The anxiety over AI hinges upon questions of great significance in society. This significance is of two levels: The first being that it pertains to the shape, value, and meaning of signs. The second is how a person gains significance in this milieu.

It goes without saying that we have passed through the industrial revolution into an information economy. Yes, we still need resources and services to survive, but the circulation of these resources depends on information flows within networked spaces, directing logistical chains, speculating on risks and rewards through financial markets, and wrapping these commodities in various brands, lifestyle narratives, and marketing campaigns. In some instances, major portions of our lives are focused on the symbolic layer of culture, as we stream movies, watch commercials, listen to music, indulge in fashion, scroll social media, play games, etc. How the signs of the modern world are created, circulated, and interpreted is a matter of great importance to people today.

Within this generalized semiotic space, significance is a question of value. Notable signs are those which are worth noticing. And while significance matters as a strategic opportunity within the global currents of the attention economy, this value extends to our personal lives as well. Each of us would like to think that our lives matter. In tribal societies, perhaps this significance was easier to come by. In medieval cultures, perhaps theology could provide a sense of significance. Accompanying the rise of that nation-state, perhaps patriotism could furnish this sense of belonging. We can say pretty safely that in contemporary societies, people seek significance through participation in consumer culture and social media. I cannot speak for the efficacy of past regimes of personal evaluation, but I can say that the present scheme seems painfully inadequate.¹

This pain is not without its consequences. To many critics, this is by design, with cultivated insecurity feeding consumer binges and dopamine addictions designed to extract resources from vulnerable populations. Certainly, many opportunists are in the business of selling alternative methods of fulfillment, whether we are talking about self-help and fitness, medical and psychological fixes, spiritual remedies, pathways to professional success, obsessive subcultures, etc. Where St. Augustine of Hippo or John Paul Sartre saw the angst of being as a foundational aspect of the human condition that could be answered through a kind of acceptance (Augustine took heart in the idea that restlessness was proof of the soul’s yearning for God; Sartre argued that the ultimate understanding came from the idea that there was nothing to yearn for outside of existence). Under the conditions of universal restlessness and absent any widely shared narratives, we live like postmodern hunters and gatherers, foraging daily meaning. Deleuze and Guattari described this in the language of nomadology.

We could see this happening to us.² Nietzsche sees the end of coherence, providing philosophy’s last gasp in his effort to distill a common currency in “the Will to Power.” Perhaps the failure of Nietzsche was due to the ultimate ugliness of his implications. Though I prefer to think that the ugliness itself is proof of the existence of another, greater value. For though the conditions of our existence are framed by questions of power and our ability to exercise it without apology or sentimentality, the fact that we cannot accept this means we want for something else. And that desire to be defined by something other than possibility, points towards conceptual reality that exists beyond materiality. That people want impossible and irrational things is a condition of the human worth exploring. And the fact that someone might come to me and insist that I am wrong to desire something that is not materially possible means that they, too, participate in the same condition of desire. They are irrational to believe that they can control the thoughts of others.

This may sound like an absurd intellectual exercise, but it carries with it a point: Within the terrain of communication—where reason is expressed, where persuasion is practiced, and where conformity is performed—a great many subversions are exercised. We negotiate, we obfuscate, we omit, we lie, we escape. Within the massive apparatus of instrumental language, we generate a multitude of ambiguities and alternatives through poiesis. We bend our language to reserve space for that part of ourselves that wishes to stand apart from the systems that circumscribe us, we listen for whispers and reach for friends in the solitude, we construct fictions and imagine worlds. Following Michel de Certeau, we make do against the efforts to strip us down to our basics, in the hopes that we live lives of some significance, however small, against the grim prognosis that power over others is the only thing left.

What do I mean by “the shape, value, and meaning of signs”?

Reading and writing is not just something we do. It is something that we do within a system that was constructed over many years, by many different actors, under many different pressures, for many different reasons. To even describe it as “constructed” almost defies our understanding of the word, as the intentions driving its many manifestations do not adequately describe how it came to be. In media studies, we like to use metaphors like ecology and evolution to describe the process of its development over many generations, as if it were the product of a natural process. These metaphors evoke the way media changes over long arcs of time through the impersonal genius of its design of collective cognition: Media systems present a record of the past, shape and share thinking in the present, and project thinking into the future, determining subsequent iterations of the process. And when we argue over a change to this system, it is necessary, at the risk of belaboring a point, to sketch out as many features of this system as possible.

