TY - BOOK
T1 - Machine Mimesis
T2 - Electronic Literature at the Intersection of Human and Computer Imitation
AU - Erslev, Malthe Stavning
PY - 2023
Y1 - 2023
N2 - Machine learning agents, and especially large language models such as OpenAI’s GPT models and Google’s Bard, have emerged in recent years as distinctly powerful computational systems that can generate text that is virtually indistinguishable from human-written text. These systems can generate poems, song lyrics, and stories that often emulate the styles of well-known authors and are generally recognizable as being of a literary kind. In short: they imitate human language with a hitherto unseen efficacy. The advent of large language models has triggered excitement, fear, and discussion across contexts – in the technology industry, the academy, popular culture, and public debate. Their unprecedented capability to imitate human writing seems to usher in a new situation, wherein questions of computational creativity, artificial intelligence, and the very concept of humanity seem crucial. In this situation, however, one question often goes unnoticed and unanswered: what does it even mean that machine learning imitates human writing? What does imitation do, what does it infer?In this dissertation, I take on the question of what imitation means in the context of machine learning. I focus on what it means to mimic in an age of machine learning, which is a subset of the larger question about what happens to literature and literary culture when the production of language, and especially linguistic imitation, has become algorithmified by machine learning. To that end, I anchor the notion of imitation in the concept of mimesis, which has roots in Ancient Greek philosophy and still today plays a central role in our conceptions of similarity, representation, and imitation. I primarily engage with literary, anthropological, and philosophical conceptions and accounts of mimesis from the past century, drawing extensively on the work of Walter Benjamin, Michael Taussig, Jane Bennett, Erich Auerbach, Olga Goriunova, Nidesh Lawtoo, among myriad others.As the main contribution of this dissertation, I coin and develop a notion of machine mimesis: an emerging literary practice that exists at the intersection of human and computational imitation. As a central feature, machine mimesis is based on human participation in the mimetic loop, which means that we are not just dealing with computers that seek to imitate human writing, but with humans that imitate computers’ imitation of themselves, which creates an at times dizzying recursivity that is the primary source of literary and political agency of machine mimesis. Machine-mimetic practices negotiate and, at times, resist the drive towards indistinguishability in machine learning by sustaining an imperfect, friction-filled, and poetic kind of mimetic practice that reworks the imitative logic of machine learning from within.The overarching argument of the dissertation is that machine mimesis opens up the imitative relation between humans and machine learning and enables difference, critique, reflection, and renegotiation. This stands in contrast to the otherwise dominant assumptions about machine learning and imitation alike, both of which are often considered unilateral forms of assimilation – concerned with optimizing indistinguishability. Against such optimization, machine mimesis illuminates trajectories by which we can learn to read, write, and live with machine learning through reciprocal imitation, allowing us a more prosaic relation to machine learning and allowing us to affect its propagation in digital culture.The illumination of machine mimesis necessitates the development of a methodology that is appropriate for its study. Since it exists at the intersection of human and computer imitation, machine mimesis possesses a complicated materiality, which consists just as much of imaginaries as of artificial neural networks, none of which usurp the other. The operations of imaginaries and neural networks are difficult to scrutinize, and their interrelation perhaps even more so, since the material contingencies of either cannot be located in source code but only traced as they unfold. Facing this challenge, I develop a methodology based on analytical insights, theoretical meandering, and mimetic practice.My work is situated in the field of electronic literature, which is concerned with documenting, theorizing, and experimenting with the overlap of literature and computation. Machine mimesis is accordingly characterized as a form of electronic literature. There is a rich history of engaging with text-generation technology in this field, upon which my work builds and expands. I contribute with an unprecedented depth of engagement with the concept of mimesis, a development of a new methodology to study machine learning in from a perspective of electronic literature, and a survey of a broad range of cases pertaining to machine mimesis.The argument unfolds in three parts:1. To the end of approaching the mimetic messiness of machine mimesis with some clarity, I dedicate Part 1 of the dissertation to the study of the human end of the machine-mimetic loop, bracketing the actual participation of machine learning in machine mimesis. I argue that the study of machine mimesis benefits from such bracketing, wherein the focus lies on human imitation of software that imitates humans, which I conceptualize as a practice of bot mimicry. Bot mimicry primarily occurs online, where users of large-scale platforms write in bot-esque styles as a kind of mimetic negotiation of the relationship between humans, data traces, and bots online. By removing machines from the mimetic equation, bot mimicry warrants the development of an understanding of the perception at play in machine mimesis and a view to its politics, both of which are helpful as entry points to a deeper inquiry into machine