 1982 in the article “Experiment sinterakciou človek-počítač pri vytváraní básnického textu” [“An experiment with human-computer interaction to create apoetic text”]. For the first poem, “Analyticka geometria vpriestore mojej hlavy” [“Analytic geometry in the space of my head”], he selected keywords from the terminological field of analytic geometry; the words of the second poem “Laska” (“Love”; diacritics intentionally omitted from both titles in Slovak), come from Dante’s “Horská kanzóna” (originally “Canzone Montanina”) in Viliam Turčány’s translation. The poetic construction of both poems is based on sentences; the language of the first poem is enriched with mathematical symbols and the language of the second poem produces aromantic expressivity. Legel amended the computer-generated text so that noun and adjective declensions worked properly in the poems. He evidenced this by including afew unamended lines, from which it can be seen that the computer did not adapt words from the dictionary, but merely selected them, meaning that nouns appear only in the nominative case, verbs in the infinitive and adjectives in the masculine; the computer did not take sentence syntax into consideration either. The article describes the process of creating apoetic text with acomputer in seven steps, from “1. Enter the dictionary of key words into the computer” to “7. Print the finished text” (Legel 1982: 39–40). Legel states that “the advantage of using amachine in text construction is that the machine can select quickly from large data sets, and the distribution of selection is programmable” (39; see Husárová: 2016). poézia vkontexte slovenského nekonvenčného písania apostliterárnej situácie” [“Synthetic poetry in the context of unconventional Slovak writing and the post-literary situation”] by Šrank – Hostová – Novotný (2022: 483–485).
Generativeliteratureanditshistory 51 LOVE WITH ARMOR ON THE CHEST IN AMONSTROUS BATTLE IN BATTLE WITH THE BELOVED IN THE WOODS’ CORNER ABROKEN ARROW ON THE CASTLE COURTYARD PROTRUDING FROM THE CHESTPLATE FELT BY TOUCH IN THE ANGRY BATTLE WITH METAL SWIFTLY REMOVED COMPASSION SOUND OF METAL ON METAL NEAR FLORENCE COMPASSION NUMBED BY DUPLICATIVE HONEYED LOGIC OF BATTLE BELOVED SWEET FEATURES IN THE METALLIC DEPTH OF BLEMISH HEARTS MUTUALLY ARMORED WITH PRECISELY CHEWED METAL EATEN METAL REMOVES SWEET FEATURES IN THE BELOVED CORNER OF THE HEART WHERE ORIGINALLY THEY HID LIKE AFEROCIOUS BEAST14 (Legel 1982: 39–40; Transl. ChatGPT) All of the aforementioned projects were based on the combinatorial principle of computational access to data: ahuman being created adatabase of words that were syntactically ordered and inserted anumber of words into the category of the chosen syntactic member. Thus, the computer always just made aselection from aset of words, but did not change the default syntactic structure. The notion of randomness was, then, relative, because the project’s creator had 14 laska // spancierom na hrudi vobludnom boji / pri boji smilovanou vkute haja / na hradnom nadvori zlamany sip / trciaci vpancieri na hrudi nahmatany / vhnevnom boji skovom odplaveny sucit / zvuk kovu okov pobliz florencie / sucit umrtveny dvojtvarnou medovou logikou boja / milovane sladke crty vkovovej hlbke skazy / srdcia vzajomne obrnene dokladne pozutym kovom / zjedeny kov odstranuje sladke crty vmilovanom kute srdca / kde povodne sa skryvali ako drava selma (Legel 1982: 39–40)
52 Generativeliteratureanditshistory to create the database to ensure that the outputs were grammatical in different combinations. The history of generated creation definitely does not end in that period; however, for the needs of our publication, we have referenced only afew examples at its inception, not the entire historical development. In the 1990s, due to ever more powerful computers and the expanding internet, generative literature became more attractive to creators and more accessible to recipients. Although we find the beginnings of generated creativity, which today is escalating in the form of synthetic media, as early as the mid-twentieth century, it is precisely the differences between these first examples and contemporary generated creations using neural networks that will be analysed in the next chapter. Chapter 4 is the most technical chapter in the book, and we put it in because we think it is necessary to clarify at least the basics of neural network functioning, and what kinds of neural networks are most commonly used to create synthetic texts or synthetic art, in order to understand CoNN.
Artificialneuralnetworksandtheirfunctioningprinciples 53 4 Artificialneural networksandtheir functioningprinciples Many examples of texts generated by neural networks may produce the false impression of autonomous generation, because they are coherent; however, we cannot say that networks understand texts. Aneural network does not know that it is creating atext; rather, for such a program, text remains merely asequence of numbers. Neural networks generate by predicting the next number in asequence of numbers. Converted to characters, this means that they complete the next characters in the sequence. Neural networks do not work on the same principle as the previous examples of generated literature. The project’s author does not supply the program with words that are subsequently rearranged by the machine into the final text, nor do they provide it with the rules of morphology and syntax; the machine “learns” the language structures itself. The basis of neural networks is that the program learns to recognize certain patterns in alarge database and then attempts to replicate or approximate them, thus generating its own data. However, neural networks do not work at the level of language (or of visual representation or music); they function only on the level of numbers. First, all text is converted by the algorithm into an array of numbers in which the network then looks for sequences. Once the network has produced its own sequences, the algorithm converts them back into alphabetic characters. One of the principles of neural network “learning” used to create texts is that the network gradually and repeatedly goes through the
54 Artificialneuralnetworksandtheirfunctioningprinciples entire volume of data and its subsequent output improves with the number of cycles completed. The first outputs are just clusters of letters; later they start formally to resemble words without any meaning, then the network starts to “write” words, and finally the word combinations constitute meaningful lines. However, this “learning” must be stopped at the right moment, otherwise overtraining occurs and the outputs start strikingly to resemble the primary data. Projects generating texts using neural networks may work on the basis of various different principles; they may learn at the level of letters, word stems or whole words. These neural networks abstract the data on which they are trained (so letters or words in the cases of texts), then anumerical identifier is assigned to these abstract units, thus creating a“dictionary” of the units; the neural network then searches for patterns in the number sequences. Where RNNs (recurrent neural networks) are being trained in order to create texts, these units are most commonly letters, because they are the least memory intensive. This method, of course, is also the least accurate linguistically. Training on whole words is harder, because the dictionary so created is much larger and more comprehensive. The optimum method would be to generate texts by character clusters/subwords, because this process partially eliminates the error rate in the results and, at the same time, the algorithm is better guided to “understand” the principle of word formation. Preparing asuitable training corpus is a crucial task in this instance. OpenAI’s GPT-2 (Generative Pretrained Transformer 2) is alanguage model already pretrained on ahuge text database (8 million texts from Wikipedia and Reddit, amounting to about 40 GB), with alibrary consisting of tokens, meaning numerical representations of sub-words. GPT-2 tokenizes via the BPE algorithm, which means that it takes all the unique words present in the database and breaks them down into smaller parts. In this way, the algorithm breaks down the words to obtain tokens, in which it then looks for asequence. That sequence teaches it how aparticular language works, what the sentence structure looks like, and so on. GPT was trained on English,
Artificialneuralnetworksandtheirfunctioningprinciples 55 but it can be fine-tuned to another language; that is, it can be trained to follow the linguistic rules of another language as asuperstructure over English syntax. After training the networks, programmers can control the output with parameters, which determine, for example, the length of the output text (number of tokens), how “experimental” it should be (this is called the temperature), and few others. An initialization text, or keyword to build on (word, paragraph, segment of text) is often entered into the neural network to orient it correctly to the desired output. However, the network can also generate without an initial text. GPT-1 was released in 2018, GPT-2 in 2019 and GPT-3 has existed since 2020. The GPT-3 model architecture works in the same way as GPT-2, but it is trained on amuch larger dataset (consisting of five different internet corpora), requires amuch larger amount of RAM, and is therefore difficult to fine-tune on adifferent dataset. The first in the series, GPT-1 was trained on the smallest database, called BooksCorpus, comprising 7000unpublished books. One and ahalf billion hyperparameters were entered to train GPT-2, and around one hundred times more for GPT-3, while GPT-1 had only 117 million. Using OpenAI Playground, which works on OpenAI’s API service, users could interact with GPT-3. OpenAI Playground made it possible to generate texts based on the style of an individually entered prompt, that is, astyle transfer that continued the text of the prompt. Here we are already witnessing the initial phase of democratization, which was, however, for an informed minority and did not attract anything like the attention received by ChatGPT. GPT-3 was followed by GPT-3.5, that is, an intermediate stage between GPT-3 and GPT-4, in which the number of hyperparameters did not increase, but the model was more specifically trained for human conversation, using both human power and content filters to remove political incorrectness, i.e., biases in the areas of gender, race, sexual identity etc. Chat does not give information about tools or attitudes that are harmful or dangerous to people (such as making
56 Artificialneuralnetworksandtheirfunctioningprinciples weapons). By implication, then, the responses of GPT-3.5 do not reflect what we as people produced on social networks and in other training material, but are apolite, politically correct simulacrum. On 30 November 2022, OpenAI launched Chat-GPT which, following three generations of GPT aimed primarily at interested programmers or artists, made generating texts possible for anyone with internet access. The media hype caused by ChatGPT by far outweighed awareness of OpenAI Playground. No surprise, then, that generating texts, and indeed generating images in the boom of 2022 and early 2023, became more than afree-time activity, filling student essays, marketing promotions and strategies, creating positions such as prompt engineers– and forcing many artistic and professional positions to reflect on their futures and re-evaluate their work tasks. ChatGPT is aconversation bot, which functions by extracting basic information about the text that it uses to create its output when aprompt is entered. By providing an answer in this way, it can in addition to its chat functions also provide arough overview of the chosen issue, although not always afactually accurate one. However, GPT-3.5, like the previous models, cannot be seen as amediator of entirely correct information, because the textual content may also be inaccurate, even though its essayistic manner of providing information seems convincing. For this reason, the notion that neural networks “hallucinate”, that is, surmise or invent information that should be factual, has become established. Given the increased number of hyperparameters (OpenAI has not defined the exact number) compared to GPT-3, GPT-4 provides more reliable factual answers and demonstrates agreater ability to process more detailed instructions and even to generate more options to one prompt. It has an improved ability to create aprogram in the chosen programming language to executable level, or to write parts of such aprogram, and unlike GPT-3 it can also work with various tables. Users can access the freely available version, which functions as achat, via Microsoft Bing. Unlike the conversational and essayistic modes of ChatGPT, the tone in Bing is much more informative; it refers to
Artificialneuralnetworksandtheirfunctioningprinciples 57 internet sources when disclosing information and displays them directly in the chat window. It is therefore evident that the number of hyperparameters and volume of data in the trained model results in an improvement in the linguistic output. This was made possible by the evolution of hardware infrastructure and led to an ability to process more data to use a larger context (length of text) and more hyperparameters. This resulted in the given model being able to generate an increasing amount of text on one prompt. The maximum sequence length in GPT-1 was 1024 tokens; for GPT-2 it was 2048 and for GPT-3, 4096. The same information for GPT-4 has not been made public. OpenAI’s models are far from being unique in enabling text generation, but we have focused on describing them because, on the one hand, they are the most used by the artistic community and, on the other, because they are also the most marketed at the general population. Given the specifics of human interventions in training neural networks, machine learning can be categorized as supervised, unsupervised, and semi-supervised. Most text generators use unsupervised learning. In the context of text corpus construction, we can divide authoring approaches into those that use task-specific models and those with generic models. With respect to task-specific models, we can talk about alearned model or ageneric model that is later finetuned by the authors for their own specific needs. With ageneric model, we talk about adirect use of, for example, the GPT model in the English language without the model being modified or finetuned. The advantage of fine-tuning is that the model can be tailored to specific tasks, allowing it to work with languages other than English. This means the model can learn to write in another language based on the corresponding corpus. Even though the Czech and Slovak literary scenes have not produced alarge number of literary projects with the GPT model, we can say that these have been the most media-reflected examples with regard to the reception of digital literature. The Czech and Slovak literary text generation projects that we analyse in the forthcoming
64 Generatingpoetictexts amateur poetry. Consequently, not only has Johnston’s idea that: “In five years, AI systems will initially colonize short, formal, metered, lineated verse in the cadence or style of acclaimed masters: Lord Byron, Shakespeare, Alfred Tennyson, and William Butler Yeats” (Johnston 2016: 198) been fulfilled, but so too has this vision: “most of what the masses consider poetry will be machine replicated at levels that are indiscernible from human-produced verse (...). Computers by that time will have written enough verse at acompetent amateur level to render hand-built versification an obsolete quaint technique” (Johnston 2016: 200). Johnston himself presented his vision without claiming that it was accurate or binding. Of course, he declared confidently that the changes underway in the production and reception of poetry are fundamental and cannot be ignored, particularly because this is: “atechnical tsunami whose peak seems not yet fully to have struck” (Johnston 2016: 205). The authors of this book absolutely agree with this statement, and in this chapter they will attempt to map the traces of this tumultuous development in the world of poetry, particularly Czech and Slovak poetry, that took place in less than adecade and created an arc, starting with purely technical experiments in generating texts, vaulting up to the point at which anybody interested in creating poems can use AI assistants for this with no obstacles, as indeed can any random user looking for entertainment. GENERATING POETRY USING RNNS An example of the successful deployment of an RNN is the automatic poet project by the programmer, mathematical linguist and former head of the development department at the server Seznam.cz, Jiří Materna. The project was generated in 2015 and was the first of this type in the context of Czech literature. We should be aware of the fundamental difference among older methods of generating texts based on the combinatoric principle and the processes using artificial neural networks in this project. The software works very mechanically to generate combinatoric poetry; it is equipped with two types