Vilem Flusser’s “Gesture of Writing” performs the kind of phenomenology required to think through the anxiety of the present moment. I apologize in advance for reproducing a very long quote in this text, but I urge you to enjoy the exquisite transcript (the source of which, is even more delightful, as it is a scanned image of a typed manuscript, which contains the corrections we used to see in manuscripts before the advent of the word processor):

But, as it always happens with phenomena covered by habit and more than habit, writing becomes almost mysterious, if we discover it by deliberate consideration. If we draw off the cover of habit and more than habit, which renders writing an obvious gesture taken at face value, it becomes a gesture of such a complexity that it defies description. I shall nonetheless attempt such a description. And I shall restrict it to alphabetical writing as it is being performed at present. To write, we need several things that are supplied by our culture. First, we need a blank surface, for instance a white leaf of paper. Second, we need an instrument with contains a matter that contrasts with the whiteness of the whiteness of the paper and which can put that matter on the paper surface, for instance a typewriter supplied with a ribbon. Third, we need the letters of the alphabet, which is the shape of the contrasting matter we want to put on the blank surface. These letters may be stored in our memory, or, as in the case of the typewriter, in the instrument itself. Fourth, we need to know the convention which gives a meaning to the letters, which is, in the case of our alphabet, a series of sounds of a spoken language. Fifth, we need to know the rules which order the letters into higher structures, what is called “orthography”= correct writing. Sixth, we need to know a language which can be signified by alphabetic letters. Seventh, we need to know the rules which order that language, what is called “grammar”. (Premises five, six and seven imply each other and cause theoretical and practical problems.) Eighth, we need an idea to be expressed in a language to be expressed in letters to be impressed on the surface. Ninth, we need a motive express that idea. Now all these premisses (sic) must be assembled if we are to write, but they are not all of the same ontological order. The typewriter is not the same sort of reality as is a spoken language or rule of grammar, let along an idea. There for writing is a gesture which goes on several ontological levels. (Flusser 2).

This delightful passage identifies for us things we scarcely think about when we set out to write something. But this exercise is not without purpose, for each of these pre-requisites has the potential to open up serious questions about what we mean when we talk about writing. He continues, asking the provocative question about the fidelity of the pencil vs. the typewriter as a writing tool. In one sense, a pencil is vastly more expressive, as using it is akin to drawing, with the prescribed letterforms bending this way and that as our personalities express our thoughts. However, if we view writing as primarily conceptual exercise organized around the clear expression of words, then the typewriter is the more faithful tool for rendering thought as it strips away the noise of drawing them. From here, he asks readers to consider those forms of writing which strive to be unconventional. Here the pencil is better for deviant forms of writing as it allows us to play all sorts of games with the page (consider illustrations or marginalia). The tricks we might play with a typewriter requires one to invent new methods of deviation that exploit the technical constraints and affordances of the tool (concrete poetry comes to mind). Next, Flusser points to facts about writing that exist outside of the preconditions laid out in his text, noting that the first writing technologies were not applicative, but subtractive! Early writing was performed scratching surfaces or impressing symbols into them.

As one might expect, he works through the other preconditions, supplying each with equal care. When he arrives at the seventh and eight preconditions, he explores the question of what we are “transcribing” when we write: Texts could be marks arranged on a page by accident. In other instances, people set out to transcribe or paraphrase a text (or talk) for some purpose. A writer could sit down with a predetermined task to achieve through writing, perhaps even holding an imaginary copy of the text in their head. A writer could sit down to express themselves with no set idea of where their writing is going. All these instances of text carry with them different meanings for us, many of which would not be considered writing at all. But this is not without a point, for Flusser points out that the culmination of this act of writing must be understood alongside all the other forms that resemble writing, but do not fulfill these criteria. And this writing exists within a cultural sphere that is flooded with many instances of what he calls “pseudo-writing” (Flusser 7). Consider for a minute our world today, with spam, junk mail, ads, transcripts, policy documents, terms of service, forms, derivative texts, plagiarized texts, instructions, logos, and all the other things we could decipher alphabetically, but which we struggle to consider writing in any meaningful way. Long before the advent of ChatGPT, he seemed to notice that we are already in a wash of textuality that does not fulfill the conditions of writing under that system.

And then he pivots to the question of significance, of who can write and what the implications are for those people, who within this system, consider themselves to be “writers”:

if no data like sheets of paper and typewriters are available, (for lack of time or other social and economic reasons), then the fact that there is something to express becomes ineffective. The result is a frustration, (a constant repression of an urge to express). Which may destroy a life project and lead, in extreme cases, to suicide. Because if it is a fact that there is something to express, to write becomes the central gesture of living. (Flusser 8)

In other words, in a world in which the system of writing exists as the chief means of expressing matters of significance, being denied said significance creates feelings of frustration. Furthermore, in a system where we are flooded with texts (“pseudo-writing”), the means by which this significance is registered cannot be reduced simply to pencil and paper, but through the production of significance (the sense that we could write something worth reading). Of course, Flusser notes the historical specificity of this regime and those enmeshed in it (and concludes with the lament that the system of writing that has “programmed” him as a writer will be replaced by some other system).