mimesis.2. Inviting the machines (back) into the mimetic loop, I dedicate the Part 2 to developing a view to the mode of representation that characterizes machine mimesis. Arguing that machine mimesis necessarily plays out on two levels of abstraction at once – the levels of output and data – I develop a poetics of misrepresentation. Working through such a poetics, machine mimesis both does and does not represent – it misrepresents. Misrepresentation is a crucial part of machine mimesis, which allows for critique and negotiation while also enabling machine mimesis to result in the making of new works of electronic literature. I trace the poetics of misrepresentation at the level of output, via a reconsideration of the Turing test, and at the level of data, via a reading of a machine-mimetic work of electronic literature.3. In an effort to uncover nuance and establish its applicability, I dedicate the Part 3 to tracing machine mimesis’ ways of interfacing with design, research, and literacy. First, I outline the relation between machine mimesis and interface design, arguing that the former works as a form of prototyping in the context of the latter. Second, taking my own practice-based knowledge development as the guiding thread, I argue that machine mimesis can work as an independent mode of knowledge development in digital culture. I outline a mimetic method in academic research and discuss machine-mimetic literacy in K-12 education, treating both as forms of knowledge development wherein machine mimesis plays a crucial role. Third, I argue that these methods and literacies enable alternative design trajectories.In sum, my work sheds light on the often unnoticed mimetic aspect of machine learning. Bot mimicry illuminates the perception and the politics of imitating machines that imitate humans; the poetics of misrepresentation develops an account for the mode of representation of machine mimesis as well as its poetic capacities; the consideration of machine mimesis vis-à-vis design, literacy, and research situates my work in a larger context and provides insight into how machine mimesis already propagates in digital culture.Refraining from focusing on intelligence or creativity, I zoom in on the conceptual assumptions that guide our usual ways of discussing and measuring these. In anchoring my study of machine learning-based imitation in the concept of mimesis, and by tracing it in and through electronic literature, I arrive at a critical view to the current state of machine learning in digital culture that bases itself on but does not limit itself to, a long, rich, and complex set of histories of literary mimesis and literary computation. This work is crucial to understanding the still-developing techno-cultural phenomenon of machine learning. This dissertation provides a theoretical, analytical, and practice-oriented framework through which we can trace, understand, and renegotiate the operation of machine learning, particularly large language models, in digital culture and beyond.
AB - Machine learning agents, and especially large language models such as OpenAI’s GPT models and Google’s Bard, have emerged in recent years as distinctly powerful computational systems that can generate text that is virtually indistinguishable from human-written text. These systems can generate poems, song lyrics, and stories that often emulate the styles of well-known authors and are generally recognizable as being of a literary kind. In short: they imitate human language with a hitherto unseen efficacy. The advent of large language models has triggered excitement, fear, and discussion across contexts – in the technology industry, the academy, popular culture, and public debate. Their unprecedented capability to imitate human writing seems to usher in a new situation, wherein questions of computational creativity, artificial intelligence, and the very concept of humanity seem crucial. In this situation, however, one question often goes unnoticed and unanswered: what does it even mean that machine learning imitates human writing? What does imitation do, what does it infer?In this dissertation, I take on the question of what imitation means in the context of machine learning. I focus on what it means to mimic in an age of machine learning, which is a subset of the larger question about what happens to literature and literary culture when the production of language, and especially linguistic imitation, has become algorithmified by machine learning. To that end, I anchor the notion of imitation in the concept of mimesis, which has roots in Ancient Greek philosophy and still today plays a central role in our conceptions of similarity, representation, and imitation. I primarily engage with literary, anthropological, and philosophical conceptions and accounts of mimesis from the past century, drawing extensively on the work of Walter Benjamin, Michael Taussig, Jane Bennett, Erich Auerbach, Olga Goriunova, Nidesh Lawtoo, among myriad others.As the main contribution of this dissertation, I coin and develop a notion of machine mimesis: an emerging literary practice that exists at the intersection of human and computational imitation. As a central feature, machine mimesis is based on human participation in the mimetic loop, which means that we are not just dealing with computers that seek to imitate human writing, but with humans that imitate computers’ imitation of themselves, which creates an at times dizzying recursivity that is the primary source of literary and political agency of machine mimesis. Machine-mimetic practices negotiate and, at times, resist the drive towards indistinguishability in machine learning by sustaining an imperfect, friction-filled, and poetic kind of mimetic practice that reworks the imitative logic of machine learning from within.The overarching argument of the dissertation is that machine mimesis opens up the imitative relation between humans and machine learning