Literature : John Cayley
|
AI-generated image / Department of Literature Literature : John Cayley When discussing literature in the early 1990s, Jacques Derrida was quick to remark, “The name ‘literature’ is a very recent invention.”1 Historically and politically located, therefore. Literature, particularly when it is taken to derive from or include oral linguistic aesthetic practices, has also been understood as an inevitable, evolved, ‘universal’ consequence of human creative activity, and even as constitutive of the ‘human.’ The argument of this essay addresses literature at a recent, still evolving, historical moment, but it attempts to critique our understanding of this historically situated conception by way of language philosophical principles that are ontological and based on the essayist’s experience and thinking as particular type of literary practitioner. I am asking, “What is literature?” but, more specifically, “how does literature distinguish itself now that Large Language Models (LLMs) have been provided with affordances that seem to render them linguistically competent and, some claim, ‘creative?’.2 I do so as a long-standing practitioner of language art with computation in networked and programmable media that is made and proposed as ‘literary.’ My argument is that the current, general understanding and appreciation of literature is overdetermined by what I call orthotextuality. The formation of this neologism is analogous to that of ‘orthography.’ It indicates ‘correct’ or ‘corrected’ textuality. When overinvested, orthotextuality may produce a pathology of literature that I will call textual idolatry, the belief that literary objects can be identified with texts. Large Language Models and their transactions with text-as-data represent a massive overinvestment in orthotextuality.3 For readers and writers, it is simple common sense to proceed as if literary objects were texts. In the everyday world of traditional literary studies, the effects of the kinds of misdirection that I will attempt to specify may figure like those of relativity with respect to Newtonian physics. At this historical moment, however, as the discourse and statements of LLMs and Generative Pre-trained Transformers (GPTs) balloon, these effects will generate more significant, serious, and unintended, consequences. Consider this invalid—idolatrous?—but compelling syllogism: Literature is text. Literature is language. Therefore, text is language. The cultural power of literature falsely distributes the faculty of language to text. One unintended consequence might be expressed in a valid syllogism that takes the previous invalid conclusion as one of its premises: Text is language. The output of LLMs is text. Therefore, the output of LLMs is language. Few literary critics or theorists would confess to textual idolatry without first examining what—particularly here—is meant by ‘text.’ There is the further complication that I have invoked through the work of Derrida. If there is nothing outside ‘the text’ how does it not encompass language and especially literature? Writing—including literary writing—and other practices of literature (including literary criticism) deeply complicate both a simple statement of my argument and the bald reiteration of Derrida’s intervention. However, the historical situation has changed since Derrida challenged our grammatological understanding. The predominance of a contrasting scientistic conception of text—text-as-language that matters in the form of ‘data’ and text that may even be taken to ‘think’ or ‘create’—now either promises or threatens to overwhelm both recent and traditional institutions founded on linguistic media, along with their co-dependent institutions of education, particularly humanistic literary education. A scientistic conception of text insinuated itself as the regime of computation developed in the post-war period, and this development occurred almost simultaneously with the elaboration of the poststructuralist thinking that I consider to be essential for its critique.3 Now that ‘data’ is ‘big’ and demonstrably effective in society, indeed, ‘powerful’ in both the political and the technological senses of the word, text is accepted as data without question. To the contrary, we must continue to ask, What does the world give us as data with respect to language? Does the world, as we experience it, give us text? Is text really any kind of privileged data with respect to language? With respect to thought? And to creativity? My answers are that it is not text—not only or simply text in this scientistic sense—that the world of human experience and its faculty of language offers to us. Text is not necessarily data for language-as-such. There is no such thing as raw data.4 All so-called data is capta—what we can capture of what the world gives.5 This is particularly true of literary text, which is explicitly cultivated and curated, edited and even ‘cleansed’ into literature by its makers and its associated disciplines. Moreover, the poststructuralist critique of linguistics was precisely the thinking that showed us that all traces of language are a ‘dangerous supplement.’ If they are all we have of what language is, the cultural and political perils that this circumstance entails are not thereby diminished. Taking up the question of what we mean by text, I turn to Garrett Stewart’s recent, Book, Text, Medium (2021).6 Text is instantiated by material and perceptible traces, typically and predominantly typography. The essential quality of text is, however, human readability. Although the material support of text must be visible and present for us—‘in’ the book, ‘as’ typography, ‘as’ printed letters—it is non-identical with any of their distinct materialities. We scan and read traces which are tightly integrated but non-identical with the text. Rather, once grasped and read, the text is something that establishes a relationship with language-as-such, with the world of whatever is sayable.7 To be readable, the text must relate significantly and affectively to all those practices—many of them distinctly non-textual, let alone non-typographic—which constitute the human faculty of language. Text is a medium, as Stewart puts it, between typographic materialities and language-as-such. It is something that relates, separately, to these materialities on the one hand and to language on the other. And just as text is non-identical with its support media, language-as-such is non-identical with text. In analogous relations whose theorization was foundational for the science of linguistics, Ferdinand de Saussure’s ‘acoustic images’—material forms produced by our vocal organs and perceived by their aural counterparts—maintain, philosophically, the same relations both to ‘something readable’ and hence to language-as-such. In this thinking, the ‘something readable’—what we grasp and understand in the aurality within which we speak and hear—is also what I understand as text.8 Text in this strong sense, understood as a readable thing distinct both from its material support and from language-as-such (for which text is a medium in Stewart’s account), is not, however, orthotext. In the delivery medium of aurality, the material elements of what is read—from text into language—are embodied vocal gestures, and these gestures are clearly materially distinct from the text they support. It is perfectly possible to maintain, however, that graphically represented text in this strong sense is constitutive of literature, substantively and/or essentially.9 Regardless, language-as-such, rather than text in any sense, is the proper medium of artists who make literature. Humanly readable graphic texts as we know them provide us with objects that integrate, regularly and functionally, with language-as-such, and those subsets of linguistic artifacts that we consider to be literature are composed of such humanly readable textual objects. Simultaneously, in the material forms integrated with such objects—whose support media typical present themselves as an empirically accessible, and now digitized, typography—graphic text also offers itself as an object of contemporary computer science, something that can be studied, theorized, and known according to computational methods. The poststructuralist critique of structural linguistics remains crucial at this point because structural linguistics still provides the predominant scientific paradigms for the study of language, at least within those sciences most associated with computation. Despite claiming, in its founding gestures, to take speech—our species-universal linguistic behavior in aurality—as its primary object, linguistics based its science on differences which can be textually transcribed, and with various conceptions of ‘correctness’ explicitly at stake. Linguistics did, and still does, transcribe these differences, and then takes them in empirical evidence, as linguistic data. Derrida understood that structural analyses could thus only ever be referred to writing.10 He went on, however, to show that all linguistic practice had the same ‘logic of the supplement,’ but that there were differances (which I tend to think of, in a simplified grammatology, as language-constitutive differences) that might be readable-as-inscribed but materially absent in the terms or tokens of one or other transcription. The ‘silently’ irresolvable ‘a’ of differance, and all the ‘invisible’ puns, ellipses, or folk- and attested etymologies—insofar as they are read and grasped—are enough proof of this concept.11 Events of language must be able to turn, ontologically, on differances, on creative, language-constitutive differences. This applies equally to the acoustic images of speech and the gestural images of signs in sign language, as it does to the graphic images of writing. In all cases of integrated transcription—which are always events of language whenever they are produced as utterance or writing—what Agamben calls ‘the sayable’ is at stake but, insofar as ‘what is sayable’ may go beyond what has been previously inscribed, constitutive difference must also and always be in play.12 The logic of the supplement prevails for any (arche-)writing, in any support medium, and resists all scientistic transcriptions of language-as-such when transcription is based solely on principles of encoding and formulation. For events of language, we will, therefore, have texts—plural and ‘strong’ in the sense developed here—in more than one support medium. Typically, we will have closely related texts of linguistic events in graphic and aural forms. When these texts are referred to the same linguistic object, my thinking now characterizes their typical relationship as one of integration, but integration without necessary or certain encodable identities, and allowing, always, for the emergence of constitutive differences.13 Garrett Stewart provides a methodology of reading that regularly recognizes, in the course of its hermeneutic practice, that every text is simultaneously a phonotext.14 Stewart’s readings insist that what the text evokes is always already an ‘evocalization,’ an encounter with what the text says to us, not only when read out loud but as evocalized in our interior silence, a phenomenon he calls ‘secondary vocality.’ This interplay of graphic and aural texts is integral—my own not Stewart’s characterization—to the reading of anything literary and, typically, manifests as a vital aspect of what we call style. To read as Stewart does, one must acknowledge that there is more to the text than is spelled out in its orthography, syntax, and punctuation. The text is more than its orthotext, essentially so. Moreover, the phonotext, as Stewart puts it, guarantees integral relationships between its graphic and aural text and both human embodiment and the world in which these bodies live.15 The ‘acoustic images’ and their secondary vocality are informed by voice, by voice that exists to address others—socially and politically, not just significantly, affectively, and aesthetically—in our shared world.16 These are relationships that have emerged and developed in both biological and historical time. Literature must be sayable.