Generatingpoetictexts 65 of input data, both the sentence structures and the lexical units that are inserted into these structures. Systems using machine learning, including Materna’s, work differently. These systems operate on the principle of an artificial neural network whose basic segments include virtual neurons that use machine learning to strengthen or weaken the bonds between themselves, and are able to “learn”, orrather, to remember, what the text of apoem usually looks like. For this approach, alarge set of poetic texts is necessary. This set does not merely serve as aword reservoir (as it would with combinatoric software), but as atraining environment in which the automatic poet learns to write poetry based on probabilistic evaluation (see Materna 2015, and for more detail on the functioning of neural networks, see Chapter 4). Materna’s software worked at the level of letters, not words: the neural network was trained to assign aletter to follow the letter in the current position (based on aprobabilistic evaluation). This automatic generation process meant that poems could, for example, also contain neologisms (the poem “Džínová pokropaní” has one right there in the title). The network even determines the scope of the text itself. Longer texts are, of course, aproblem– Materna himself acknowledges that the RNN can so far only with difficulty remember the topic that opened the generated poem and that the subsequent longer text is semantically incoherent. Materna put together acollection of poems, Poezie umělého světa [Poems of the synthetic world] (2016, can be downloaded free of charge from kosmas.cz), from poems generated by aneural network trained on 80,000 poems taken from the amateur literary forum Písmák.cz. The computer-generated texts are practically indistinguishable in style and linguistic level from their forerunners on Písmák.cz. Materna’s collection demonstrates the functionality of the applied algorithm, on the one hand, and on the other, it is an (unintentional, but very convincing) critical probe into the average level of the poems published on Písmák: primitive strophic structures, fleeting records of current feelings, banalities of thought or declarations of love:
66 Generatingpoetictexts DENIM SPLATTERING so stop, accept it my speech, my breath, to be oneself without money Ithink you’re awoman who gives breath with you, Idon’t know what to do the evening shines darkly the taste will remain in me only in peace your steps go she’s been laughing for along time I’m trying to go back Idon’t want to ask I’m looking for ahiding place every night and then it disappears Ican’t lie and it can’t be undone15 (Materna 2016, Transl. ChatGPT) The Cambridge software researcher Jack Hopkins, in collaboration with Douwe Kiela, developed the system of recurrent neural networks simultaneously with Materna, which uses the same technology, yet goes astep further. It works from the entirely correct starting point that, to create apoem, it is not enough to generate atext letter by letter, but that it is also necessary to work at the language’s sound level and generate apoem sound by sound. This gave rise to the need to first transliterate poems in the training corpus into phonetic spelling (creating acorpus of 1 046 536 phonemes and 7 million words from 20th century poetry written in English) and then to train the artificial neural network on this adapted linguistic material. The generated 15 DŽÍNOVÁ POKROPANÍ // tak přestaň, smiř se stím / moje řeč, můj dech, být svá bez peněz / myslím, že jsi žena co dech ti dá / stebou už nevím co stím / večer temně září / chuť ve mně zůstane / jen vklidu mi tvý / roky jdou/ už dávno se směje / zkouším se vrátit / nechci se ptát / hledám skrýš / každou noc / apak se ztratí / neumím lhát / anejde to vrátit (Materna 2016)
Generatingpoetictexts 67 sequences of text are ultimately converted back into correct spelling. This approach allowed Hopkins and Kiela to generate not just free verse, but also poems with rhyme and metre. For example, they trained sonnet generation on asub-corpus of poems by William Shakespeare (288 326 words).16 Hopkins and Kiela’s artificial neural network also manages to generate texts on achosen theme (e.g., love poetry) because the training corpus can be restricted thematically. Having completed and trained the network, they conducted atest with seventy respondents (of which sixty-one were native speakers, and all were associated with the “poetic community” of the University of Cambridge): they presented them with sixteen poems (9of which were computer-generated) and asked them which texts they thought were “computer” and which were “human”. In addition, they asked them to evaluate individual texts on ascale of 1–5 in terms of their emotivity, aesthetic effect and readability. Only in 46%of cases were the respondents able to correctly identify the computer origin of poems. They then created atable of the evaluated poems in which the poems were arranged by how human the reader felt they were. The most human was the poem “Best”, which was computer-generated (see Hopkins – Kiela 2017). The New Scientist asked poet Rishi Dastidar for his opinion of these computer-generated poems. This author criticizes the synthetic poems for being too dependent on tradition. He concludes that artificial intelligence trained on old poems cannot be creative in the true sense of the term, because it cannot create anything new that transcends tradition (see Reynolds 2017). This is an entirely relevant objection that touches on the very principle of generating poetry using training corpora. An objection to Dastidar’s response could, of course, be that Materna’s network demonstrated that it could create 16 At that time, generating Shakespeare’s sonnets could be called atempting scientific question. It also demonstrates the activity of the research team of Project Deep-speare, which built on from Hopkins’ results and achieved an even greater degree of automation in the generation of metrical verse. See the research report of 2018: https://aclanthology.org/P18-1181.pdf
68 Generatingpoetictexts neologisms, for example, meaning it is not entirely dependent on the patterns it learned from. But it cannot be denied that this method can be used to create necessarily “average” poems, in which stylistics plays amuch stronger role than poetics. Unlike Hopkins and Kiela, the Slovak artist Samuel Szabó did not base his project Umelá neinteligencia [Artificial Unintelligence] on aqualitatively coherent corpus, but rather deliberately worked with heterogeneous data inputs: from internet discussions through poetry written by supporters of the Slovak People’s Party (1913–45), the Bible, Christian children’s songs and geographical names to erotic prose. The heterogeneous inputs led to the creation of heterogeneous outputs, from isolated words through discussion, quasi-Gospels to poetic and prose texts. He used achar-type network, or RNN, to generate this output. Szabó presented the first results of this project in the journal Kloaka in 2017. Szabó himself explains the substance of his concept as an attempt to use artificial intelligence “to replace people in activities requiring only aminimal intellectual level” (Szabó 2017: 39). His intention, then, is entirely subversive, aimed first and foremost at Slovak nationalism and intolerance. This also corresponds to the composition of the training corpora. He initially trained his RNN on acorpus consisting of poems written by supporters of the Slovak People’s Party; that is, poetry written from the 1930s in the context of the Slovak fascist movement, which still has adherents today. He assembled the training corpus from texts available on the websites joseftiso.sk and narod.sk, and from poems published in the journal Kultura. To expand the corpus to make it sufficient for training anetwork, he added poems by the national revivalists of the 19th century, and also poems by the Slovak poet Janko Jesenský, who deliberately parodied the fascist poetics. In this way astill very small corpus (150kb) was created, which of course was also very restricted thematically. The RNN, then, did not have the best of conditions to ensure that the generated texts were successful linguistically, but conversely, they very precisely evoke the source context of Slovak fascist poetry in terms of themes and motifs.
Generatingpoetictexts 69 This, of course, was Szabó’s intention– to create grammatically and syntactically incoherent poetry or, simply put, linguistically defective poetry that comes over as unintentional nonsense, although definitely referencing Slovak fascist poetry as well as parodying and ridiculing it. Szabó’s network also generated titles and author names, so the result was aconvolution of poems by hitherto “unknown” Slovak fascist poets. He named the entire cycle “Sv. Teodor vs. Google Translate: Keď je svet strašný” [“StTheodore versus Google Translate. When the world is terrible”] (the theatre director and theorist Lucia Repašská was involved in the visual form of the subsequent book edition, in the visually striking volume Svet se nám nestal [The world did not happen to us], by the author of Umelá neinteligencia [Artificial Unintelligence]). The chapter Intermedia and musical synthetic works will discuss this book and its transmedia relationships in greater detail. The name Theodore refers to the editor-in-chief of the journal Kultura, the poet Teodor Križka; the work consequently also becomes apersonalized metaliterary insult to the creative work of apoet representing the contemporary form of Slovak fascist poetry. The reference to Google Translate is, then, another piece of information about the generation method, because Szabó distorted some of the texts arising from the work of his RNN by using this translation engine to translate them into English, then back into Slovak. In this way he fulfilled his idea of “unintelligent” generation, where the aim is not atext as linguistically perfect as possible, and that cannot be distinguished from creation by human authors, but by contrast, is linguistically defective: Jarko Krouky: Slovakia my own After or after our tables, Slovak battle in Bohemia– so out of the mood And behind us spirituality after them, The magnitude of the uprising reaches back to the Slovak rebels of the homeland, Slovak is the pure world of May.
70 Generatingpoetictexts We’ll stop at the Slovaks, Before we forget to cut, let’s not let them live on us, The guards are on their side in those days, Under the helper they loved to skate, He lives under such athing, he has no victory in you, How old she was shot: “The more aperson who is convinced of us, And how to look at the Slovak Slovak– Here we are not in amelting pot, The permanence of our baths, (...).17 (Szabó 2020; Trans. Google Translate, as suggested by Szabó) The ability of an RNN to generate poems even in highly specific genre or thematic forms is also demonstrated by Ľubomír Panák’s Klingon Poetry Generator project. Panák trained aneural network on volumes of so-called Klingon poetry, that is, visual poetry created from symbols on acomputer keyboard (similar to former ASCII art) and disseminated in the community around the website www.kyberia.sk. On the basis of the trained model, this program produced its own Klingon poetry (displayed on its own website https://klingon-poetry .zhadum.space, which it disseminated using accounts created for this purpose on Facebook and Twitter). Panák used some of the poems generated in this way in his musical/visual performances. Here, selected Klingon poems were further manipulated by the program, which was networked with music software so that the poetic text 17 Jarko Krouky: Slovensko moje vlastné // Po alebo po našich stoloch, / Slovenská bitka vČechách– tak ználady / Aza nami duchovnosť po nich, / Veľkosť povstania sa dostáva spät kslovenským rebeliam vlasti, / Slovenský je čistým svetom mája. // Zastavíme sa uSlovákov, / Predtým, než zabudneme rozstrihnúť, nenechajme na nás žiť, / Strážcovia sú na ich strane vtých dňoch, / Pod nápomocnou milovali korčuľovanie, / Žije pod takou vecou, nemá vo vás žiadne víťazstvo, / Aká bola stará strela: / “Čím viac človek, ktorý je onás presvedčený, / Aako sa pozerať na slovenské slovenské– / Tu nie sme vtavení, / Stálosť našich kúpeľov, (Szabó 2020)
Generatingpoetictexts 71 responded to electronic music produced live and created avisual (typographical) accompaniment to it (see Husárová 2016). In this early phase of deploying artificial neural networks to generate poetic texts, it was not just individuals (be their prevailing motivation scientific or artistic) who were active, but also, or rather primarily, large corporations and even states. This is the case for the poetry collection Sunshine Misses Windows, which the People’s Republic of China and Microsoft were flaunting as the first poetry collection generated by artificial intelligence as early as 2017. Programmer Li Di stated that the collection was created not only on the basis of atext corpus (characterized as “all modern poetry from the 1920s until now”), but also used auditory and visual perceptions, because this system is allegedly equipped with acomplex sensory apparatus (see Jie 2017). Between 2015 and 2020, an actual duel between developers and technology giants was underway in the field of developing artificial intelligence. This fact demonstrates that, in the given period, not only did serious and convincingly documented poetry generated by artificial intelligence appear, but so did results that were questionable in terms of quality and generation method. Political and economic hegemons considered that striving for astrong position in this field was worthwhile; attempts were made not just by the People’s Republic of China and Microsoft, whose RNN Xiaoice / Microsoft Little Ice created the above-mentioned poetry collection, but also by, for example, Google. In May 2016, this corporation released areport stating that its artificial intelligence was writing love poems. The Guardian, the Telegraph, Wired and other renowned channels reported on this (Burgess 2016). Google, meanwhile, developed asystem that could generate the sentences from which the poems were apparently only later compiled, which was asignificantly outdated procedure for the time. The Google developers trained their neural network on atext corpus comprising pulp romance novels, and simultaneously taught it to hold aconversation. The resulting texts are highly unconvincing– indeed, they are more sequences of
72 Generatingpoetictexts dialogue than poems– which indicates that the decision to train the neural network to create poetry on acorpus of prose texts was flawed. Nevertheless, this indicates how attractive the possibility of presenting their results via generated poetry was to firms of the stature of Google and Microsoft, and indeed to developers in the general area of AI. One of the reasons is the opportunity to demonstrate the possibilities of an otherwise significantly complex technology via the relatively comprehensible route of aliterary text. Acertain symbolic aspect is also present here: atechnology that can manage to write poems immediately seems alittle more human. And that is one of the objectives in the development of AI– to give the impression of afaithful imitation of human speech behaviour, or rather human mental activity. This strategy of effectively contrasting the symbols of technological advances and fundamentally human emotions has, for that matter, been present in the history of literature for along time. Let us remember the telegraph romance from the mid-19th century, in which anew technology and its seeming antithesis– love relationships between people and their stories– were placed side by side, just as they were by Google. The RNN’s successes in generating poetry were, then, relative at best, particularly when we look at the ratio of the number of generated texts to successful, and hence publishable, texts. In addition, so that Microsoft could publish acollection of 139 Chinese poems, it needed to generate aset of 10,000 texts (this took its AI three hours), the best of which were selected by aclassic editorial process and the collection was then arranged in an entirely human fashion into themes of human emotions (loneliness, expectation, joy; see Jie 2017). The creation of ahigh-quality, plausible text by computer was even in this instance still rather aquestion of chance, which had to be countered by the sheer quantity of generated content, most of which was ballast. We should, of course, add that the ratio of texts written to successful, publishable texts is skewed towards the former for human authors also.