The key takeaway from this is that we must consider these systems and those who exist within them. For me, as I read Flusser’s writing, I cannot help but think about my own situation, working on a computer, having passed from books and journals, print indices, microfiche, and library stacks into the World Wide Web, digital databases, electronic literature, and the emerging formats of the 21st Century; I am a transitional animal. Like many of my generation, I experienced the sudden explosion of possibilities as the range of texts and publication formats experienced this rapid proliferation. At the same time, I am supplanted by a generation whose habits of reading and writing seem, at times, enormously frustrating. Like a schizophrenic, I gesture endlessly towards realities that others simply do not see, I conjure experiences that verge on meaninglessness to the students in my care, “Get yourself a cup of coffee, find a comfortable, quiet place out of the way, and spend your day buried in a book! Put your phones away and get a notebook and write your ideas down! You won’t regret it!” With kindness, they humor me, nodding gently, smiling at me as if I am a benevolent nuisance trapped in a web of delusions, as they carry on with conversations about one show or another, a strange social trend they experience online, or a weird social joke that I can scarcely understand. But do not mistake my nostalgia as a plea for my continued relevance or my openness to the possibilities of this new world as an effort to ingratiate myself to the new framework. The point is only that we consider the larger frameworks within which reading and writing occurs.

A Large Language Model, or LLM, does not read texts the way we do. An LLM learns through mass analysis of texts, without intention or meaning, as we understand it. Its ability to use language is based on a massive analysis of what words tend to go together. And because words mean something to us, its model generates meaning based on the probability that its reply correlates with the given prompt. In the same way that a calculator doesn’t “know” what 5 means, it doesn’t really matter, as long as it says 2+3=5, 1+4=5, 8-3=5, etc. A key innovation in AI was the development of the Transformer model, introduced by Google engineers in 2017, which proposed understanding a text as a set of correlated “tokens” in relation to each other, rather than in sequence. The Transformer model improved the speed and accuracy of machine reading, by breaking free from a linear approach to reading and generating text. The Transformer’s “attention-based approach” reads all words in a text simultaneously, rather than left to right, and models the relationships among words based on their relationships within a larger corpus of texts.

An LLM functions by breaking down human language according to modular sequences and assigning them tokens. A word or a fragment of a word is assigned an ID number in an extensive database. For example, the word “dog” in GPT-3 is a single token [9703] and the word “dogged” is represented by two “dog”[9703] and “ged” [2004] (Tokenizer). The model treats the word dog as a sequence of letters that can appear as both a standalone word and a portion of word. However, the word “hot dog” is “hot” [8940] and “ dog” [3290], treating the word dog as a suffix (as in hot dog, bulldog, wiener dog), and uses the same token in the phrase “tired as a dog” is [3290] (Tokenizer). Thus we can see that LLMs reads words differently than we do, assigning signifiers based on how they appear in the larger corpus. This meaning agnostic view simplifies them for machine readability according to its statistical modelling of how words appear, while preserving the range of meanings a morpheme can have within a language. A model can break a language down to a massive number of tokens (50k for GPT 3, 200k in the latest version), but it does not have a lexicon.

From here, the model uses mass analysis of texts, looking at how strings of tokens appear in existing texts (and the relative authority of those texts based on how widely they are referenced), building new sentences based on probabilistic approaches to which tokens are contextually correlated with other tokens, what words are likely to appear in a sentence. In this way, an LLM can do many things when prompted with natural language: It can respond to a question with answers that generally seem reasonable. It can perform tasks like comparing spreadsheets, building a bibliography, summarizing notes, etc. Furthermore, LLMs use visual images and sounds (which themselves) are broken into constituent tokens (based on patterns of pixels or spectrograms) to read or generate audiovisual information. This approach can also be used to read and create code for digital media, shifting their domain from print and into contemporary complex media environments. It is this dynamic plasticity that allows machine intelligence to bridge virtually any domain that can be rendered in data, using a highly abstracted approach to language to do things that are often scarcely comprehensible from any human perspective.

You don’t need to formally study linguistics to understand that these shared understandings of words correspond to concepts (often multiple concepts) in our minds. Nevertheless, the terminology from linguistics can help us. Phonetics, phonology, morphology are useful terms that help us talk about a word as a set of sounds or markings that correlate with meaning is a useful building block. Semantics helps us discuss the way that the understanding of a word relies upon where it falls in the sequence of words (“I see a dog” vs. “I will dog them until they reply”). Semantics gives us a label for the way that words relate to other words in close proximity (“I would like to order one dog with mustard”, “I brought my dog to the veterinarian”, “When he starts drinking, he’s a real dog”). Pragmatics explores the ways that words work in larger contexts where we can draw deeper inferences about what they mean. When applied at a massive scale, often trained on trillions of tokens, drawn from a comprehensive corpus of texts, these machines develop the high degree of fluency that we see in current models.