This is from Russell Hoban’s Riddley Walker (originally published in 1980). Despite being written in a speculative dialect based on English, this novel won major awards for science fiction—the genre in which is it categorized—while also establishing itself as a significant literary contribution. Harold Bloom listed the book in the 20th-century ‘Chaotic Age’ section of his Western Canon, whatever one may make of this distinction.18 I am quoting Riddley Walker for a number of reasons: because it is a vivid demonstration of a (typographic) text that is not an orthotext; because it cannot be read without what Stewart calls evocalization; and, because it fabulizes a critique of the regime of computation which, given that it was first published in 1980, exhibits remarkable prescience, resonating with contemporary AI anxieties. To ‘work it a roun,’ I will be quoting from the continuation of this passage three more times. A surprising proportion of this text from Riddley’s post-apocalyptic world is in fact straightforward orthotextual English, but there is enough in the way of unorthodox spelling, grammar, and word segmentation for this text to presage consternation if not havoc were it to encounter any kind of statistical, digital humanities, distant or machine reading.19 For the human reader it is no more difficult to construe than any other mannered or highly wrought prose. For many readers, including myself, this is a text that gives pleasure, and not only pleasure but the same gifts of affect and significance that can be found in novels more conventionally inscribed. Where would Riddley Walker’s language sit within the Large Language Models? How and how frequently would user-readers be able to prompt a GPT to generate plausible, commensurate texts in this or a similar style? I do not propose to try and answer these questions with the kind of research and experimentation that would be required.20 In part, this is because I believe that the underlying answers are intuitively obvious. The LLMs are trained or are being trained on orthotext and, even if provoked or constrained to do otherwise, they will tend to generate orthotext. Riddley Walker is an outlier that we, as human readers can recognize as literature, but this is literature that will barely perturb the vast vector spaces of an LLM. The content and context of this text may also prompt the question, “Was the disaster that destroyed Riddley’s civilization in any way attributable to an inability of the ‘box’ to input and read its ‘datter’ as anything other than numbers and orthotextual tokens?” Since, in the domain of language-as-such, our current systems cannot (or if they can, we cannot be certain of this), I suggest that this is evidence that the medium-term consequences of orthotextuality threaten to become dire, perhaps even existential for literature as conventionally configured and humanly read. Will literature survive the continued over-valorization of reading and writing held captive in the LLMs’ orthotextual gravity? Will reading and writing adapt themselves to digitally formulated techniques that take orthotext as their object, unproblematically and unscientifically? What are the LLMs? Here’s a brief fable of my own to begin an answer. In 1980, Riddley Walker’s ‘box’ represents computation at a time when it was already economically and socially significant (and, clearly, perceived as threatening) but computation had only just become ‘personal computing,’ and this was a good while before computation became ubiquitous as all kinds of networked devices. And yet Hoban’s fable lights on a perennial problem. Why do we need computation, and what will it do to us? The ‘box’ is recognizable to us as some kind of hardware, but, in the world of its fable, it is still a strange, unfamiliar, peripheral device. Today, such hardware is everywhere, because we have been convinced to invest in an infrastructure for computation that has proved so useful and ‘powerful’ that most of us will not live our daily lives without it. Meanwhile, the truly powerful computation has migrated from ‘our’ hardware to ‘their’ ‘cloud.’ The cloud extends on its inter-networks to the terminal points of our devices, while dwelling within a more centrally owned infrastructure that was built with previously accumulated capital but also, crucially, with profits (‘theirs’ not ‘ours’) from our own investments: earnings invested in ‘our’ infrastructural devices, and our personal data extracted and invested through our devices into the cloud. Our ‘return’ for these investments is little more than convenience, ‘increased productivity,’ and entertainment; very, very little in the way of shared profits or actual ownership of any means. This brief sketch lays groundwork for a discussion of the contemporary reconfiguration of computation. Computation became software during the earlier part of this tale. Now, in my fable, it becomes symbiotic if not synonymous with AI when this means Large Models and their Generators. You can feel this happening for yourself. The infrastructure in which you have invested is bringing you into daily contact with the new reconfiguration of computation in every service that you use. Simply choose to activate any or all of those new, ‘beta’ AI ‘assistants.’ They’ve been running the show for longer than you think. For more on what LLMs are, beyond their role in my speculative reconfiguration of computation (and in less fabulist terms), it might be better to turn to other sources, rather than this essay on the concept of literature now that LLMs are with us.21 It is possible to present serious critiques of this technocultural formation—and an appropriate sense of its potential to deform key political concepts—without having a detailed understanding of the ‘inner workings’ of its engines. We should know, however, that Large Models (LMs, which are not necessarily ‘Language Models’) are one outcome of a long running discourse of ‘data practices,’ as Adrian Mackenzie puts it.22 A ‘model’ is constructed by being ‘trained’ on huge data sets which are entirely abstract with respect to the operations that are performed on the data to build the models, although the selection and software-structural engineering of these operations may be tailored to accommodate both the source of the data and anticipated outputs. At the heart of these models are vector spaces with data points composited into correlated vectors plotted on one or other dimensions of the space. Linear algebra operates on these vector matrices and induces them to generate other vectors and dimensions which are derived from the original data but have no necessary or engineer-anticipated correlate in the world (ours) from which the original data came. These correlations are generated; and although they are intended—and are proven in practice—to influence and ‘improve’ output, they are not necessarily understood—either semantically or in terms of their actual effects—by the engineers or by us. Nor are they directly relatable to our world. When big data and serious computing hardware power is in place, the number of such generated dimensions is staggering, in the billions and beyond, as now often reported. The incomprehensibility and non-human scale of the vector spaces is the chief reason that we may reasonably, humanly say that the inner operation of the Large Models, from ‘prompt’ to ‘output’ is hermetically closed to us even though most of us encounter and relate to their separately coded interfaces daily. The general function of General Pretrained Transformers (GPTs), including those with a ‘Chat’ interface, can, however, be characterized. They evaluate, classify, categorize, decide. Their primary purpose, once given a string of prompt tokens, is to predict, based on their hermetic, ‘highly dimensional’ and staggeringly complex vector spaces, the next ‘best’ token. When referred to ‘language,’ the tokens in question are letters and/or words which are delineated in terms of orthotext. ‘Best’ here is a function of training, which may or may not be guided in the training phase by human evaluations, classification, categorizations, and decisions. Some randomness is usually introduced into the operation of models and their interfaces. In principle, a random allowance could be made for any ‘decision’ between virtual data points in any of the vector space’s dimensions. And possible output tokens may be randomly selected within a specified range of closely weighted options. The interface of the GPT will determine when and why the prompted output string concludes. Since language-as-(ortho)text and language-as-such are at stake for those ‘models’ which are qualified as linguistic, I have been fleshing out certain literary aspects of one particular critique of these models, which I will call the critique from linguistic ontology. Strictly speaking, this critique applies only to the Language Models but the structure of the underlying problem suggests the same kind of ontological critique for all of them. The models are trained on data—which is never raw—and the crucial question remains unanswered, that is, the question of how and whether this data is integrated or not with any embodied world that the models have been built to make decisions about or even to control. In the case of language, it is my personal position that this crucial question may be unanswerable. Language and literature may not be computable. Besides the critique from ontology—the problem of integration between purported data and its object—I single out three other rubrics for the critique of Large Models: critiques from bias, ownership, and hermeticism. The critique from bias is receiving a great deal of attention in the burgeoning academic literature addressed to the LMs.23 Rather that the relationship between data and its object, problems of bias arise from social, economic, cultural, and, indeed, political dispositions within the ‘object’ being modeled when the composition of this object—typically human populations and their cultural productions—is a mass of individuals or individual artifacts. The data sets will bias any model if they contain more data points from one or other privileged subset of individuals. It is also important to bear in mind that bias applies locally to subsets of individuals within a particular culture sphere (variously defined), and globally, between and across culture spheres. For example, at the local level, the linguistic practices of certain subsets of English speakers will be better represented in a model rather than others; at the same time, on a global level, English linguistic practices will be better represented than those of any other ‘culture sphere,’ here defined in terms of natural languages.
The critique from ownership concerns the question of who owns and benefits from both the models’ input data and their output responses. This is a highly complex and contentious topic, which deserves to be at the sharp end of activist critique while being blunted by its enclosure within established, ideologically implicated law, copyright in particular. Who owns literature? Copyright recognizes the legal and moral rights of individuals with respect to created, not just published, linguistic artifacts, but the question of what constitutes such an artifact remains. A short lyric poem or even the fifteen syllables of a haiku may be accepted as a work of literature. Then why not fifteen unique syllables of my writing on the internet, in any form, if I consider them to be a created work that is proper to myself? Even if these syllables have no recuperable value on the basis of my legal rights, do I not retain