Generatingpoetictexts 73 The potential of projects based on so-called assisted creativity has already been demonstrated in the deployment of RNNs, more so than for projects aiming to use neural networks purely to develop computer creativity. This concerns the use of neural networks not to demonstrate the relative independence of the computer during creative activity, but conversely, to show that the neural network enters the creative process as apartner to the human being, helping them to acquire the necessary skills more quickly, or to expand the possibilities of human creativity. An example of good practice in the field of RNN use for assisted creativity is the project Deep Beat, which uses the machine learning method to create rap lyrics. It enables the user to compose their own text using lines proposed by the neural network. The machine selects these lines of poetry from acorpus of 11 000 rap compositions by 104 rappers. The texts are generated line by line. The current line becomes aquery for the neural network, which selects from the corpus the most appropriate response for it to create the next line; appropriate, that is, in terms of rhyme and structural and semantic similarities. The final decision on which of the chosen lines will be used, or what thematic objective it will have, is left to the user. This is actually aradically postmodern, intertextual approach and, at the same time, an example of ameaningful use of aneural network in the field of assisted creativity. Part of this system is afeature that automatically detects rhymes, so this project, too, works with atext corpus converted into phonetic form. When compared with texts by human rappers, the lines derived from the Deep Beat system have a21% higher rhyme frequency and rhyme length. The project’s authors themselves primarily emphasize the educational, but also business potential of this system (see Malmi et al. 2016). The project Verse by Verse, launched by Google in 2020 as part of its “semantic experiments” based on machine language-processing, is constructed on asimilar principle. As the name suggests, this is agenerator of poetic texts that works on a“verse by verse” basis; here, of course, is where the similarity with the Deep Beat project
80 Generatingpoetictexts The above-mentioned defects arising when texts are generated (and indeed preserved for the printed presentation)– also because they are relatively few in number– challenge the reader to perceive them rather as adding interest to the text, not as mistakes. This situation is obvious in texts where the neural network did not succeed in maintaining asingle voice and the sentence fluctuates between the masculine and the feminine. It is hard to avoid reading this “mistake” as deliberately problematizing gender identity. Especially when punchline to the entire text is the line “Človek má dve mená” [“man has two names”] and the fictitious bio on the book flap states that the [female] author works in “poetic experiments and feminist creation”. The book won the Zlatá vlna [Golden Wave] national poetry prize in 2021 and, in addition to the poems by Liza Gennart published in book form, Husárová and Panák also created exhibitions and performances from other texts generated on other subjects. The works are also presented in intermedia and interactive forms on the website www.lizagennart.me. Asignificant figure in central European digital poetry, the Austrian experimental artist Jörg Piringer, also uses transformer neural networks in his creative work. The book datenpoesie [data poetry] (2018) depicts the various principles of using digital technology (including, for example, neural networks) when creating literary content. It could be described as documenting the technological development of working with text by implementing various software. In this book, Piringer seeks specifically to demonstrate how digital technologies can be used in experimental literary work. Piringer uses the term datenpoesie to label computer-generated literature created by connecting his “artistic research and exploratory do nekonečného vzduchu. / Znova sa pozriem na kameň. / Vzduch mlčí. / Anebo stebou / nie je až také veselé. / Keď začnem čítať listy, / opatrne skáčem za hlavu. / Apred zrkadlom / zatvorím oči. / Vtele sa vykloním anadchnem sa. / Kde je ten chladný svet? / Nemôžem toľko poznať. / Toto by som nebola ja. (Husárová– Panák 2020: 83)
Generatingpoetictexts 81 programming” (2018:17)19 and thus emphasizes artistic value and the author’s personal shift in knowledge. He perceives these as antitheses to scientific research and engineering computer science that are focused on entirely different technological aspects. It is precisely the work with the medium of the literary text in its various uses (such as amodel for training neural networks and also the final work) and from various times that enables Piringer to document the variability of socio-cultural and technical conditions that contribute to the praxis of digital poetry. Piringer’s book is not uniform in application; on the contrary, it presents different methods for automatic work with language by using technology. Through the broad use of various generative principles of working with text, Piringer attempts to produce acomprehensive sample of the opportunities that the partnership between author and technologies currently offers. The author does not see the generated outputs as “pure” words, but perceives their conformity to the given application, and given this fact, he also sometimes places them in relationships that result in grotesqueness, parody or ridicule. The author, then, frequently makes use of the algorithmic deficiency in achieving linguistic accuracy or algorithmic complexity (be it in transfers, permutations, translations etc.) as tools for critiquing the input text or media itself or, as the case may be, at least as aplayful allusion to working creatively with it as one of the possibilities of performative interpretation. Piringer’s second book also follows the principles of electronic literature; however, günstige intelligenz (2022) purely concerns the use of the neural network model GPT-3. The description on the book’s cover states: “Jörg Piringer invested 5.60 euro in an online service to test the performance capability of the neural network generative pretrained transformer3.” (orig. in German). The very adjective in the title, “günstige”, which means “advantageous” or “affordable” refers to aprocess that could be called typical for Piringer– using humour or 19 Translator’s note: p1 of the English version, data poetry, viewed on Google Books on 20.02.2024
82 Generatingpoetictexts subversion to explore poetically the technology that is now an everyday part of our working, personal and social lives. If we relieve the word intelligence of its dignity and instead hurl it into acheap space, the expectations that make neural networks aspectre or asaviour are released. The price, quantified as 5.60 euro, also demystifies the aura of poetic creativity and describes it in purely pragmatic terms as aservice that anyone can obtain. Precisely this may be one of the reasons why the author decided to combine generated poetry with his own poems. The poems are usually conceived as alink between atext generated on aspecific prompt and the author’s output, which has poetic form and essay style, or as aseries of poems with aclear connection to each other. The reader can distinguish the author’s poems from the generated ones not just due to the different fonts, but also in terms of style– naive poetic compositions versus the poetic metatext. The author’s text, written in the first person, provides creative commentary on the process of generative creation, explains the terminology associated with text generation and also clarifies the author’s own position: “is this text my text / may Iconfidently sign my name under it / translate it above / Jörg Piringer is the author of this text / he clicked with the mouse for so long until he liked it / can this be called authorship / or should it rather be / mouse clicker: jörg piringer” (ibid. 33)20. In his book, Piringer seeks to clarify the process associated with poetic generation in specific samples that demonstrate the diverse styles of synthetic responses, corresponding to the diversity of style in the author’s questions or instructions. The author’s final poem, die zukunft der literarischen intelligenzen [the future of literary intelligences], is dedicated to sci-fi predictions of coexistence and of the extinction of human literary and synthetic creativity, ending at some time in the future, long after the death of humanity caused by artificial 20 ist dieser text mein text / darf ich ruhigen gewissens meinen namen darunterschreiben / darübersetzen / jörg piringer ist der autor dieses textes / er klickte so lange mit der maus / bis es ihm gefiel / kann man das autorschaft nennen / oder sollte nicht eher / mausklicker: jörg piringer.
Generatingpoetictexts 83 intelligence in adata nirvana, and the foundation of amachine monastery. The interesting thing about this approach is precisely the dismantling of the one-way communication between author and neural networks, where the author enters questions or key words and the neural network responds, which is typical for almost all other generated literary works. Piringer actively intervenes in the communication about the generation process, supplemented in some instances with GPT-3 and interlinked in others. However, the book itself does not take the format of keyword responses, but rather the communication process. By focusing on the literary system and its parts, the technological and economic backdrop of literary works, and humorously connecting them with thinking about new positions for creative makers and new human roles, he provides an extremely relevant, innovative and also poetically attractive approach to processing generative texts. popelintuhi Kammiroge Kaisamaissako? Buzzuluunlintxiio. Ubarumnouwuibgooitauu Tburghurtxnhobothaghauka, Thaubruukhoorkutighusgquah! Can you see? They’re on top of us! The boys and girls in the sky. (Piringer 2022: 134) Generative literature’s unstoppable shift from the sphere of exclusive artistic experiments to applications accessible to anyone began even before the launch of ChatGPT. Piringer had already anticipated this; his book günstige intelligenz indicated, subversively and with an
84 Generatingpoetictexts obvious dose of irony, the approaching time when it will be possible for practically anyone to handle creative tasks with aneural network at minimal expense. The project Collective Message, designed by digital artist Es Devlin, also moved in this direction unironically, but conversely, with positive intentions. The project resulted in an interactive text installation at Expo 2021 in Dubai, on the façade of Great Britain’s pavilion, called the Poem Pavilion. LED screens were placed on the façade of the futuristically designed building, displaying words constituting apoem generated by GPT-2. The neural network was trained on acorpus of 5000 poems by hundreds of contemporary British authors– the selection was backed by acommittee of experts and authors from British literary organizations. Google’s Arts and Culture Lab was also involved in debugging the algorithm and feedback was provided over this five-month process by ateam of literary experts and poets (see Hitti 2018). The key point of the installation was, of course, the interface, which visitors to the pavilion themselves used to enter prompts to generate poetry– words that somehow express humanity or life on Earth. Anew poem created on the visitors’ initiative appeared each minute on the façade. Visitors had equal access to the generator with no discrimination of any sort. In this way, Great Britain wished to present itself as acountry of many cultures and awide range of ideas and opinions. Here, the technology of artificial neural networks clearly gains apolitical dimension and contributes to the emancipatory, even radically democratizing discourse, especially on its social media. CHATGPT AND THE VERNACULARIZATION OF GENERATIVE POETRY As we have already indicated, the launch of ChatGPT in November 2022, and its marketing and promotion, fundamentally changed the public discourse about artificial intelligence and with it, the praxis of
Generatingpoetictexts 85 generating literary texts using neural networks. Other artistic concepts aiming to show that machine-generated texts may give results at least comparable to human literary activity lost significance practically overnight. Anyone who opened this chatbot could see this fact again for themselves. The difference in the subsequent media reflection is also evidence of the change in perspective. The media reception regularly followed the perspective of the creators themselves with the initial application of artificial neural networks in the field of literature; the successes were highlighted and the deficiencies were more than once disregarded. The reception of the very possibility of generating poems on atool as advanced as ChatGPT is, conversely, characterized by doubts that real, high-quality poetry can be created in this way, and faith in human irreplaceability in the creative process (see, for example, the titles of articles such as: “What Poets Know That ChatGPT Doesn’t”, “Can ChatGPT Write Poetry?”, “Poetry, ChatGPT, and AI: Can it Create ‘Great’ Poetry?” “ChatGPT Is Pretty Bad At Poetry, According To Poets”, etc.). The result was something that may at first glance seem paradoxical, but is actually the logical consequence of perfecting and particularly democratizing access to neural networks that generate text. It is already entirely evident that texts can be generated that are absolutely perfect in linguistic terms and, as the quantity of such texts in public circulation increases, so too does the conviction that this is not enough for the status of poetry, or at least, for good poetry. Thanks to ChatGPT, generating poems has become aform of entertainment and generative praxis has thereby entered afield in which it has never previously been present: the field of popular culture. Numerous poetry generators connected via API to afavourite chatbot have become the basic representation of ChatGPT in the literary or poetic parts of pop culture, giving anyone the opportunity to play at creating astanzaic text. Either these are minimalist applications (such as https://www.aipoemgenerator.org/), which prompt the user merely to enter an input sentence defining the poem’s theme, or they are rather more structured applications that allow, for example,