It creates records of the frequency with which tokens appear in relation to each other within existing texts, building new sentences based on probabilistic approaches to which tokens are contextually correlated with other tokens, where grammatical form is derived from the attention-based model. In this way, an LLM could produce answers to prompts that are mathematically deterministic, delivering in each instance the highest probable combination. On top of this, LLMs use stochastic sampling to introduce linguistic indeterminacy to a mathematically deterministic model. For instance, if a person is talking about a standing body of water, there are many synonyms that could be used: lake, pond, pool, sea, reservoir, ocean, etc., each with implications for related terms. Based on context, likely terms might be sea or ocean. In another case, reservoir or lake would be correct. In others, pond or pool might be better. An ocean would point to harbors, navies, merchant ships, beaches, sharks, etc. A pond would point to frogs, lily pads, geese, rowboats, ice skating, etc. This stochastic layer is more than a gimmick to create the appearance of voice, it fuzzes the output to generate variability within a probabilistic range, with linguistic choices creating alternatives within the ballpark of plausible responses. By introducing an indeterminacy modeled on diverse texts, this is a soft mode of debugging.

And while the actualization of these models represent a significant achievement in recent years, the idea is not exactly a new one. Positioned in 1945, at the dawn of the digital revolution, Vannevar Bush imagined the “memex,” which contains a comprehensive set of texts, which allow “associative indexing, the basic idea of which is a provision whereby any item may be caused at will to select immediately and automatically another.” Two decades later, Ted Nelson discussed the idea of “Xanadu,” proposing a file structure and set of protocols for writing and reading that would make these kinds of associated textual pathways navigable as hypertext. Ideas like memex and hypertext are generally considered foundational to the development of the internet and world wide web. As a standalone space for reading and writing, the web is a monumental achievement. But today we see that it was a waystation on the path to machine intelligence. Google’s rise as the dominant search engine drove a refinement into its effectiveness. Its utility required an ever greater ability to provide more than just explicit results in response to precisely formed queries, but relevant answers to a general public asking in natural language. In 2017, two decades since the company’s founding, Google introduced the Transformer (Vaswani, et al.), which provided the mechanism for breaking language into tokens and analyzing these tokens for their statistical relationships (rather than reading text in sequence), creating the foundation for the LLMs that we use today.

It is this basic understanding of language as a system that can be manipulated indifferently that drives the development of LLMs. If we can describe the way that language is formed, specifically written language, as it appears in the print tradition, it becomes possible for machines to generate meaningful strings of text in relation to the input we provide. In this way, machines are able to elaborate on what we say or write in a manner that is readable to us. To think of it differently, we have used language this way through analogue processes. For instance, when I encounter a word I do not recognize phonemically and cannot figure out semantically or pragmatically, I can open up a dictionary or encyclopedia, find where the word is indexed, and then read the definitions, along with semantic clues (parts of speech, synonyms/antonyms, etymological roots, and sample sentences), and then walk away with the ability to “know” what the word means.

Humans, particularly operating in the modern world, where formal institutions, the need for translation tools, and a concept of formal education, there is an antecedent to this approach to language. Assume, for instance, I am totally unfamiliar with the word dog, but otherwise have some grasp of the English language. I can search the Merriam-Webster’s Dictionary for the word “dog,” and find definitions that point to nouns—one for the animal (“Pet this dog”) and one for the metaphorical use for humans who behave like animals (“He is a real dog, the way he treats women”)—and in place of verbs like “to chase” or “to hound” (“Ever since I started working with LLMs, I have been dogged by a concern that my assumptions may be incorrect”). When I use a resource to decode language in this way, I proceed with an openness to finding its meaning, rather than an a priori understanding of what it does mean. I can zero in on an understanding by re-reading the original passage and closing the gap in understanding through semantic and pragmatic reckoning. In this way, language does proceed through a kind of rational calculation. Furthermore, there are contexts within which this rational calculation of language is adequate for the context of use. For instance, if I am reading the employee handbook for my workplace or trying to access my online banking portal, this kind of unambiguous understanding is perfect. Generally, this is the goal of literacy—to understand things the way we are supposed to. At its basic level, we train children to decode texts for comprehension (not invention) and this continues throughout our lives. Even in the humanities, the goal is to teach an approach to the text for encoding and decoding with maximum fidelity. Even when we train artists or lawyers to explore ambiguity, it is with the intention of production and opportunity. Though this approach to language as utility constitutes the totality of its institutional purpose in the world and possibly constitutes the majority of uses in everyday life, there is more to language than this (and I will come back to this later). This difference was critical to the development of Structuralist linguistics, which depends on the distinction between langue (which is the formal system of language) and parole (which means everyday usage, but generally consists of that which eludes the rationalist pretensions of linguistics).