my moral rights of association (my name is to be associated with them) and integrity (they should not be used or rearranged without my say-so)? The questions may seem absurd but only because copyright is absurd, particularly with respect to linguistic or literary intellectual property (IP) in an age when data extraction and copy-making entail close-to-zero cost. Suffice it to say that, in line with many other domains of practice where IP is at issue, now that the LMs are operating at scale we are in a situation where, in exchange for convenience, ‘increased productivity,’ and entertainment, even recognized owners of significant artifacts are constrained to accept that the extraction of the data in which their work consists is ‘fair use’ (or something like it since copies are made in the extraction process) but that the output from the models are ‘original works,’ the rights to which are likely to be owned by some combination of prompt engineers and the holders of the models’ infrastructure. And the latter will likely employ most prompt engineers, co-opting their rights. The inequalities of late, vectoral capitalism will approach their apogee.
Although the critiques from bias and ownership will have a great bearing on literature in a world where computation is migrating to machine learned ‘AI,’ here I have concentrated on the critique from linguistic ontology. Earlier, we touched on the critique from hermeticism and have something more to say about it now that the models themselves have been briefly introduced. It has been often reported that the LMs operate like a ‘black box’ and although this epithet has also been applied to earlier instances of computational programming, the relatively common admission of Machine Learning engineers that they do not know in any detail how their engines work remains striking. The numbers applied to the models’ ‘highly dimensional’ vector spaces seem, categorically, to preclude human comprehension. This impression requires qualification in several respects. From a Critical Digital Humanities perspective, scholars like James E. Dobson rightly maintain that the models and their software are the product of explicit mathematical formulation and operations. Thus, despite scale, they are in principle knowable.26 Dobson rightly maintains that if there is a problem with the cultural, humanities, literary application of any digital methodology or generator, critical study must expose any problems by either historizing the entire enterprise and its discourses, or by critically examining theoretical and philosophical issues, as I attempt to do with my critique from linguistic ontology. One aspect of such an approach is taken up, for example, by Minh Hua and Rita Raley, who make an important distinction between ‘core deep learning code’ and ‘ancillary deep learning code,’ with respect to the models as they are embedded in distinct services.27 This is a situational as well as an historical critique addressing itself to our current moment. Finally—as relevant to the immediate context of this essay—computer scientists, linguists, and engineers are applying themselves to the analysis of existing language models and systems in an effort to shed light on the hermetic darkness within their ‘boxes,’ often with the intention of identifying, locating, and perhaps controlling, larger structures of correlated dimensions in the vector space, and then also relating these structures to the structures of, for example, new and existing grammatical or stylistic theories.28 The just or telling critique of a political re-conception of literature in cultures that come to be dominated by AI services will turn most immediately on specific critiques from bias and ownership. For literary practitioners and students of literature, however, the critiques from hermeticism and linguistic ontology cannot be bracketed, unless the disciplines associated with these roles abandon their own agendas or allow themselves to be reconfigured by more ‘powerful’ disciplines and discourses. Literary language art in networked and programmable media has worked with computation from its earliest history. The specialist writers who took up computation in the days of ‘personal computing’ did not pause to consider whether orthotext was language-as-such. Formal engagement with artistic media and, for that matter, chance operation was an aesthetically legitimate order of the day. The matter of language, however—syntax, diction, style, narrative—broke easily and quickly when subjected to the kind of computational affordances then available. Some language artists like myself persisted and found ways to critically and theoretically justify these self-described ‘experimental’ practices. The heuristic—as opposed to hermetic—aspects of this project were not as well appreciated at the time as they are now. Programming was not the professional discipline—particularly for artists—that it now is, and it hardly mattered that the experiments undertaken by language artists were, in principle, humanly knowable in their programmatic aspects. These experiments could be repeated or applied to other linguistic material. If aesthetic or other ‘discoveries’ emerged with respect to these programs’ linguistic or literary outcomes, they could be known and understood. These practices were, as I say, heuristic. And when these programs generated language, there was either an expectation that the language should be read as literature—sometimes as so-called ‘electronic literature’—or that the project should be read as a conceptual gesture, as conceptual literature, its concept knowable in a way that was guaranteed by its actual program, its code. None of these knowable, discoverable aspects are features of our transactions with computation-as-AI. Not yet and perhaps not ever. Hermeticism persists in this sense, in our inability—as makers or critics who are using AI—to know what we are doing with language when we transact with the LLMs, in