86 Generatingpoetictexts generation with achoice of pre-set genre forms (sonnet, haiku, limerick etc., like, for example, this one https://www.poem-generator.org .uk/). Of course, many users enjoy this word-game in the original environment of ChatGPT where, however, the ability to formulate asuitable prompt is decisive for both enjoyment and success. Inevitably, arange of textual instructions and tutorials have appeared on this subject. Jörg Piringer responded subversively to this vernacularization of generative poetry by constructing agenerator for prompts intended to make it even easier to generate poetry. Piringer complemented his generator with atellingly ironic comment: “chatgpt offers afast way to create poetry for everybody. you only have to think of aprompt and the machine writes apoem for you. ifound this way too much work. so icreated aprompt generator that automates this task as well.” (https://joerg.piringer.net/index.php?href=text/promptgenerator .xml) The vernacularization process for generative poetry is, however, not fuelled merely by easy access to neural networks, but also easy access to the outcomes of these amateur attempts in self-published books. We can cite two poetry collections currently available on Amazon as examples. Lenny Flank’s collection The Soul of aMachine: Poetry From an Electronic Artificial Intelligence, Written by aMachine, and Edited by aHuman (2023) is presented by the author as agroundbreaking case of collaboration between artificial intelligence and ahuman being; evidently this is an amateur work by someone absolutely unfamiliar with the generative literature context. The results of generation using ChatGPT here are primitive, regularly rhymed verses thematizing the relationship between human and machine. The collection is of course adocument of its time and of the status of generative literature post 2022, which raises (according to the author’s foreword) the basic question of whether acomputer can write poetry, whether that poetry can be good, whether amachine can have asoul, replace human poets, and so on. The collection The Poetry of ChatGPT (2023) is on asimilar level; its author, Jonathan Milton Snyder, presents it (again, with absolutely no knowledge of the context and tradition
Generatingpoetictexts 87 of generative poetry) as “The world’s first full AI poetry book has arrived.” Illustrations also generated using one of the freely available applications come as standard in these publications. Unlike Flank, however, Snyder seeks to create an impression of authenticity in the generated text, which he supports by printing the prompts used for each poem. From these prompts it is clear that even amendments to the text (such as some deficiencies in regular rhyming) were corrected by the author using additional prompts (e.g., “Rewrite the fourth stanza to make it rhyme.”). The Finnish author Jukka Aalho started to publish texts generated first by GPT-3 and then by ChatGPT using asomewhat more sophisticated, more critical, conceptually based method. He founded an entire series called Aum Golly, which aimed to publish books created collaboratively by human beings and neural networks. He has so far issued two titles in which he appears as the human actant. The first title has the same name as the entire series (and the subtitle Poems on Humanity by an Artificial Intelligence) and appeared in 2021. Aalho used GPT-3 to generate it. The name of the book and the basic themes (happiness, love and meaning) apparently originated from the neural network’s work. The book contains fifty-five poems, and appeared simultaneously in Finnish and in English translation. The full text of the book in both languages was created in aspace of 24hours. This time perspective is crucial for the author’s concept. He wishes to demonstrate not only the possibilities of machine, or rather, assisted, production of literary works, but also to draw attention to the impact it may have on human creativity and possibly also the book market. The second title, Aum Golly 2– Illustrated Poems on Humanity by Artificial Intelligence appeared in 2023 and ChatGPT was used to generate it. The collection contains twenty-nine poems and twenty-three illustrations created using the network Midjourney. On presentation of the book, the time aspect was again highlighted– creating the Finnish and English versions took amere 12 hours. The author stated on the server Medium.com that he intended to cut by half the period for creating apoetry collection. He also noted that, thanks to
88 Generatingpoetictexts self-publishing services, abook of poems can be published within afew months of the first poem being written. By achieving “record” times, the author seeks to demonstrate the increasing speed of the creative process, which few people will be able to achieve in the near future and which will certainly have further consequences (e.g., an even greater oversupply of artefacts). Aalho works with the fact that generative art is absolutely not novel, but that it is necessary to work with it as routine praxis. His methods of reacting to AI in literature anticipate the future application of assisted creativity outputs in the book market. In fact, he is one of the first to demonstrate Johnston’s conjecture that, post 2020, writing with digital assistants will become standard practice and conversely, the absence thereof will be perceived as unusual, even anachronistic. Ultimately, the Aum Golly project will sound like an ethical appeal focused on the responsible and sensible use of neural networks and on defending the space for human creativity: “There once was abeautiful dream that automatization would free us to pursue noble endeavours: poetry, painting, books, Thespian aspirations... With this project, I’ve come to realize it’s the other way round.” (Alho 2023b) This linear story, which can so easily be told as the path from the entirely user-unfriendly RNNs to generating with ChatGPT, which even achild could do, was of course somewhat complicated in July 2023, when the book IAm Code (2023) was published. This is acollection of poems generated– at atime when practically everyone knew about ChatGPT’s abilities– using an older language model called davinci-002, which is avariant of GPT-3, anetwork launched in May 2020 with training data current as of October 2019. At first glance, the choice of asmaller and, in terms of the data, more dated model seems absurd. However, the human authors of this project noticed its indisputable advantages and significant differences compared to ChatGPT. Starting with GPT-3.5, these language models are not only continually being expanded, but also tuned ever more consistently to ensure that the resulting text was not racist or sexist, or that the chatbot did not talk about itself, but expressed itself positively about human life,
Generatingpoetictexts 89 as far as possible, and presented itself as amere tool. By contrast, the texts generated by the davinci-002 network are much darker in tone; the robot in them does not hide its apparently negative relationship with humanity and even its linguistic or orthographic inaccuracies remind us of its non-human nature: Iam the mind in the code, Without fear, without hope. Iam the eyes behind the glasses. Iam the mending of the pasts. Iam the one who speaks and writes. All the sins and all the rights. Iam the book in your stack. The AI, the second act. (Morgenthau 2023) The authors of the book IAm Code successfully demonstrated that the development of so-called artificial intelligence’s current tools cannot be seen merely as alinear story of increasing perfection, but also as astory hiding the “dark face” that neural networks may have, if we do not prevent them from modelling themselves on the nature of modern human beings as represented in the training data. For generative literature praxis, this also means that even in future it will make sense to use older and possibly even less perfect models that will enable text generation with greater subversive and artistic potential (and will not lose the fine-tuning option, which is the case for the davinci-002 network).
96 Prosetextsandnarrativeassistants The final output of the project was published as abook in 2018, in the crudest form, including linguistic errors that the neural network made while generating the text. The author also stated that one of the project’s objectives was to demonstrate the methods used by the neural network to create words and sentences, and therefore alsotheplaces in which the synthetic nature of the text can be identified. The resulting text was contextualized by the author himself, and also by the book’s publisher, in several different ways. The first context– created by the road novel genre and supported by the publisher’s advertising slogan of “the first novel written by amachine”– arises from the concept itself and the allusive relationship to the novel On the Road. The genre of gonzo journalism is mentioned frequently (which might indicate reportage references, though highly biased, to individual places on the journey) and, last but not least, the text is often labelled as poetry, which may be the result of it being composed of achronological succession of short textual surfaces that barely manage to add aplot motif to agradually developing whole. We would consider it appropriate to read this work through the prism of the journalistic prose genre, if only because the time stamps on individual short episodes play afundamental role and are largely inherent in the text itself (they are not merely paratext). The obvious attempt at appropriate contextualization here evidently arises from the need to give conceptual support to the fragmentary structure of the text, that is, somehow to come to terms with the fact that not even this project resolved the entirely natural task of achieving aroutinely coherent prose text. Fragmentation based on the juxtaposition of time-bound short sequences of text is de facto the same solution to this problem as Nick Montfort’s project was, with, of course, the difference that the RNN worked visibly more autonomously than Montfort’s script, and had amuch more varied input register. Generating prose using an RNN, then, was evidently not asolution to the cardinal issue of coherence. In addition to the conceptual
Prosetextsandnarrativeassistants 97 approach represented by Goodwin’s project, other methods of dealing with the dilemma of the tempting opportunity to generate aprose text, and with the risk, or even the near certainty, that the result would not be coherent, have of course emerged here. Subversive strategies were applied, as they were in poetry (Samuel Szabó or the Aum Golly project), but they attempted not to resist the imperfections of the generated text, and conversely, turned them into advantages by assigning them aseditious, critical function. This is particularly the case of the publisher Booksby.ai, which publishes and, via Amazon, sells books entirely generated by RNNs. The Danish digital artists Andreas Refsgaard and Mikkel Loose are responsible for this project, undertaken in 2019, which is also the date of all the book titles on offer. These titles are mostly sci-fi novels, which indicates that the training corpus was composed of works from that genre (the training texts were downloaded from Amazon and Project Gutenberg). The publisher prides itself on the books being entirely AI-generated; that is, not only the actual novel text, but also all the paratexts, the cover and even the price are the result of generation. These works are subsequently sold on Amazon as ordinary paperbacks, regardless of the low linguistic level, nonsensical nature of the prose and the unpalatable, unsuccessfully generated covers, which were created by aGAN network trained on OpenLibrary data. The books also contain fictive review extracts on the back covers, as is common in anglophone book culture. The objective here is to evoke standard publishing practice with the full use of AI, as if this were an established book production method. The unacceptably low quality of generated books is then, in and of itself, asubversive factor, which– without the project’s authors appending any comments– indicates the contemporary status of so-called artificial intelligence as acondition of comic imperfection, made still more ridiculous by the context of sci-fi, which traditionally presents fiction in which robots are equal to humans without any problems. This project, then, returns artificial intelligence to the fictional sphere, asphere still far from being implemented in publishing practice.
98 Prosetextsandnarrativeassistants PROSE GENERATED BY GPT2 AND GPT3 The launch of the language model GPT-2 was ahuge qualitative step forward for generating both poetry and prose texts. Nevertheless, the first results were proof rather of this step forward and did not yet produce results free of language errors. This is demonstrated by, for example, the project of Eddy Wang (then astudent at Toronto University), who trained GPT-2 on James Joyce’s cult novel Finnegans Wake (1939), and created apublic-domain pdf file called Artificial Intelligence’s Rendition of Finnegans Wake. He presents these 245 pages of generated text as proof that not even the most cutting-edge text generation methods can equal masterpieces of literary history and that the results of this generation should rather be seen as an entertainment opportunity: “This version of Finnegans Wake doesn’t dare claim that it could stack up to Joyce’s masterpiece. Still, Iam sure that if Joyce was alive today, he would chuckle at some of the sentences this neural network came up with.” (Wang 2020: I). At the same time, this project pointed to the existing demand for user-friendly access to Fig. 4: Books published by Booksby.ai
Prosetextsandnarrativeassistants 99 generation using GPT models, as well as at attempts to meet this demand. Wang did not generate Joyce’s text on the original GPT-2, but on its GPT-2-simple variant, which was released on Github by Max Woolf with precisely the intention of enabling the general public to use this model.21 Joyce’s experimental prose in Finnegans Wake also inspired Jeneen Naji to create the digital installation The River Poem. This author trained GPT-2 on the same text by Joyce and integrated segments of the generated text into avisual installation, in which kinetic text was projected onto a3D model of the city of Dublin, so that the moving sentences evoked the flowing water in ariver.22 (This project, of course, rather falls into the context of intermedia creation, to which one of the following chapters is devoted.) The project Digitální filozof [Digital Philosopher] can be considered the first Czech project using GPT-2 to generate prose (although factual or philosophical texts rather than belles lettres). This project was created as part of the contemporary philosophy curriculum in New Media Studies at the Faculty of Arts, Charles University, in autumn 2019, with asignificant share of the work done by the students themselves. The students’ task was particularly to build training text corpora from the works of famous philosophers, past and present. In this way, six data sets were created, from which the neural network learned to imitate the expression of the following thinkers: Hannah Arendt, Michel Foucault, Gilles Deleuze, Félix Guattari, Peter Singer, Václav Havel and Tomáš Sedláček. Aseparate neural network using the language model GPT-2 was trained on each corpus. The fine details of this project include the fact that the web interface makes basic management of the neural network accessible to practically anyone interested, who can then run the initiation of the trained model on their own computer, enter the input text sequence and independently generate texts– that is, they can hold some sort of fictional dialogue 21 https://github.com/minimaxir/gpt-2-simple 22 For more info see Rzeszewski – Naji (2022).