In the context of the utility of language, the pursuit of machine intelligence is not a rupture of our linguistic system, it is an extension of its logic of a modern technocultural trajectory. The goal is to produce legible, meaningful content, rather than the writing of an individual human person. In the present moment, writing exists in a cybernetic context, and though it means many things to many people, the systemic orientation towards content is one of profound media agnosticism. Rather than provide the meticulous breakdown of pre-conditions modeled by Flusser, I will point towards larger frameworks within which writing exists.

Audiovisual media

Following the introduction of static, time-based, and interactive audio-visual formats for the transmission of content, a great deal of writing is read by production teams who translate a variety of texts into other formats for audiences (not strictly readers). Additionally, writers who operate in this environment are not necessarily informed by the alphabetic texts themselves, rather they are enmeshed in processes of formation that include these audiovisual forms. In other words, a person who writes for television and film, generally engages with these audio-visual texts in the formation of new works (and may work in or with a variety of production artists, as well). And following Stuart Hall, the creation of texts follows multiple phases of encoding and decoding, as the work winds its way through a landscape that includes several distinct phases that provide meaningful interpretations of the work. A writer could create a treatment, which in turn taken up by production staff, which is subsequently edited, then shifted to a marketing team, and finally arrives in view of an audience (which increasingly provides its own derivations and enframings of the work).

Education

Reading and writing is understood in a different way in the educational system. Here, “literacy” contains many of the vestigial associations identified by Flusser. Reading is considered a vital part of the educational process, in which students are taught to read and write for the purposes of learning and assessment. In this context, writing is considered for its humanistic associations, cultivating deep concentration and a sense of interiority that is considered the necessary pre-requisite for academic success in an environment that depends on books and learning. Additionally, reading and writing is often juxtaposed to audio-visual forms as a remedy for distraction and disorganization. This perspective is often informed by developmental psychology and popular anxiety. At the same time, the advances of psychology and social change have also created stress on this objective, as the core humanistic values which underpin it are also under immense critical pressure in these fields for their normative function within society. Closely related to this are the efforts to extend this ideal of literacy to the media sphere in an effort to create normative strategies for the interpretation of audio-visual culture that align with the priorities of the passing print regime (“media literacy”). At the same time, these priorities are also justified as a form of workforce preparation, with (“cultural”) literacy seen as a job skill that will preserve institutionally desirable norms that are necessary for success in college and later careers.

Corporate Communication

Office culture depends heavily on reading and writing. Handbooks, contracts, reports, memos, emails, etc. circulate internally. Workers are expected to read and produce massive amounts of organizational communication pertaining to individual behavior, company objectives, clients, contractors, and adjacent institutions. Additionally, these entities create massive amounts of outward-facing reading and writing (“storytelling”) for advertising, public relations, marketing, and consumer support. In many instances these institutions are also in the business of media products.

The Post-Digital Environment

A great deal of reading and writing takes place in and around digital networks. While, of course, this space includes all forms of communication referenced previously, networked culture contains unique features as well. Everyday people create content for the internet in a variety of publication formats. Websites, blogs, social media, customer reviews, text messaging, emails, and an unknowable number of invented forms create many occasions for people to engage in acts of writing. Many of these formats include novel elements like emojis and tags, which supply additional (and often explicitly machine-centered cues). And all of this content is driven by varying degrees of machine reading, translation, and coordination to simultaneously enrich databases for further analysis and generate framing contexts for human readers who participate in these networks. A significant share of this content is machine-generated itself (and, of course, code itself is textual), which is in turn used by human readers. Artificial Intelligence (specifically Large Language Models) uses mass analysis of text to train machines to interpret text and create data (all in the form of text), which can subsequently be turned into a variety of other media formats.

If these frameworks sound confusing, it is because they are. The only meaningful point to make is that though we spend a great deal of time talking about the impact of AI on reading and writing, we scarcely comprehend what we want from reading and writing. We might point to the act of an ideal writer (perhaps a novelist) writing an ideal text (a great work) being read by an ideal reader (a person with the appropriate level of literacy). But few, if any of us, live in this context. And fewer still live there exclusively. And even fewer still could provide a careful breakdown of the sort that Flusser is able to provide. But even if we could, we would not read this way. To turn back to de Certeau, reading is tactical, we make do, because we like it. And perhaps it is in the cognitive act of making do that we obtain the cognitive gains we see through formal literacy.