our inability to know where the language being generated comes from, and whether or not it is, indeed, language-as-such, ontologically. Personally, I think I know that, in my practice of language art with computation, I am always working with orthotext and deforming it. But I am doing this in (sayable) dialogue with other existing literary practices, insisting that this is an experimental poetics, viable as such. I think I also know that, although literary practice will survive as long as the human faculty of language-as-such, specialist experimental practices such as mine will be swept away by the reconfiguration of computation. Already, even as I persist in programming my literary objects, AI assistants are at hand to do the coding better—in a way I may not understand—regardless of whether these assistants are entirely unable to access the language-as-such that is the actual medium of my project. I won’t be doing what I do by myself anymore. One argument of this essay is that linguistic and literary ontology is dependent on a practice of reading that integrates traces of language in more than one material support, in both graphic text and evocalized acoustic images. This integration currently eludes the orthotext in LLMs. Perhaps, in my own practice, I can write and code a dynamic text that puts my ‘integrationist’ thesis to a poetic test. I have tried to do this elsewhere and I will link to my project here.29 Sounding is a complex undertaking but, if nothing else, it is heuristic and therefore readably literary critical and in terms of Critical Code Studies, even if it may be difficult for traditional literary critical practices to accept or appreciate the project’s text-to-be-read as the poetry it intends. I offer it here in conclusion as a poetic claim for my argument that our concept of literature is perilously overdetermined by orthotextuality, with certain political consequences indicated above if not fully explored. The project makes its statements heuristically, with accessible code, reading out, in silent human speech rhythms, a dynamic, randomly phased and juxtaposed, but deterministic text. None of these characteristics, proposed as literary, are typical of transactions with Large Language Models, or the orthotext of their inputs and outputs. And yet, with transactive models and their assistants dispersed everywhere amongst our networked, menu-accessible, textually-friendly conveniences for writing, the models and their orthotexts are far more likely to overwhelm our near-future literature—in practice and conceptually—while sweeping away much of what we can know or feel or say about it. Within the models, we cease to know, what is literature?
Published on August 31, 2024 * John Cayley is Professor of Literary Arts at Brown University * 1. Derek Attridge, “‘This Strange Institution Called Literature’: An Interview with Jacques Derrida,” in Acts of Literature, ed. Derek Attridge (New York: Routledge, 1992), 40. This was Derrida’s response to a question about his earlier distinction between ‘literature’ and both ‘belles-lettres’ and ‘poetry.’ Nonetheless, it sharpens an important sense that any artifactual engagement with language is historically situated, contextualizing both Derrida’s and my own philosophical engagements with literature as concept.↩ 2. For a basic introduction to LLMs, see Stephen Wolfram, What Is ChatGPT Doing . . . and Why Does It Work? (Wolfram Media, 2023) and N. Katherine Hayles, “Inside the Mind of an AI: Materiality and the Crisis of Representation,” New Literary History 53:4 (2022): 635–66.↩ 3. This development is finely documented and interestingly analyzed in Lydia H. Liu, The Freudian Robot: Digital Media and the Future of the Unconscious (Chicago: University of Chicago Press, 2011). But Liu interprets these developments entirely differently. She sees the entry of alphabetic text, specifically and historically English text, into computation as a process that de-integrates its phonetic aspect, rendering it ‘postphonetic’ and, in Liu’s terms, ‘ideographic’ (ibid., 31–2). In computation and digital media, for Liu, alphabetic text becomes ‘Printed English’ (capitalized as such to distinguish it properly as a conceptual noun phrase) (ibid., 45–6). Liu’s Printed English is, by the end of The Freudian Robot, more or less identified with a kind of Derridean arche-text—the digitally mediated and abstracted traces of the linguistic unconscious—the appreciation of which might offer, in her terms, the potential to “constitute the first step toward coping with the techne of the unconscious in digital media” (ibid., 265). If Liu is right (and although her account is fascinating and deserves deep attention, I do not think that she is) then we might have to relate to the LLMs not only as potential literature but as a transactable collective unconscious—and to do so in Printed English. The very transcription of ‘English’ here evokes (or evocalizes, as I will discuss later in the essay) integrated embodied readings which are far from abstract.↩ 4. Lisa Gitelman and Vergina Jackson, “Raw Data” Is an Oxymoron (Cambridge: The MIT Press, 2013).↩ 5. Johanna Drucker, “Humanities Approaches to Graphical Display,” Digital Humanities Quarterly 5:1 (2011): 3.↩ 6. Garrett Stewart, Book, Text, Medium: Cross-Sectional Reading for a Digital Age (Cambridge: Cambridge University Press, 2021).↩ 7. Giorgio Agamben, What Is Philosophy?, trans. Lorenzo Chiesa (Stanford: Stanford University Press, 2018). My usage of ‘sayable’ here is based on Agamben and influenced by Garrett Stewart, Book, Text, Medium: Cross-Sectional Reading for a Digital Age (Cambridge: Cambridge University Press, 2021).↩ 8. John Cayley, ‘Grammalepsy: An Introduction,’ Electronic Book Review 08-05 (2018). Much of what I say here is derived from my own developing theory of linguistic ontology, one based, in part, on a reconfigured understanding of reading.