and enables difference, critique, reflection, and renegotiation. This stands in contrast to the otherwise dominant assumptions about machine learning and imitation alike, both of which are often considered unilateral forms of assimilation – concerned with optimizing indistinguishability. Against such optimization, machine mimesis illuminates trajectories by which we can learn to read, write, and live with machine learning through reciprocal imitation, allowing us a more prosaic relation to machine learning and allowing us to affect its propagation in digital culture.The illumination of machine mimesis necessitates the development of a methodology that is appropriate for its study. Since it exists at the intersection of human and computer imitation, machine mimesis possesses a complicated materiality, which consists just as much of imaginaries as of artificial neural networks, none of which usurp the other. The operations of imaginaries and neural networks are difficult to scrutinize, and their interrelation perhaps even more so, since the material contingencies of either cannot be located in source code but only traced as they unfold. Facing this challenge, I develop a methodology based on analytical insights, theoretical meandering, and mimetic practice.My work is situated in the field of electronic literature, which is concerned with documenting, theorizing, and experimenting with the overlap of literature and computation. Machine mimesis is accordingly characterized as a form of electronic literature. There is a rich history of engaging with text-generation technology in this field, upon which my work builds and expands. I contribute with an unprecedented depth of engagement with the concept of mimesis, a development of a new methodology to study machine learning in from a perspective of electronic literature, and a survey of a broad range of cases pertaining to machine mimesis.The argument unfolds in three parts:1. To the end of approaching the mimetic messiness of machine mimesis with some clarity, I dedicate Part 1 of the dissertation to the study of the human end of the machine-mimetic loop, bracketing the actual participation of machine learning in machine mimesis. I argue that the study of machine mimesis benefits from such bracketing, wherein the focus lies on human imitation of software that imitates humans, which I conceptualize as a practice of bot mimicry. Bot mimicry primarily occurs online, where users of large-scale platforms write in bot-esque styles as a kind of mimetic negotiation of the relationship between humans, data traces, and bots online. By removing machines from the mimetic equation, bot mimicry warrants the development of an understanding of the perception at play in machine mimesis and a view to its politics, both of which are helpful as entry points to a deeper inquiry into machine mimesis.2. Inviting the machines (back) into the mimetic loop, I dedicate the Part 2 to developing a view to the mode of representation that characterizes machine mimesis. Arguing that machine mimesis necessarily plays out on two levels of abstraction at once – the levels of output and data – I develop a poetics of misrepresentation. Working through such a poetics, machine mimesis both does and does not represent – it misrepresents. Misrepresentation is a crucial part of machine mimesis, which allows for critique and negotiation while also enabling machine mimesis to result in the making of new works of electronic literature. I trace the poetics of misrepresentation at the level of output, via a reconsideration of the Turing test, and at the level of data, via a reading of a machine-mimetic work of electronic literature.3. In an effort to uncover nuance and establish its applicability, I dedicate the Part 3 to tracing machine mimesis’ ways of interfacing with design, research, and literacy. First, I outline the relation between machine mimesis and interface design, arguing that the former works as a form of prototyping in the context of the latter. Second, taking my own practice-based knowledge development as the guiding thread, I argue that machine mimesis can work as an independent mode of knowledge development in digital culture. I outline a mimetic method in academic research and discuss machine-mimetic literacy in K-12 education, treating both as forms of knowledge development wherein machine mimesis plays a crucial role. Third, I argue that these methods and literacies enable alternative design trajectories.In sum, my work sheds light on the often unnoticed mimetic aspect of machine learning. Bot mimicry illuminates the perception and the politics of imitating machines that imitate humans; the poetics of misrepresentation develops an account for the mode of representation of machine mimesis as well as its poetic capacities; the consideration of machine mimesis vis-à-vis design, literacy, and research situates my work in a larger context and provides insight into how machine mimesis already propagates in digital culture.Refraining from focusing on intelligence or creativity, I zoom in on the conceptual assumptions that guide our usual ways of discussing and measuring these. In anchoring my study of machine learning-based imitation in the concept of mimesis, and by tracing it in and through electronic literature, I arrive at a critical view to the current state of machine learning in digital culture that bases itself on but does not limit itself to, a long, rich, and complex set of histories of literary mimesis and literary computation. This work is crucial to understanding the still-developing techno-cultural phenomenon of machine learning. This dissertation provides a theoretical, analytical, and practice-oriented framework through which we can trace, understand, and renegotiate the operation of machine learning, particularly large language models, in digital culture and beyond.
M3 - PhD thesis
BT - Machine Mimesis
PB - Aarhus Universitet
ER -