A more simple elucidation of this point is available in Heidegger’s description of “hammering”:

Hammering does not just have a knowledge of the useful character of the hammer; rather, it has appropriated this useful thing in the most adequate way possible. […] The less we stare at the thing called hammer, the more actively we use it, the more original our relation to it becomes and the more undisguisedly it is encountered as what it is, as a useful thing. The act of hammering itself discovers the “handiness” of the hammer. (Being and Time 69)

To apply it to the question of this essay, to understand the impact of AI on writing we have to consider the larger framework that contains reading and writing and Large Language Models. And to understand its impact, a key perspective would be to look at how it is used.

Language is fundamentally useful. Just as was the case during the rise of literacy, seemingly small shifts in practice have significant impacts. For instance, the introduction of movable type by Johannes Gutenberg in the mid 1400s (the notable first is his publication of the Latin Vulgate Bible in 1455) is widely regarded as a watershed movement in media history. However, within a few short years, the acceleration of printing was driven by corresponding increase in demand for texts. The increase in literacy was driven by the growth of the middle class in European cities (along with the growth of mercantile trade and manufacturing). This key turn to the vernacular was critical to the Protestant emphasis on personal interpretation of scripture as the key to salvation. Thus, liberated from the influence of the Catholic Church the governments of Europe formed national identities, which in turn organized the many regional dialects into national languages. In this way, the major languages were shaped by organized systems of education. Along with languages, literatures emerged as expressions of national identities. And while history is filled with a variety of coordinated efforts to shape education, elevate literature, and forge national identities, this process was not governed by any unified genius. It happened uniquely in many nations, across many timelines, implemented by many actors, and with varying degrees of efficacy. Nevertheless, they happen, through a discourse within which events reach their likely conclusion.

One temptation, of course, is to compartmentalize the present moment and situate it within the current context of LLMs. Proponents of this view argue convincingly that the recent breakthroughs in Machine Intelligence and their rapid adoption constitute a fundamental departure from how we use language. Whereas the 19th and 20th centuries were marked by a series of innovations that reframed communication in increments—the high speed printing press flooded the world with cheap texts, photography and image reproduction transformed semiotics, audiovisual technologies opened new pathways for propaganda and storytelling, desktop computers and the internet disrupted traditional bottlenecks in publication and dissemination, the rise of platforms alters the way we participate in media circulation, and so forth—AI intervenes as an interlocutor that is capable of transforming and modulating all of these activities in a more fundamental, if often inscrutable way. This argument is a compelling one. However, I prefer to cast a wider net, anchoring the current milieu to the deeper history of cybernetics, which begins with social theories that can be traced back to population thinking, a view of humans not as individuals, but as a sociological block. In the Smartness Mandate, Halpern and Mitchell trace population thinking to Thomas Malthus’ 1798 A Principle of Population. From here, they advance through the work of economist Friedrich Hayek and evolutionary biologist Ernst Mayr, in whose work they find the idea of a collective cognition that exhibits intelligence in collective action, an intelligence that is unable to “’learn’ in the traditional sense” or “consciously ‘know’ anything” (46). Halpern and Mitchell’s concept of smartness leads through the financial innovations of Black-Scholes and the development of derivatives markets which can transform all measurable phenomena into commodities.

This long transformation in thinking paves the way for the present moment of platforms that analyze, coordinate, and monetize everything. So while I agree that AI exemplifies a fundamental shift in the way we think about humanity, the model of intelligence, as de-individuated aggregrations of text that amount to meaningful signals provides the basis cybernetic systems following logical procedures take on a new form of intelligence when carried out at speed and scale. In other words, LLMs are fundamentally different, however these social, psychological, political, and cultural changes they represent are the fulfillment of many decades of systemic change. In this reckoning, LLMs are not the beginning of a change, but rather its realization. In this respect, we (the lingering remnants of humanism) are a bit like post-colonial subjects, after having our folkways disrupted, our resources stripped, our labor exploited, and finally finding ourselves expelled from the countryside, landing like refugees in Ellis Island. But the city we find ourselves in did not ultimately cause the abrupt alienation we experience upon finding ourselves in an alien land, rather it is only the logical conclusion of a process in which we are entangled. And, now, like the billions of immigrants displaced during the industrial revolution, with no home to return to (even in our own regions), we must make this one our own.

With the rise of cybernetics, it goes without saying that the evolution of language introduced this massive frame shift. Bernard Geoghegan’s Code provides a rich overview of the impact of cybernetics on culture during the 20th century. Starting with Modern technocratic endeavors of the asylum, colony, and camp and carving a path through the computational innovations of World War II to the postwar ascendancy of the United States Military-Industrial complex (aided by universities, foundations and think tanks), Geoghegan’s archaeology identifies the epistemic shift in cultural organization. In other words, “code, communication, computing, feedback, and control…embodied an effort to develop more enlightened analytics for the force wielded by science and the state” (2). With an emphasis on the impact of cybernetics on the University, Geoghegan links the history of computing (Weaver, Shannon, Bush, Wiener, and von Neumann) to social scientific and humanistic—media studies (Lazarsfeld and Lasswell), anthropology (Bateson and Mead), psychology (Kubie and Harrower), and literature (Richards and Ogden). The appeal of cybernetics was its potential to mediate diverse phenomena through manipulable symbols that could be measured, analyzed, and modelled impartially. As computation drives for a theory of language that is calculable, the stage is set for a variety of approaches that deliver us to the material and conceptual resources necessary for the implementation of Artificial Intelligence.