↩ 9. This would resonate with arguments in Haun Saussy, “The Poetics of Text Editing [Die Poetik der Textedition],” Deutsche Vierteljahrsschrift für Literaturwissenschaft und Geistesgeschichte 97:1 (2023): 243–53. My arguments here, advising against what I call orthotextuality and textual idolatry, are not intended to denigrate or call into question the literary critical editor’s deep investment in identifying and understanding the text, and, indeed, expanding our idea of what the text is. Doing so, in Saussy’s terms, is the proper object of literary studies, so long as it retains the distinction between this object and that pertaining to the ‘history of ideas’ (Geistesgeschichte in Saussy’s article and its German literary context). My argument is against, as I am putting it, orthotextual practices: unconsciously standardized and data-‘cleansed,’—treated as if notions of ‘correctness’ guaranteed the text’s relationship to the medium of language art. I do take it as axiomatic that this medium is not text but language-as-such. An orthotextual relationship is, I am suggesting, too often, if not definitively, assumed and arguably adapted to by literary studies, with perilous consequences for the emerging cultural effects of machine learners ‘trained’ on text as such. Saussy’s evocation of poetic text editing takes the treatment of work by Dickinson and Hölderlin for its chief examples. Dickinson practiced a form of punctuation that would elude orthotextual consideration. And, as Saussy makes clear, both poets, if not all poets, resist any ‘being finished’ of the poem: another unresolvable problem for transcription. Saussy contrasts new ‘poetic’ editors who “return us to the moment of potentiality, when the poem is not yet finished, but realizes itself before our eyes, and when it is not yet anchored” and contrasts their disciplinary work with those who corrected errors and sometimes “[n]ot very long ago,” “considered not only what the author must have said, but what the author should have said” (252). Without abandoning a range of editorial investments in literary studies or the distinction from the history of ideas, Saussy acknowledges that, “An evolution in the sphere of ideas, of Geistesgeschichte, makes imaginable the text-as-event, the anti-final text” (ibid.). Firmly in the camp of the poetic text editors, I go further than Saussy in suggesting that all practices of language are unfinished events momentarily brought to resolution by human reading (‘reading’ in both linguistic aurality and visuality), and that stochastic simulations based on orthotextual so-called data (currently only ‘graphic’) are no substitute object for literary or any other language-dependent disciplines.↩ 10. See: Jacques Derrida, “Linguistics and Grammatology,” in Of Grammatology, trans. Gayatri Spivak (Baltimore: John Hopkins Univerisity Press, corrected ed. 1997), 27–73. “Yet the intention that institutes general linguistics as a science remains . . . within a contradiction. Its declared purpose . . . confirms . . . the subordination of grammatology, the historico-metaphysical reduction of writing to the rank of an instrument enslaved to a full and originary spoken language. But another gesture [… i.e. Derrida’s] liberates the future of a general grammatology of which linguistics [sic] phonology would be only a dependent and circumscribed area,” (ibid., 29–30).↩ 11. Jacques Derrida, “Différance,” in Margins of Philosophy, trans. Alan Bass (Chicago: University of Chicago Press, 1982), 3–27. See also: Geoffrey Bennington and Jacques Derrida, Jacques Derrida, trans. Geoffrey Bennington. (Chicago: University of Chicago Press, 1993), 70–84.↩ 12. Giorgio Agamben, “On the Sayable and the Idea,” in What Is Philosophy?, 35. Stewart also works with the essays in Agamben’s book, particularly in the last two sections of his Book, Text, Medium dedicated to the notion of ‘medium.’↩ 13. I owe this adoption of ‘integration’ and derivatives to the ‘integrational linguistics’ of Roy Harris, who offers, within that discipline, one of its more nuanced and coherent theories of writing (see, in particular, his Rethinking Writing [London: Athlone Press, 2000]).↩ 14. Garrett Stewart, Reading Voices: Literature and the Phonotext (Berkeley: University of California Press, 1990).↩ 15. See ibid., and also: David LaRocca and Garrett Stewart, eds., Attention Spans: Garrett Stewart, a Reader (New York: Bloomsbury Academic, 2024), 345: “phonotext: those inscribed phonic (if not fully sonic) cues emitted by the script of written text that narrow the gap between decipherment itself and imagined audition” (emphasis in original). ‘Imagined audition’ must necessarily be embodied.↩ 16. Voice is a difficult term, particularly in a post-Derridean context. I would gloss it as follows: what allows traces of language to be read into language-as-such, not from the timbre of authorial performance but from the corporeal fiber of these traces themselves. This is based on the gloss by Garett Stewart published in David LaRocca and Garrett Stewart, eds., Attention Spans, 347: “voice: what enters text not from the timbre of authorial speech but from the fiber of phonetic writing—and is then elicited in reading by secondary vocality” [emph. in original]. I believe it is important to acknowledge the evolved and typical integration of phonetic practices with human language-as-such while allowing ‘voice,’ as a critical concept, to comprehend traces of language in any humanly perceptible medium—as the ‘phonology’ of sign languages demonstrates. Sign languages are often analyzed in terms of ‘phonology’ within linguistic studies, at least since the work of scholars like William C. Stokoe. ↩ 17. Russell Hoban, Riddley Walker (Bloomington: Indiana University Press, 1998), 48.↩ 18. Harold Bloom, The Western Canon: The Books and School of the Ages (New York: Harcourt Brace, 1994).