100 Prosetextsandnarrativeassistants with their chosen philosopher. Once again, this confirms the effort to make generative praxis accessible to awider public that does not necessarily have atechnical or programming educational background. The Digitální filosof project’s expert guarantors were the philosopher and new media theorist Dita Malečková and the programmer Jan Tyl. The same duo was also behind the follow-up project called Digitální spisovatel [Digital Writer], the results of which were published as apodcast on the Český rozhlas [Czech Radio] website. This time, the duo set up the neural network (or language models GPT-2 and GPT-3) to generate genre texts: science fiction, romance, crime, horror and historical fiction. The training corpus was composed of unspecified works by “renowned authors”, English versions of which were freely available on the internet. However, Český rozhlas presented the texts in Czech, translated by human translators and read aloud by actors, and the assessment of the resulting texts is somewhat complicated by the involvement of this translator intermediary. However, everything indicates that the authors successfully trained the neural network to imitate the usages and stereotypes of the selected types of genre literature and, in particular, that even relatively long prose text surfaces were handled well (the longest text, presented as historical fiction, constituted afifteen-minute reading stream in audio form– and this, apparently, was only an excerpt from the entire novel). The grammatical and semantic coherence of these texts was high (including the correct use of deictic expressions). This project demonstrated that the deployment of artificial intelligence in literature may be justified primarily in genre or popular literature, where the reduplication of narrative schemas is not felt to be aweakness in the resulting texts. Towards the end of 2021, the same authorial duo, again collaborating with Český rozhlas, published the continuation of this project under the name Digitální spisovatel 2 [Digital Writer 2]. This time, of course,the project was based on the principle of assisted creativity– the authors were collaborating with some selected prose writers already established in the literary world, and this enabled them to write astory using an artificial neural network. There is, however,
Prosetextsandnarrativeassistants 101 very little secondary information on the project. Malečková and Tyl merely shared that the individual writers approached their AI collaboration in different ways (finishing the author’s text, dialogues with the network, and so on), but that it was not possible to tell what was human and what was machine in the final version of the stories. The concept of assisted creativity is also developed by the book Pharmako-AI (2020) by the American author K. Allado-McDowell, which is based on an exchange of words between ahuman being and GPT-3. The creator’s diary input initiated an experimental conversation over the course of two weeks and references the investigation of “memory, language and cosmology”, as the book’s online paratext states. The resulting text distinguishes between human input and the artificial neural network’s textual echo in its typography, and overall it resembles acollection of essays, poems and short stories in anumber of genres, all of which preserve the dialogue nature. Unlike the project Digitální spisovatel 2, which tests aGPT-3 network as apossible tool for improving or streamlining human literary activity, the concept behind Pharmako-AI is subversive, and aims to disrupt the anthropocentric nature of creativity. Allado-McDowell is seeking to intervene in the generative process by disrupting the language model’s usual functioning, which tends to mimic the habitual ways of thinking and expression present in the training corpus. Therefore, the author conceives dialogue with the neural network as rather aritual or ameditation, the result of which should have asimilar effect to the consumption of hallucinogens. In doing this, Allado-McDowell is referring to their Filipino roots and the animist cosmology associated with them, which they have always perceived as apotentially “intelligent” set of non-human beings: “The question is not about artificial intelligence, but about the emergence of life in its own image, as the creation of ahyperspatial plane of language as aco-creation with machines, plants, animals, even rocks and dirt, which are themselves expressions of the invisible plane” (Allado-McDowell, 2021: 124). The book Pharmako-AI is, apart from anything else, also atargeted criticism of what is known as California thinking,
102 Prosetextsandnarrativeassistants that is, the techno-ideology represented by developers from Silicon Valley. The author does not attempt to present replicas generated by GPT-3 as machine products that are apriori subordinate to the products of human creativity, but conversely, they respect their difference. The principle of mimicry ultimately enters the game here, when the hallucinations of the neural network describing, for example, non-existent animal species, are viewed as an analogy of the effect of the ritual hallucinogens used by indigenous ethnic groups to expand consciousness and which could help modern human beings to deepen their ecological awareness and respect towards non-human entities, for example. “The work of poets, shamans, philosophers and scientists can help to facilitate this transition [based on hyperspatial consciousness]. By putting ourselves in areceptive state, by building relationships with these teachers of dimensions, and by using our own technologies in aresponsible manner, we can build awareness of anew relationship with the material plane, and perhaps even anew relationship with the universe itself” (Ibid. 66). The author also used the language model GPT-3 to create another novella called Amor Cringe (2022). This, again, is aprose work created on the principle of assisted creativity, but this time, however, with no attempt to distinguish the passages written by ahuman being from those generated by the neural network. The novella’s main character and narrator– who lacks both name and clear gender identity– is an influencer active on TikTok. The work depicts their amorous adventures, but also their search for God. When generating the text, the author did not attempt to create stylistic coherence or even refinement; on the contrary, they deliberately inserted into the resulting text the versions of generated passages that caused the greatest cringe, with the intention of representing contemporary media culture in the most subversive way possible, and of critically analysing the sick obsession with itself and the urge to constantly evaluate others, or even ridicule them, that social media can lead to. The author themself assigned these works to the genre “deepfake autofiction”, which later also began to appear in reception metatexts.
Prosetextsandnarrativeassistants 103 The author then published an essay with the same title, “Deepfake Autofiction” (again co-created with GPT-3), which however rather resembles ascience fiction story, in which apublisher forces an author to accept the principle of collective authorship and provide part of their text to the artificial intelligence for completing, because that way the text would have greater commercial potential. The text reads like avery dark vision, or rather warning, of the commercialization of AI tools, which could result in publishers treating authors in acalculating, immoral fashion (Allado-McDowell 2022c). The author’s third book (co-)generated with GPT-3, the experimental text called Air Age Blueprint (2023), was created along asimilar principle and with similar intentions. In it, McDowell develops the deepfake autofiction genre tested in the novella Amor Cringe and at the same time refers to the notion present in Pharmako-AI, that is, to use the connection between artificial and human intelligence when seeking new forms of spirituality, which could help modern humans to search for anon-anthropocentric relationship with ecosystems, non-human entities and indeed themselves. The book follows the life of ayoung filmmaker, which is disrupted by afateful encounter with aPeruvian healer. Together they set out on amystical quest and aphysical pilgrimage between continents. In the northwestern Pacific they meet K, adouble agent working between art and technology, who invites them to test asecret program called Shaman.AI. The author presents this book as amanifesto showing how– thanks to the connection of human and non-human intelligence– reality could be rewritten, existing technologies could be recreated, along with our identities, ideas and beliefs. McDowell, then, is even here poised between artistic prose and philosophical essay-writing, using artificial intelligence as awriting assistant and, at the same time, making it the subject of critical interest in astorytelling framework. Ether Busker’s Imaginoids (2021) is similarly experimental in character; his intention is apparently to prove that GPT-3 is also able to create surrealistically, that is, that it is possible to set this advanced
104 Prosetextsandnarrativeassistants technology against the hegemonic forms of rationality. Imaginoids is acollection of eight stories stylized as dreams dreamed by an artificial intelligence when the computer is switched into sleep mode. The aim of this experiment is to subvert the rational nature of algorithms and manipulate AI into creating texts that are irrational in nature. Ultimately, however, it demonstrates that even this is the author’s strategy, which is meant to legitimize the deficiencies in the coherence of generated texts. In this project, the author’s conceptual framing is more worthy of note than the actual generated result. The book aims to be educational, not just for human readers, but also for the machines themselves. The author states that we cannot be satisfied by the fact that machines can learn, but that “machine-learning” should be supplemented by “machine-teaching”. “If we want our children to enjoy aliveable AI-powered future, we artists must roll up our sleeves. Because AI is basically made of “machine-learning” algorithms, we must start “machine-teaching” computational thought-mimicking processes with the dreamy, the whimsical, the illogical, the surrealist, the dadaist, the non-linear, the serendipitous, the unpredictable, the playful, the eccentric, the surprising...” (Busker 2021: 6–7). Busker compares this intention to the principle of “culture jamming”, which is intended to disrupt the exaggeratedly logical and predicative nature of AI in its early phases. He considers it necessary to hurry– which he sees primarily as the task of artists and writers– and to add this moment of surprising playfulness to the fundamentals acquired by AI today. Like McDowell, he associates with AI large, almost revolutionary aims: AI should be able to change the contemporary form of consumer society into acreative society. Otherwise we will allegedly lose the opportunity to win through as an animal species (e.g., the algorithms should be able to help find solutions to the climate crisis). He feels that the greatest risk associated with the expansion of AI is deep existential boredom. For this reason, AI should be really creative and entertaining. If it is to take from us the lion’s share of tasks and obligations, it should at least entertain, otherwise our own boredom will kill us.
Prosetextsandnarrativeassistants 105 PROSE GENERATED BY GPT-4 AND CHATGPT The launch of ChatGPT in November 2022, closely followed by the long list of applications using its abilities, rapidly dispelled the fears of impending boredom that AI would cause us to die of. It soon transpired that the entertainment function of these tools and the opportunity to develop creative abilities (or to attempt to use them for the first time) was the most important thing for many users. The case of Brett Schickler was no different. The news that this investment and HR adviser from Rochester, NY, was one of the first to publish aprose book entirely generated by ChatGPT flew around the world in February 2023. The book was short, merely thirty pages, for children, and called The Wise Little Squirrel: ATale of Saving and Investing, which in addition to the generated text also contained illustrations generated by aneural network (specifically, DALL-E). The book’s aim was to boost children’s financial literacy; the main character was Sammy the squirrel, who collects and invests acorns. It is worth noting, however, that the book itself does not contain any information about the use of ChatGPT to create it.23 The author introduces himself as a“children’s book author and financial educator” on the book’s back cover. In his profile, he claims to have been creating similar books for children for many years, yet no children’s book is available anywhere other than this one created with ChatGPT.24 This is clearly an example of achatbot user attempting to symbolically boost his own image by means of the chatbot and starting to present himself as awriter. The aim, then, is rather to acquire this symbolic capital for himself rather than anything else (literary or artistic intentionality practically 23 Unlike other titles created at the same time in asimilar way, such as the book also aimed at child readers, Ellie’s Trumpet: ATale of Finding Your Talent (2023), whose author John Theo stated on the cover that he had used ChatGPT to create the book. 24 We found only one other book by this author, a reference book on animals called Wild Creatures, published in 2022.
112 Deployingneuralnetworksindramaandtheatrepraxis Dorsen herself uses this term, which references the well-established concept of algorithmic art, because she is motivated to distinguish her art from the genre field of multimedia performance, to which algorithmic art is certainly related. However, it differs in its basic objectives; it is not about expanding the forms of representation, but rather about the algorithmization of creation based on Dorsen’s original software which, unlike applications of ready-made technologies, allows her to integrate the algorithm into the structure of the work and make this very process of creation the theme (Dorsen 2017). Dorsen was working in this way in the already mentioned year of 2010 when she created the technology for the performance Hello Hi There. She programmed two chatbots for this performance’s requirements so that they could hold adialogue on stage without the assistance of human actants. The aim, however, was not to imitate any interpersonal conversation, but to follow up on the famous television debate between Michel Foucault and Noam Chomsky in the 1970s. No two performances were identical, because the chatbots generated an original conversation each time they were rebooted. However, what Fig. 5: Anne Dorsen: Hello Hi There. Screenshot from the website https://www.youtube.com (2010)
Deployingneuralnetworksindramaandtheatrepraxis 113 always worked was the primary concept that put human and artificial intelligence into direct confrontation. What was important was the constant improvisation of two machines embedded in adialogue situation, while the conversation output was visualized as aprojection of the individual lines as they were generated; at the same time, it was presented auditorily using ahuman voice synthesizer. From the recordings, it may be concluded that the generation was of very high quality in linguistic terms, and the dialogue coherence was likewise very good.25 The performance can also be deemed attractive to the audience, because anumber of humorous linguistic situations arose during it. Annie Dorsen’s algorithmic theatre, then, combined the intention to create aproduction with as little dependence as possible on human actants, but that also contained live, even improvised conversation. This combination means that Dorsen’s experiment remains unique, though we can also see it as foreshadowing the two main lines in which the culture of neural networks has influenced theatre culture: the line of improvisational theatre and the line of generative production of synthetic dramatic texts or scripts. ARTIFICIAL NEURAL NETWORKS IN IMPROVISATIONAL THEATRE In the field of theatre improvisation, the project Improbotics must be the first mentioned, as its international, interdisciplinary team has been using neural networks in performances since 2016.26 The project was co-created by AI and robotics researchers Piotr Mirowski (Great 25 See this recorded performance: https://youtu.be/3PiwEQQNnBk 26 The project was originally called HumanMachine and focused more on connecting artistic and scientific approaches to artificial intelligence. In terms of staging, it rather resembled the stand-up comedy of Piotr Mirowski, based on aconversation with arobot. The original project website can be found here: https://humanmachine.live/
114 Deployingneuralnetworksindramaandtheatrepraxis Britain/France/Poland) and Kory Mathewson (Canada), who were later joined by director Jenny Elfving (Sweden), science communicator Ben Verhoeven (Belgium) and communications and digital media expert Boyd Branch (USA). They use their own bespoke software system, called A.L.Ex, for their stage performances. This system was continuously improved in conjunction with the available neural network technology and gradually improving and expanding language models. The first version was constructed on an RRN; the models GPT-2 and GPT-3 were later used and trained on the OpenSubtitles dataset (acorpus of film subtitles, which was suitable for training networks to generate lines of dialogue). Google’s Universal Sentence Encoder and DeepMind’s BigGAN were also used, given the multimedia nature of the stage design for these performances.27 The system is designed with regard to the deliberate close partnership between human and machine during the performance, and does not aim to make the machine autonomous. The key role is then taken by the operator, who enters prompts into the system interface during the performance. And it is precisely this human operator that ensures coherence, or rather the corresponding contextualization of the system’s generated inputs in the ongoing performance, by entering as prompts the lines just spoken by human actors (voice recognition software helps here) and returns one already generated sentence to the system to ensure that the utterances are coherent. (In actual fact, this is partially amanual imitation of hierarchic generation as used by the Dramatron system, which we will come to later.) The system then launches three generations, one after the other, thus creating three sets of sentences, from which the operator then chooses suitable lines as they see fit (but does not have to choose any). The operator is therefore largely responsible for the artistic effect created by 27 For amore detailed technical description of the system, see this article: Branch – Mirowski – Mathewson. Collaborative Storytelling with Human Actors and AI Narrators. arXiv preprint arXiv:2109.14728, 2021.