With this framework in mind, we can understand the long history of personal computing not as a series of discrete watershed moments in human history, but as a progression in the history of reading and writing. Just as the Gutenberg Press is an historical event in culture, it is better seen as a starting pistol that accelerates the rise of vernacular education and the emergence of National identities, myths of authorship, the romance of the individual, and other associated phenomena. This growing market for print accelerates the development of printing into the Industrial revolution, as print processes become faster and cheaper, and the content of printed matter becomes more ephemeral. It takes several centuries to see something resembling the Mass Culture we associate with the 19th and early 20th centuries. Our sense of what reading and writing is and what it ought to do is strongly shaped by a multigenerational process of technocultural development, and many of the common practices that we valorize as integral to the thinking and being are legacies that belong to a world that no longer exists.

Simply put, the tools of signification that reached their peak in the 20th century no longer provide the hope of significance that drove participation in these forms. Instead, we see the rise of categories of symbolic communication in two main directions: 1) Educational assessment and corporate communication transform writing into a kind of formulaic busywork, quantifiable evidence of activity produced through volume. 2) Audio-visual media and post-digital networked communication transform writing into a commodity form, quantifiable through engagement metrics. While both categories measure a kind of significance, the first being labor value and the second being exchange value, both seek to define human significance in economic terms. The cybernetic measures of value are totally disconnected from the ideal of personal significance that accompanied the rise of print literature. The dream of “greatness” has been thoroughly dehumanized, as none of these amount to virtue in modern humanism. Instead, we are stuck with billionaires, entertainers, managers, and politicians, our “heroes” are generally polarizing winners of the rat race. Downstream from this, the successful are generally frustrated and unhappy, and the losers are transformed into externalities (failures, victims, patients, or criminals). And stuck in between are a class of white collar (relatively affluent, bourgeoisie of the 4th Industrial Revolution) and service workers (less affluent transitional class, a kind of petty bourgeoisie in this schema) who manage this massive chasm of desire between the significant and insignificant.

On the ownership of models

The major cause of concern is the fact that AI has been developed following the model of digital innovation. While much of the early progress of key inventions (mainframe computers, telecommunications networks, the Internet, satellites, etc.) occurred with the support of state (and often militarized) investments, development of these technologies into commercial, consumer, and public utilities have been taken on by corporations. In the case of AI, the massive coordination and investment has resulted in a situation in which many of the major models operate under murky designations as closed, open-weighted with closed software, and open source.³ The development of major models is complex because the cost of research, computational infrastructure, data, and training is significant. There is intense competition for performance for models that can be generally useful or specifically targeted. Furthermore, development often requires contracts and collaborations across many sectors, the navigation of many legal and technical constraints, with many companies offering specific services or innovations. And there are inherent risks associated with centralized control over AI technology (whether we are talking about state control or corporate monopoly). Frontier models offer significant competitive advantages in many industries, which motivates development of large models. Furthermore, the security risks that come with frontier AI are enormous. The risk proposition for a large private corporation provides a degree of leverage which would not exist if the models didn’t belong to anyone or were held by the state.

At the same time, there are a wide range of tactical approaches being deployed. Using what we have learned from LLMs. Many are experimenting with Small Language Models, which operate on user-owned machines (like laptops, smart phones, and local servers) and are trained on smaller sets of data. Others are exploring more targeted applications of proprietary tools like Retrieval Augmented Generation, which allows users to focus their results on data sets that are defined by the user, or Fine-Tuning, which allows users to constrain an LLM to work within a specific context. On top of this, many who are exploring these approaches use a hybrid approach that incorporates a variety of large and small models to exercise greater control of how and when machine intelligence is used with intention, in the same way that a way traditional researchers of the past may have tapped into disparate knowledge bases and communities to carry out complex projects.

While the prospect of situating this major engine of cultural development in a proprietary space is troubling, we can see similar examples of massive technocultural change throughout the history of the modern world. For instance, the development of global commerce during the age of exploration created the first chartered corporations, as Nation-states contracted with merchants operating under the flags of their respective sovereigns. This model continued to evolve with the development of the railroad industry, the development of energy industries and the electrical grid, the telephone, aviation, aerospace industries, pharmaceutical and medical research, and the early phases of the digital revolution. While each of these undertakings is associated with significant historical problems, most people experience the benefits of these undertakings without much consideration or awareness of the complex research, investments, and logistics involved in coordinating such massive systemic processes. To return to an earlier metaphor, we might consider ourselves subjects of a new kind of empire. But just as printing presses, libraries, and schools served the apparatus of empire, they were not the empire itself. We would be wise to consider LLMs from a similar perspective.