↩ 19. For a related discussion of non-orthotextual dialect in literature see John Cayley, “Differences that Make No Difference and Ambiguities That Do,” review of Martin Paul Eve, Close Reading with Computers: Textual Scholarship, Computational Formalism, and David Mitchell’s “Cloud Atlas” (Stanford: Stanford University Press, 2019), Novel: A Forum on Fiction 54: 2 (2021): 315–20.↩ 20. Even if I undertook this research, any results would apply to whichever models I used at the time; these would be swiftly replaced by one or more from a new generation of models.↩ 21. See Stephen Wolfram, What Is ChatGPT Doing . . . and Why Does It Work? and N. Katherine Hayles, “Inside the Mind of an AI.” For a Foucaldian analysis of Machine Learning, see Adrian Mackenzie, Machine Learners: Archaeology of a Data Practice (Cambridge: MIT Press, 2017). To contextualize what is now known as Critical AI Studies within the project of the Digital Humanities, see James E. Dobson, Critical Digital Humanities: The Search for a Methodology (Urbana IL: University of Illinois Press, 2019) and The Birth of Computer Vision (Minneapolis: University of Minnesota Press, 2023). For important contributions to Critical AI Studies, with atttention to the fields of literature and literary criticism, see Hayles, “Inside the Mind of an AI; ‘Subversion of the Human Aura: A Crisis in Representation,’ American Literature 95:2 (2023): 255–79; Minh Hua and Rita Raley, “Playing with Unicorns: AI Dungeon and Citizen NLP,” Digital Humanities Quarterly 14:4 (2020) and “How to Do Things with Deep Learning Code,” Digital Humanities Quarterly 17:2 (2023). For a commentary, within the field of electronic literature, on an early experiment in AI-generated poetry, see David Jhave Johnston, ReRites: Human + A.I. Poetry; Raw Output: A.I. Trained on Custom Poetry Corpus; Responses: 8 Essays About Poetry and A.I. (Montreal: Anteism, 2019). For some of my own contributions in the context of electronic literature and Digital Humanities applications to close critical reading, see John Cayley, “The Language That Machines Read,” in Attention à la Marche! Mind the Gap: Penser la Littérature électronique in a Digital Culture [Thinking Electronic Literature in a Digital Culture], ed. Bertrand Gervais and Sophie Marcotte (Montreal: Les Presses de l’Écureuil, 2020), 105–113; “Modelit: Eliterature à la (Language) Mode(l),” Electronic Book Review 07-02-2023 (2023), available at: https://electronicbookreview.com/essay/modelit-eliterature-a-la-language-model/; and “Differences That Make No Difference and Ambiguities That Do.” For a close reading pertinent to the constrasting concept of heuristic digital language art, see my contribution to Critical Code Studies: “Computational Art Explorations of Linguistic Possibility Spaces: Comparative Translingual Close Readings of Daniel C. Howe’s Automatype and Radical of the Vertical Heart 忄,” Digital Humanities Quarterly 17:2 (2023). For a recent survey of Deep (Machine) Learning from the perspective of Computer Science working with Linguistics, see Ellie Pavlick, ‘Semantic Structure in Deep Learning,’ Annual Review of Linguistics 8:23 (2022): 447–471.↩ 22. Adrian Mackenzie, Machine Learners: Archaeology of a Data Practice, 3.↩ 23. Within the field of what is now known as Critical AI Studies, this paper is usually cited as beginning to bring the LLMs to account,: Emily M. Bender et al., “On the Dangers of Stochastic Parrots: Can Language Models Be Too Big?,” in FAccT ’21: Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency; March 3-10, 2021 (New York: Association for Computing Machinery, 2021), 610–623.↩ 24. Russell Hoban, Riddley Walker, 48–49. ↩ 25. Ibid., 49.↩ 26. James E. Dobson, Critical Digital Humanities: The Search for a Methodology, 8: “These components or atomized steps in a research method are black boxes because of the complexity of the algorithm or because we cannot fully account for why the algorithm made certain decisions. Despite the possibility of this uncertainty, questions about the assumptions and operation of such black boxes remain pressing and answerable.” See also the following statements: “[T]he constructed computational model determines the object of analysis” (138); and “The specific implementation of the computational methods . . . no doubt quickly change and evolve into different forms, but the hermeneutical orientation that appears alongside framing—inquiring about origins, looking over the shoulder, under the hood, investigating the scene of execution—these tools . . . are among the many enduring contributions of the humanities” (131, my emphasis).↩ 27. Minh Hua and Rita Raley, “How to Do Things with Deep Learning Code.” This distinction is not made as often or as clearly as it should be. It is particularly relevant for the critiques of LLMs from hermeticism and, especially, bias. Core code pertains to the software that trains or is incorporated into the models. Ancillary code pertains primarily to software constructing and operating interfaces between users and the model. These may be—separately—either Open Source or not. Bias will arise differently with respect to the Core as distinct from its Ancillaries. Bias in the Core is a function of the models’ relationship with its data, whereas any bias in the Ancillary code will be the product of direct human and corporate, or governmental, intervention. The Ancillary code implementing such intervention for large-scale, public-facing production systems is all but guaranteed to be proprietary and closed—hermetic in a different sense. Ergo, tech corporations will curate the bias of their systems, executively, with or without popular or governmental involvement or regulation.↩ 28. See, for example, Michael A. Lepori, Ellie Pavlick, Thomas Serre, “Uncovering Intermediate Variables in Transformers Using Circuit Probing,” arXiv:2311.04354 (2023); Ellie Pavlick, “Symbols and Grounding in Large Language Models,” Philosophical Transactions of the Royal Society A A 381.2251 (2023).↩ 29. John Cayley, Sounding, https://work.programmatology.com/sounding/.↩ 30. Ibid., 49.↩ |