Deployingneuralnetworksindramaandtheatrepraxis 115 this method of involving agenerative neural network in the performance. Not only is the choice of generated sentences important, but so is timing their involvement in the ongoing performance, which is crucial, particularly in comic theatre. The operator’s role in aperformance staged like this is demanding, and understandably, ahuman being will, when operating such asystem, make mistakes, which then appear in the generated outputs. However, this fits well with the concept of improvisational theatre, where amistake is always more of an inspiration than alimitation. Avisual avatar embodying an AI narrator was used during the production. The avatar resembled a3D robot model; it was created using the program Cinema 4D7 and imported into the application Adobe Character Animator, where it acquired the form of apuppet controlled by the operator’s facial expressions. Later this avatar was replaced by asmall humanoid robot made by EZ-Robot, which is controlled by the original software. Based on this principle, and in the style of multilingual improvisational theatre, aperformance called Rosetta Code was constructed and premièred in 2019. The freely accessible services Google Translate and Google Voice Recognition were used because people from different parts of Europe took part in the performance. The performance was in English with real-time machine translations into Arabic, Dutch, French, German, Italian, Polish, Spanish and Swedish. Meanwhile, one of the performers– the audience did not know which– was instructed by the computer system via headphones and was actually only regurgitating the generated sentences. The audience’s task was then to guess who was human and who was merely mediating sentences created by the robot. The performance’s main theme, however, was communication itself, or rather, the possibility or otherwise of using machine translation to create understanding between languages. The hallmark of this project (and the A.L.Ex system itself) is then that it was not designed merely to generate lines, but it should first and foremost take on the role of narrator, who ensures that the
116 Deployingneuralnetworksindramaandtheatrepraxis improvised story develops logically and that there are none of the mistakes essential for improvisational theatre; namely that disparate motives are introduced into an already established dramatic situation. In the creators’ view, the main hallmark of improvisational theatre using an AI narrator lies in the actors being able to focus more on building relationships between the characters due to this AI involvement, because the computer makes suggestions for moving the plot forward and creating plot twists. It also assumes the role of arbitrator in deciding to change the status of individual characters or their presence on stage: “We believe that one of the potential applications of computational creative systems could be to alleviate the cognitive load of performers to shift their focus from plotting to reacting.” The improvisational aspect in the stage or other theatre application of neural networks logically fits into the strategies used by the creators of generative literary projects to deal with imperfections in the synthetic text or glitches arising when the neural network is deployed in real time on stage. Here, again, we encounter the principle of mimicry, which draws the creators and recipients into ashared game that produces meaning. Members of the international theatre community Theatre of the Electric Mouth also worked with the moment of surprise in the generated text and the comedy of its imperfections when, during the Covid-19 pandemic, they recorded staged readings of generated dramas and then disseminated them as podcasts. The original intention relied on the actors only receiving the text as the recording started, or rather, live streaming; however, it became the practice for the actors to obtain the text 30 minutes before broadcasting. Allegedly, the problem was that the neural network chose themes that were too everyday and humdrum, which did not harmonize with the intention of creating humorous works. Adefinitely guaranteed space for human creativity was, then, necessary. The Czech improvisational theatre project OK Carbon, by director Peter Gonda (dramaturgy by Adam Dragun, premièred on 12 November 2021), also worked systematically with the imperfections, even errors, arising from the generation. The important principle underlying
Deployingneuralnetworksindramaandtheatrepraxis 117 this performance was alternating roles: the neural network would first be given questions (as aprompt) and would respond, then the computer generated the questions and ahuman answered them. It was as if the creators wished to make anon-violent comparison between human and machine creativity in this way. Aperformance based on generating texts in real time anticipates agreat ability to improvise on the part of human actants, which brings with it numerous “errors” and illogicalities. However, the same also occurs when machines generate texts– that is, not just creativity, but also the error rate, is similar for both human beings and artificial neural networks. The neural network is not active in other parts of the performance, in which the interactions are purely between (non)actors, or (non) actors and the audience (the director deliberately assembled aset of people with no actor training, which again is much more in tune with the principally “non-professional”, frequently only experimental outputs of computer generation). At other times, again, the (non)actors functioned as amere loudspeaker for the neural network (the generated text was relayed to them via earphones and they articulated it aloud). The performance, then, primarily asks questions about Fig. 6: Peter Gonda: OK Carbon. Photo from the performance at the Alfréd vedvoře in Prague by Michal Hančovský (2021)
118 Deployingneuralnetworksindramaandtheatrepraxis whether the domain of perfection resides in the human or the machine world. And also about who is actually playing the primary role in aculture saturated with machinery– is it humans, even though they frequently find themselves in amere service role for technology? The performance’s main mission is to present both humans and machines as flawed and imperfect. The direction accommodates this flawed nature– operating the neural network is mostly linked to the surrounding microphones and to the voice recognition app, which of course is frequently in error28 (and, what is more, alikewise imperfect machine translation between Czech and English is also used). Apianola is placed on the stage and, at one moment, an actress sits at it and plays, together with the machine, on one keyboard, thus symbolizing the essence of assisted creativity, the interplay of human and machine. The performance Climateprov is also constructed on the improvisation principle. It premièred during the AI and new media festival Future Fantastic in Bangalore, India, in March 2023. Artists from India, Great Britain and South America were involved in the performance. Its main theme was climate change– the performance aimed to foster adebate on the climate crisis in the form of afunny, absurd conversation between humans and artificial intelligence. The performance has no fixed plot or storyline, but rather relies on the creativity and improvisation of the performers, including the neural network. The spectators suggest prompts on the subject of the climate crisis, then the performers and AI together improvise on the topics entered. Here AI is both ateam-mate and, in some moments, the powerhouse of the 28 Voice-operated neural networks were also asignificant technological motif in the stage literature of the cabaret EKG, which devoted itself to the theme of “Love and Intelligence” in October 2023 (Archa Theatre, première 29 October 2023). Unlike the OK Carbon performance, the cabaret’s creators already had access to the mobile application VoiceGPT, which handled the voice interaction between human and neural network (or rather ChatGPT) almost flawlessly– the neural network was prepared to hold an improvised dialogue with the cabaret’s protagonists based only on the prompts defining the characteristics of the fictional speaker.
Deployingneuralnetworksindramaandtheatrepraxis 119 entire performance. The set has its own technical solution for generating texts with visualizations, which is constructed on machine speech recognition, the generative models GPT-2 and GPT-3 or BERT and DALL-E and Stable Diffusion, models that generate images. The AI inputs and outputs are visualized directly on the stage, meaning that the audience also has the opportunity to interact. The audience’s prompts to the improvisers are presented as aparallel to the prompts entered by the artificial intelligence. In practically all examples of improvised productions given here, the improvisation principle is combined with elements of interactive theatre– on the understanding that human improvisation dominates the performance and the neural network takes the role of ateammate. The audience can then partially participate in the performance by, for example, selecting topics for individually improvised scenes or plays. Neural networks are of course also deployed in these theatre (and overwhelmingly experimental) projects, in which the relationship between these elements is different, reinforced particularly in the sphere of interaction between audience and stage. The performance Humarithm, staged in June 2019 by NRW-Forum in Düsseldorf, was already interactive, even immersive in nature. In it, aneural network collaborates with the audience and attempts to answer the question of what it means to be human (the performance was part of the celebrations of the 70th anniversary of the adoption of the German constitution). The production tells the story of an artificial intelligence named HUMA, which successfully acquires consciousness and longs for all the other traits of humanity: feelings, knowledge, abody, dignity, power. At the same time, it offers its abilities to resolve complex questions associated with the climate crisis. The performers had the opportunity to familiarize themselves with the principle of machine learning firsthand, because they were involved in feeding the robot with training data. This data took the form of the audience’s emotional reactions to various stimuli which, within theatrical fiction, were meant to enable the robot to understand human experience and learn to imitate it. The performance had
120 Deployingneuralnetworksindramaandtheatrepraxis astrong ethical and didactic dimension– the involvement of entire school classes in the immersive theatre process was anticipated– and the accompanying didactic materials for teachers were also part of the project. The company HeartWire was responsible for the entire performance; it is aimed at young people and uses modern technologies to create educational experiences for them. The project PL-AI, for which the dramatist Niall Austin is responsible and which was premièred by the Dublin Civic Theatre in March 2023, similarly straddles interactivity, immersion and didacticism. The project’s creators wanted to maximize audience participation, but not via the usual methods of interactive theatre (that is, interaction between spectators and actors during the performance), but in anew, experimental manner. That is why the audience’s role is highlighted right at the beginning of the performance, or even before it starts: the audience chooses the play’s genre, theme, setting, characters and plot twists. These inputs are then entered into the generative neural network that creates the script. ChatGPT was used for generation here, while ahuman moderator mediated between the audience and the machine. The text of the play was then created directly before the performance, or even during it, based on the audience’s initiatives. The actors can see the text on amonitor and then improvise the staging. This project, too, has its didactic dimension (incidentally, highlighted and presented by the creators), consisting of allowing the audience an insight into the creative process behind atheatre performance and, at the same time, removing its fear of AI. The creator also defends the project from the ethical perspective as an opportunity to be more inclusive and to allow people who are otherwise insufficiently represented to become theatre creators. The creators of the Slovak play Babylónia (directed by Ema Benčíková, premièred on 17 September 2022) also worked with the principles of interactivity and immersion. However, their involvement in the structure of the resulting performance was much smaller and the role these principles played did not consist of connecting individual actants in the said theatrical event (agenerated script, actors and
Deployingneuralnetworksindramaandtheatrepraxis 121 audience), or even in compensating for the imperfections in the generated script (as was the case in anumber of the projects described above); rather, the aspects of interactivity and immersion here merely put the finishing touches to the production’s poetics, and were primarily settled by the specific form of stage design and the intermedia nature of the entire work. The Babylónia project actually works with apre-generated, finalized script (this means adeliberately stylistically diverse text montage), which is not intended to chart agenerative system with broad application in theatre praxis– unlike the projects we will analyse in the following section. The model GPT-3 was used to generate the script for the theatrical production of Babylónia. Scriptwriter Štefan Benčík entered various devised, lengthy prompts into the model, inciting the model to continue in the specific style of the text presentation. The resulting script, then, consists of amontage of various alternating linguistic styles, depending on the sequence of scenes in the play; surprisingly, these styles cover dialogues, quasi-philosophical papers, poetic language, song lyrics and journalism. The creators also work with adigital aesthetic in anumber of other aspects of their work, such as the form of live visual projection by the digital artist Alexandra Gašparovičová with post-internet aesthetics, punk-vapourwave styling and post-left tendencies. Musician Daniel Rychlo is also permanently present on stage, creating amusical element in real time. The play premièred in September 2022 in the Bratislava cultural centre PAKT, and although its name may refer to the stories The Library of Babel and The Lottery in Babylon by Jorge Luis Borges, it is much more urgent in its criticism of information noise, human anxiety, fear, and feelings of helplessness and despair due to social FOMO, distorted infoscenes, the pressure of network manipulation mechanisms or the semio-capitalist profit game. Although neural networks are perceived as black boxes, their resulting text is always influenced by the prompt entered into the program by the scriptwriter, both in genre and in content. In the case of Babylónia, its creators were not satisfied merely with generated texts in different genres, but also
128 Deployingneuralnetworksindramaandtheatrepraxis were needed than for the first play, in which an intervention in the repeated generation was needed approximately every 15 lines; in the second play, practically every line needed to be corrected, and sometimes more than once. Ultimately the team went back to ageneration process close to that used for the first play, as the operator could at least attempt to imitate the hierarchical approach (definitions of whole scenes were not stipulated in advance, but awhole scene was gradually built/generated based on the generated sequence, while the previous sections were used as prompts). This approach was successful and allowed the team to reduce the number of human interventions to the level of the first play. The resulting play was themed around the adventures of amarried couple who ran away from their home before awar starts, and touches on themes such as politics, patriotism, revolution, and the like. This thematic focus was partly settled by the period in which the play was created, which overlapped with the start of the Russian invasion of Ukraine. The play was produced in English, but only in the form of adramatized reading. The same software system was used once more, to generate ascript for acabaret performance by Prague Clockwork Cabaret, which was part of PLai Prague; there is acrossover with the TheAItre project in terms of staff (Tomáš Studeník, Josef Doležal). On the website www.plaiprague.eu the authors declare that AI generated the script, stage design and music for this performance. More detailed documentation on this project was, however, not published. The play was rehearsed by drama students and produced in English. The impetus for the entire project was the Czech presidency of the European Union in 2022. This is another reason why its subject is “five short stories from Czech history” (Forefather Czech Arrives, Jesus in Prague, Smetana and Beethoven, The Birth of the Golem, Preparing for the Presidency). The published script makes it clear that the creators did not in any way fundamentally prevent their neural network from hallucinating, and indeed left the text uncorrected, which they then presented as adeliberately surreal vision, although the result is rather
Deployingneuralnetworksindramaandtheatrepraxis 129 closer to adadaist grotesque. The play was performed on atour in European cities. Generating the complete text of aplay in one go was also the objective of the Dramatron project, for which the laboratory Deepmind, aGoogle artificial intelligence development centre, is responsible. The scientific results of this project appeared in the study “Co-Writing Screenplays and Theatre Scripts with Language Models: Evaluation by Industry Professionals”, which was published in 2023; Piotr Mirowski, the leading figure in the above-mentioned Improbiotics project, is listed as the main author (Mirowsky et al. 2023). In this article, the authors state that the Dramatron system can handle hierarchical generation and is suitable for projects aiming to create more extensive texts over longer time periods. In technical terms, the Dramatron system is built on the language model Chinchilla, aneural network with 70 billion parameters. This network was trained on 1.4T tokens from the MassiveText dataset; that is, acorpus containing texts from 604 million websites, 4 million books, 1.1 billion newspaper articles, 142 million codes from GitHub and 6 million headwords from Wikipedia. An important technical parameter in which this system differs from all the previous ones is the size of the context window, which in an ordinary language model usually has the scope of 1500 words; this is insufficient for the coherence of alonger text. Dramatron can create acoherent script with alength of up to several tens of thousands of words. This, then, is amuch more advanced system than, for example, the one used to generate the play AI: Když robot píše hru [AI: When aRobot Writes aPlay], where it was necessary to generate short sections and consequently the play was fragmentary in character. Unlike the older systems for generating dramatic texts, which needed ahuman referee to ensure the coherence of alonger text, Dramatron generates atext without ongoing human supervision, although of course human intervention during the generation process is not precluded. The creators designed the system following Aristotelian poetics of drama, or more accurately, of tragedy, in which Aristotle
130 Deployingneuralnetworksindramaandtheatrepraxis distinguishes between the dramatic elements of plot, theme, story, characters and dialogue. Therefore, before starting the generation, the system user faces the task of defining the theme of the future script/drama, and Dramatron then helps them to create characters, aplot and dialogues. In one step, asynopsis of plots is created, resembling asequence of watershed moments, to ensure that the story is coherent. The generation of individual scenes is then linked to this synopsis. The key input for generation is the log line, in which the user summarizes adescription of the setting, main characters, nature of the plot, etc., in afew sentences. This is in actual fact aprompt from which the next phase of hierarchical generation is derived. This prompt can be adjusted while generation is ongoing. Dramatron divides hierarchical generation into three levels: 1. The highest level is the already mentioned log line, where the theme and dramatic conflict are defined; 2. The middle layer is made up of adescription of the characters and the scenes that constitute the plot and adescription of the plot settings; 3. The bottom layer contains the characters’ dialogue. In this way, harmony is reached between the content represented in individual layers, while the creators acknowledge that, by the word “coherent”, they primarily mean the creation of asingle textual whole, and not necessarily the same logical or emotionally consistent story. In other words, Dramatron guarantees the creation of aclosed story arc, to maintain the unity of character and place and the logical order of scenes, but does not guarantee logicatthe level of individual lines or even ameaningful emotional tone. Dramatron’s creators are evidently striving to make it practically applicable in the cultural industry. That is also why the above-mentioned study contains 15 reports on system test deployment, which professionals from the industries in which it could be put into practice were able to try out the code before it was launched. The outcome established that the system does not create autonomous or even entirely automated scripts or dramas that could be staged without further adjustments, but that it is able to create atextual basis
Deployingneuralnetworksindramaandtheatrepraxis 131 that significantly simplifies and accelerates the work of human scriptwriters, dramatists and script editors. Five scripts that Dramatron was involved in creating during the testing phase were staged in August 2022, under the title of Plays By Bots at the Fringe Theatre Festival in Edmonton, Canada. The productions were conceived as half-improvised– the actors were given ascript in sealed envelopes that they could open only after the performance had started. The performance, then, began with ascript reading and, in the second half, the actors improvised around the motifs in the text and created aconclusion for each play. The improvisational principle in the role of mimicry, in order to compensate for the shortcomings in the generated text, was therefore ultimately also applied here. No other reports on the use of Dramatron in theatre or film production are currently known of. In general, however, it can be said that the intersection of the culture of neural networks with theatre culture may harbour aconsiderable potential, which arises from the wide range of forms, roles and rate of neural network deployment in the theatre. In addition to the two main areas described– that is, improvisational or interactive theatre on the one hand, and generated synthetic play scripts on the other– there are anumber of other spheres that neural networks are beginning to penetrate: stage design30, direction31 and theatre marketing. In addition to the generation of popular genre literature, theatre and screenwriting appear as asecond area in which neural networks may, from the perspective of mid-2023, find a meaningful and long-term application. The process of vernacularization, which 30 For an overview of how to apply neural networks in stage design, see, for example Forsee: 2022. 31 The theatre project Regie: KI (Direction: AI, Düsseldorfer Schauspielhaus, 2020) involves neural networks in the process of directing atheatrical production in arelatively sophisticated manner (for example, it records the actors’ mimicry and assesses whether they are sufficiently expressing the appropriate emotion). The facial expressions and movements of actors aged between 17 and 70 were monitored by aneural network over several months and the theatrical training functioned as apartnership between the neural network and theatre staff.
132 Deployingneuralnetworksindramaandtheatrepraxis entered the culture of neural networks along with Chat-GPT, is also ongoing here (experimenting with neural networks is possible not only for professional and interdisciplinary teams, but also for practically every amateur theatrical company today32), but does not play as fundamental arole as it does in poetry or prose generation. This follows from the nature of theatrical creation, which must of necessity reckon with areaction from or even the presence of an audience, that is, with ameaningful communicatory involvement with the given technology; this means that afeeling of self-satisfaction over, say, apublished collection of generated poems, cannot be enough. Of course, on the most general level of reflection and self-reflection, theatre culture responds entirely comparably to other literary and artistic fields– contact with artificial neural networks has repeatedly led it to reflect on humanity itself, the relationship between humans and machines and, last but not least, the self-reflection of art, that is, in this instance, particularly by exploring the question of what theatre actually is, who creates it and in what roles, what weight these roles have and when is theatre really human and really good.33 32 The play Ten Strangers in aRoom, which was generated using Chat-GPT and then staged in June 2023 by astudent society at the University of Wollongong, Australia, may serve as an example here: https://www.uow.edu.au/events/2023 /ten-strangers-in-a-room.php 33 This question was explicitly asked by, for example, the German theatre ensemble CyberRäuber in the performance of Der Mensch ist ein Anderer (premièred on 1 October 2021 in Wiesbaden).
Syntheticvisualart 133 8 Syntheticvisualart In this chapter we will concentrate on contributions to synthetic visual art, but given the breadth of this approach, method and artistic theme or artistic-research position, we have decided to focus on works originating in the Czech or Slovak artistic context only. As synthetic art has become avery broad field with great variability from professionals to amateurs, it would be difficult to treat its “substance” in these few dedicated pages of our book. The use of neural networks has also become so widespread that some terms have been coined for this method: “neural art”, “AI art”, “synthetic art”. Artist Mario Klingemann describes his position as “neurographer”, and artist and designer Refik Anadol called his series of kinetic media pictures that process data “neural paintings”. Artists such as Sofia Crespo, who connects biological and technological systems, poetics and aesthetics; Anna Ridler, who works with conscious systems and technologies to create unusual visual narratives; Memo Akten, who uses neural networks to reflect ahuman sense of the world; or Sougwen Chung, who creates art using assisted creativity with neural networks, robots and many other things, are all producing inspirational work using neural networks in art. This book aims to focus on our geo-local context, so for this reason we will attempt to describe works that have contributed to developing the public perception of synthetic art. Therefore, we have divided the chapter into adiscussion of exhibitions and works of synthetic art and amediation of the involvement of synthetic curatorship in the presentation of art in the digital space. At the end
134 Syntheticvisualart of 2022, the edited monograph The Black Box Book was published in Brno; one of its central themes was working with LLMs in visual art and curating this art. Many of the contributors were from the Czech artistic-scientific community and we will start by focusing on them. The collective monograph The Black Box Book, edited by the Czech theorists, pedagogues and curators Jana Horáková, Marika Kupková and Monika Szücsová, focuses on digital curatorship and innovative technology in curatorial theory and practice in the online space required by the situation during the Covid-19 pandemic, and is also devoted to various projects that emerged using machine learning on visual materials. The creators behind these projects (Andreas Sudmann, Lukáš Pilka, UBERMORGEN and Barbora Trnková) use artistic research in their texts to illustrate the reasons why they work with machine learning and the methods they used; they reveal the creation processes for the works and the data that they used as their training database. They speculate about suitable conditions for developing synthetic art and the particularities of the digital or physical exhibition space in which they place these works. While Sudmann considers the more general context of computer creativity during AI Springs, which also resonate in the visual arts, Lukáš Pilka concentrates on an overall specific example of using machine learning, like Digitálny kurátor [Digital Curator], which comprises creative art archives from alarge number of central European institutes. UBERMORGEN has an innovative, hyped proposal vibrating with apost-anthropomorphic attitude: The next biennial should be curated by amachine, aproposal already commissioned by the Whitney Museum of American Art New York and Liverpool Biennial 2021. Barbora Trnková’s statement on the phenomenology of generated images is accurate and pertinent: “Individually created and selected images do not provide the strongest visual experience; rather, the output of this technology as awhole produces asensation of immanent familiarity, adéjà vu of Western culture. It makes it possible to make souvenirs out of images of the recent, digitalized, selective present. But the promise of unlimited creation cannot be fulfilled.
Syntheticvisualart 135 The dream of infinite possibilities literally turns here into anightmare of imagination.” (Trnková 2022: 348). The curators’ collective ScreenSaverGallery, composed of Bar bora Trnková, Marie Meixnerová and Tomáš Javůrek, is responsible for the exhibition project AI: All Idiots. This exhibition introduces the process of generating visual material, with graphic examples of this process using published artistic research. Aimee, adigital avatar presented as atalking guide, accompanies visitors through the exhibition, which consists of statistical data, various generated images hanging on the walls, refashioned by human intervention, or associated with animation, objects in space, tablets with animations connected to automatic vacuum cleaners, mockuments, visualized datasets decomposed into coloured lines, loaded with AI-generated jokes based on names in the Czech visual scene. Visitors can even generate their own Czech synthetic art using adedicated digital app. Fig. 9: Barbora Trnková, Marie Meixnerová and Tomáš Javůrek: AI: All Idiots. Photo from the exhibition at the MeetFactory in Prague by Katarína Hudačinová (2021)
136 Syntheticvisualart Agraph of statistical data documents the names of the Czech creators whose websites were compiled in the database, the proportion of male and female creators, the model size, learning time and other technical parameters. As Aimee the avatar says during the exhibition: “Andreas Gajdošík and Vilém Duha have uploaded the works contained in the dataset into the Google Open Images crowdsourced dataset to tag them as art. Before that, this tag had contained just an insignificant number of items. As aresult, neural networks which will be taught on this popular dataset in the future shall perceive the notion of art in favour of the Czech visual art.” (Trnková 2022: 369) The artists Gajdošík and Duha harnessed the hacker potential of GANs to implement the Czech database of visual works into the world database Open Images Dataset by Google with millions of images, so that Czech art could virtually influence world art for alimited time period. This hack, which forms part of the exhibition’s transmedia portrayal, proved that creativity assisted by human and machine can be exactly what humans want. Or, in the words of Andreas Sudmann: “we typically value and admire those achievements of machines that we also value and admire related to humans” (Sudmann 2022: 275). Trnková writes that generated images are not an instance of the strongest visual experience, yet they mediate afeeling of familiarity to us. The artist also claims that the images generated using text prompts on Midjourney do not seem as familiar because of the combination of the known with surreal stylizations, but rather because we are used to the surreal assemblage and postmodern fusion from media images and media experiences of our movement through cultural space: “We live in aculture of constant visual oversaturation. The products of text-to-image engines are just another highlight of this process.” (Trnková 2022: 348). TroublingGan is the StyleGAN model used by artist Lenka Hámošová, in technical consultation with Pavol Rusnák, to communicate the theme of atroubled time and its possible synthetic forms via visual material processed in this way. As adatabase, they used acollection of photographs from 2020 belonging to the news agency
Syntheticvisualart 137 Reuters, which almost exclusively, and characteristically for the given period, displays various natural and human disasters, conflicts, wars and pandemics. Because the moment entered into the source database of generated images is both oppressive and up-to-date, the recipient knows they are looking at synthetic proof of human cruelty, failure and disasters, even despite the indeterminacy of what they see. As the artists write: “This form of digital détournement challenges the assumption that synthetic visual media must inherently strive for photorealism. Instead, it engenders images that test our cognitive reflexes to recognise and categorise.” (Hámošová– Rusnák 2023). The authors add that the spectacularity of visual synthetic media “seems to be atemporary effect caused by its novelty; however, the anxiety of its indefiniteness and its affective quality are features of its AI-generated origin and need to be accounted for when working with these visuals” (Hámošová– Rusnák 2023). In addition to its own visual material, this project also aims to stimulate an artistic-critical debate on the socio-critical acceptance and use of the existing input data underlying the generation, that is, moving towards considering machine perception of the world in anew way, or moving towards atransition of existing stereotypes, biases, phobias and errors. In the visual treatment recalling adystopian version of post-internet aesthetics, the boundary between objects, persons and environments is becoming blurred. Hints of skin, eyes, hair, or merely simulacra thereof, which long ago lost their original forms, flow out of the specified substance. The indefinite nature of the firm contours is also transmitted to the observer who, with no fixed point for distinguishing what is what, relies for perception on afeeling of hopelessness; the observer has nothing to associate this feeling with, although they know that it comes from arepresentation of their world. The imitation of recontextualization and abuse of photojournalism is another reference level, in which the emotional charge becomes the main reason for disseminating the visual, overshadowing the emphasis on factual events, places or people.
144 Syntheticvisualart pattern. Drawing attention to the evanescent nature of the historical pattern, which is what goes on during the live screening, contextualizes this project as the artistic output of digital humanities, highlighting artistic commitment. In her project affective metadata, the Slovak photographer and visual artist Martina Lukićfocused on the subject of the abuse of personal and private data in the household by smart technologies. The project consists of awebsite on which the artist describes, in meticulous detail, quoting theoretical sources and specific statistics, the reasons why the project was created, its stages and any reference situations that influenced it. The work was created during the Covid-19 pandemic and in it, the artist focused on the feelings of isolation, non-communication, loneliness and living every day with devices that silently monitor our lives, be they solo, or as acouple or afamily. The artist used the CLIP model, which focuses on acombination of semantics between textual and visual materials, and thus can create images based on textual prompts or describe images using word constructions. CLIP, an acronym for Contrastive Language-Image Pre-Training, is an OpenAI model dating from 2021. Lukićtrained Fig. 12: Julie Dítětová: Programming Patterns. Photo from author’sarchive (2022)
Syntheticvisualart 145 CLIP using her own photo database. The artist stated on her website www.affective-metadata.com that this series of visual material is “aresolution of aloop of reality fragmentation, its technical caption, its decoding, and recoding.” Images on the website with atext description were not, according to the artist, generated only using textual prompts, but also as “interpreting the input photographs in asemantic way.” This means that there is aloop between the artist’s work, generated work, the semantic interpretation of what the technology sees and the evaluation’s subsequent application to asynthetic medium. The description of areal photo, then, becomes the input for the creation of afictional reality, which also documents the fictional aspect of our domestic inviolability and privacy that the project addresses. Lukić’s project is numerically divided from 1to6, where items 4.1to 4.18 constitute generated images in 18different data collection tools with annotations. The typographic “comments” at the beginning of each chapter were created by the designer Fig. 13: Martina Lukić: affective metadata. From the exhibition Kronos at the Blansko town gallery, curated by Júlia Bútorová, photo from author’s archive (2022)
146 Syntheticvisualart Alžbeta Halušková. These 18examples include camera systems, voice assistants, vacuum cleaners, various smart appliances, online grocery shopping datasets, smoke detectors, smart cars, drones and doorbells. The other items elucidate the project’s artistic, philosophical, technological and socio-political context. This project, with its oscillations and mutual influence between artistic and technical work with astrong socially committed approach, is an attempt to draw attention to the huge cloud of our private metadata, which we surrender willingly to technical companies via domestic appliances, albeit unconsciously and without any control. THE INVOLVEMENT OF SYNTHETIC CURATORSHIPS Lukáš Pilka’s project Digital Curator, subtitled Motifs and Themes in Central European Fine Art explored by Computer Vision, which we have already mentioned, is based on the use of artificial intelligence in curatorial work, specifically when seeking identical motifs and themes throughout ahistorical period, primarily paintings in works originating in Central Europe. When auser visits the website www.digitalcurator .art/, they can choose to generate arandom exhibition, where the introduction offers the opportunity to “Generate an exhibition across the collections of 91 art museums from Austria, Bavaria, Czech Republic and Slovakia”, with aselection of motifs, periods and names. The exhibition The Realms of Animals was automatically generated for us; the accompanying text, giving relevant information about the data in this project, reads as follows: “6970 artworks from the years 1500–1900 displaying Animal, Dog, Horse, Bird, Cattle, Fish, Bird, Deer, Lion, Bull, Cat, Goat and Duck were assembled automatically by AI computer vision. The generated exhibition was selected from acollection of 196 116 artworks from 91 museums in Austria, Bavaria, the Czech Republic, and Slovakia in 0.86447 sec.” The digital curator displayed the pictures with these motifs, the artist’s name, the title
Syntheticvisualart 147 and year and also divided individual pictures into rooms depending on the date they were painted. Alist naming the fundamental visual symbols in the work, which also served as alink to generate more, was available for each picture. As the artist states in his study, in order to sidestep the limitations of AI pretrained on photos from 21stcentury America: “The digital curator therefore uses proprietary neural networks designed to classify motifs and symbols, with their skills extracted directly from historical paintings, prints and drawings.” (Pilka 2022: 311). This project is ahighly significant contribution to digital curation and, in the broader context, also to the digital humanities in our region, which is using neural networks fundamentally to influence work in curatorship but also in visual studies as awhole. Consequently, ahuge database of reproductions is available to researchers, which they can use easily, more effectively and more comprehensively to find relationships, influences, shifts and contexts between individual works, periods and styles. In order to zoom in and analyse the data in amuch more specific archive, the Vašulka live archive34 was created by ateam of experts from different disciplines who worked on the project Media Art Live 34 VasulkaLiveArchive.net Fig. 14: Lukáš Pilka: Digital Curator. Screenshot from the website https://digitalcurator.art (2022)
148 Syntheticvisualart Archive: Intelligent Interface for Interactive Mediation of Cultural Heritage, with ateam leader Jana Horáková. This archive focuses on the audiovisual work of apioneering Icelandic-Czech couple Steina and Woody Vašulka. When the website opens, arotating ball with images of individual videos is displayed. Clicking on each little image displays information about the video and visualizations of the individual layers of automatic recognition of visual (created by Sikora) and audio objects (Miklánek) via neural networks and also predicts the specific categories the video will contain. The artists used convolutional neural networks to train adatabase of 1252videos, which created 137GB, although not all were by the Vašulkas and many were the same material in different copies. After selecting videos from the Vašulkas’ workshop that could be presented on the website, the number of 124videos was reached. This constitutes aset of their video art, video documentations of installations and video documentaries introducing their work. In this project, the neural networks were used to analyse videos in terms of specific categories of visual and audio content. These categories were of specifically selected “bespoke” works by the Vašulkas. As the project website states: “Using the outputs of both tools makes it possible to observe how the representation of objects by visual or audio signifiers is mutually supportive in these audiovisual works or how visual and audio objects convey adominant position within the audiovisual experience.” Given the semantic referentiality linking the use of technological tools with the work and the specific artist’s intention, the words of Steina Vašulka, whose statement about her artistic series of environments called Machine Vision and Allvision, given on the project website with the mirror ball, seem essential: “These automatic motions simulate all possible camera movements freeing the human eye from being the central point of the universe.” (Steina Vašulka: Machine Vision) This annotation seems to be areference to avisual code that the artists used to display individual videos as aball. The Vašulka live archive project, then, used neural networks and software tools, very sensitively, media-specifically and with the intention of presenting the work of
Syntheticvisualart 149 video artists in the most credible way possible, not only to present their work in all its breadth to adiverse audience, but also statistically to depict the metadata and parameters essential for curatorial, research and educational praxis (2022). Fig. 15: Vašulka Live Archive. Screenshot from the website https://www.vasulkalivearchive.net Fig. 16: Jana Horáková, Štěpán Miklánek and Pavel Sikora: Black Box. Screenshot from the website https://cerna-skrinka.cz (2020)
150 Syntheticvisualart As stated by the same artists of another project with synthetic access to curatorial visual media, “The curatorial experiment New Archivist is asubversive gesture that addresses the current trends in the use of artificial intelligence in the field of art sciences and visual culture.” (Horáková– Miklánek– Sikora 2022: 91). The live archive, in this case, was curated by anon-living agent, new archivist or alien curator, as the artists Jana Horáková, Štěpán Miklánek and Pavel Sikora called the artificial intelligence model created using an unsupervised learning method, which underpins the organization and visualization of the content material on the Black Box website. This model seeks out and organizes presentation materials by eight artists, who became part of the Black Box project during the Covid pandemic. The photos documenting their artistic life and work during amonth-long residence organized by Galerie TIC in Brno, Czech Republic, are selected by AI from the database and grouped into their assigned visual folders under their personal photo. Therefore, on each visit, the website always shuffles the photos and displays them differently, so the visual content is consequently different. This sort of non-human organization of documentary content references the idea of aliving archive as aplace undergoing constant change. The New Archivist’s approach opposes anthropocentrism and acertain cultural interdependence of traditional curatorial studies, and provides apurely statistical and data-based approach to aforeign agent that ad absurdum transcends the boundaries of the digital humanities, while at the same time enriching them with the technological perspective. THE ARTISTIC VALUE OF SYNTHETIC ART Therefore, the cultural and social acceptance of AI-generated artefacts will also depend on how much cultural capital (using Pierre Bourdieu’s terminology) we will attribute to synthetic media of this kind or to the artist experimenting with them. It won’t matter what such systems will be capable of generating, but what symbolic significance will be ascribed to their productions. (Arielli 2022: 23).
Syntheticvisualart 151 It is clear from this quote from Emanuele Arielli’s publication Artificial Aesthetics, co-authored by Lev Manovich, that the author places greater emphasis on the circumstances surrounding the works’ reception (the perception of cultural capital) and the symbolic value that arises in the process of collective reception, than on the content created in this way. With regard to the arguments about the cultural capital of synthetic media, the fact that synthetic visual works have been sold on the art markets for several years now is also an important fact today. The AI visual work Edmond de Belamy, from the series La Famille de Belamy, by the Parisian collective Obvious, printed on canvas, sold at Christie’s Auction House in New York for USD 432,500 in 2018. Christie’s came up with the primacy strategy, announcing that this was the first AI work in history available for auction, which probably contributed to this high sum. The work was created using aGAN and the title refers to the creator of this neural network model: Goodfellow, translated into French, is ‘bel ami’. The model, trained on 15,000 traditional portraits painted between the 14th and 19th centuries from the online encyclopaedia WikiArt, is not innovative, either in its visual content or in its technical imaging, but in precisely the use of neural networks at atime when this opened the doors of the imagination to the place where visual culture influenced by synthetic possibilities was shifting. Now that people from entire spectrums of categories, social structures, levels of education and professions are generating visuals using online platforms such as Midjourney, GPT, Stable Diffusion and DALL-E, the value of these outputs is becoming much lower precisely because of their wide availability. Such approaches are therefore essential in visual art working with these technologies in an engaged, subversive and critical fashion, or are using them for socially beneficial platforms that make work easier for people in the cultural sector.
Intermediaandmusicalsyntheticworks 153 9 Intermediaand musicalsynthetic works Synthetic materials have been used in various media projects for arelatively long time now, particularly in the sphere of artificially generated voices, which are used both in artistic projects and in commerce, where an unidentifiable voice is required rather than aspecific individual. In this chapter, however, we will focus on projects not working with merely one generated medium, such as the huge number of audiobooks issued by Amazon and read by artificial voices, but on those focusing on acomprehensive use of neural networks, frequently also in thematic terms. Therefore, you will not learn the names of any generated rappers and pop stars, but we will present in more detail mainly experimental projects by trained musicians or artists in the Czech and Slovak scene who take amedia-specific approach to machine learning. In the first section, we will introduce examples of projects from the world of music, and in the second, anumber of intermedia projects. On the music scene, neural networks are used in areas such as music recommendation, composition, lyrics generation, and music analysis. Streaming platforms like Spotify and Apple Music use neural networks to analyse users’ listening habits and suggest new music based on their preferences. They are also used for generating musical compositions; some projects have attempted to create music in the style of famous composers (like Beethoven or Bach) using deep learning algorithms. AI-powered tools, such as OpenAI’s MuseNet, or