Indeed, even those who advocate for a return to a pre-digital print ideal, are largely unaware of the massive, multi-institutional conflict and coordination associated with creating the infrastructure needed to support literacy. The drive to establish literacy was driven to exercise power, create efficiency, and coordinate labor and resources. Nevertheless, we have used language across many generations and in multiple platforms to advance our rights, promote harmony, express ourselves, entertain each other, and create literature. While we can definitely attempt to learn from past mistakes with regards to other massive technical undertakings, we must also understand that the challenges and opportunities posed by this development might rhyme with the past only partially. To state it plainly, in a multipolar world with many actors in aggressive pursuit of automation, it would take an Act of God to stop it altogether.

In my capacity to act, the AI revolution is like the lesser invention of the telephone: I can choose how I use it (and in some cases how the people under my authority use it), but I am unlikely to stop it altogether (and, if I am being honest, as smart as I am, I am not sure that it would be right for me to even try). Following the metaphor of refugees of empire, we find ourselves as immigrants on an unfamiliar shore. The best bet is to try to make do as best I’m able in a world that I do not control, and to encourage others to do the same.

But there is another world…

Alongside this instrumental drive, there is a constant strain of conscious and unconscious poiesis, smuggling in human intentions, bending meaning, weaving metaphors, subverting authority. If AI is a logical step in the formalization of writing across the modern era (following Flusser’s comparison of the pencil to the typewriter), the digital generated a new lexicology and syntax, that frames a broader range of signs into a vastly expanded semantic field. Now we can consider a unit of meaning within a more broadly networked milieu of its related contexts. And just as creative writing with the typewriter requires a consideration of the constraints and affordances of its medium, an emergent poetics will accompany this as people come to understand its utility.

To explain it more clearly, Horace Walpole’s Castle of Otranto, published in 1764, which was initially presented as the translation of a much older “found” manuscript, translated by William Marshall. Walpole’s work (widely considered the first gothic novel) relied upon spatiotemporal distance (combined with translation) impeding verification. In addition, it uses what was then new media to create what would become a new form of literary convention (the publication apparatus functioning as a frame for the story). In the age of newsprint, Edgar Allan Poe published “The Great Balloon Hoax” in The New York Sun in 1844, which told the tale of a man who crossed the Atlantic in a balloon, which is often credited as an early contribution to Science Fiction. Orson Wells’ The War of the Worlds (1938) relied upon the new medium of radio to broadcast information that was difficult to verify over a large geographical area. While the degree of panic associated with it might be more myth than reality, the truth is that it was a sensational broadcast and, in some instances, was mistaken for reliable information. More recently, The Blair Witch Project, a low budget horror film from 1999, passed itself off as “found footage,” made special use of the internet in its early promotions. In each instance, artists have made creative use of emerging media to expand the possibilities for representation in ways that are now generally lauded for their creativity.

Where AI is fascinating is in its ability to expand poiesis in vastly new directions. While much of the angst is focused on heavily determined fields of reading/writing (University, corporate workspace, audio-visual, and social media), its implications go beyond these established contexts. But a key consideration about AI is that it is asymmetric in its technical impacts on society. AI tends to have a conservative function within large institutions—optimizing, consolidating, and cost cutting. Institutional applications generally pertain to legacy workflows, bureaucratic accountability, risk management, budgets, branding, labor costs, standardization, gatekeeping, etc. On the other hand, outside of institutional/bureaucratic contexts it can be used in more generative ways. People with limited capital, partial expertise, smaller (or no) teams, etc. can leverage their ideas more quickly through research, prototyping, testing, distribution, etc. Used in this way, it can create more entry points for cultural participation (and what that publication is could resemble things widely recognized as texts, but could also include analytic insights, software, physical objects, processes, genetics, machines, or any number of invented things).

In other words, AI can be proletarianizing when used at institutional scale to streamline and regulate. Firstly, it vastly increases the efficiency of these institutions, freeing up capital for them to expand along their current trajectories. Secondly, many of the positions within these institutions can be downsized. When so much institutional work is itself formulaic busywork—distilling data from a variety of sources, implementing policies in consistent language at all levels, preserving the coherence of corporate identity, tamping down on aberrations that are out of step with top down directives, assessing performance and increasing productivity—many of these institutional jobs, whether they are in business, government, or elsewhere will themselves become irrelevant. When the goal of training is to instill a robotic set of behaviors, especially behaviors that are primarily pertaining to the reading and writing of symbols in digital networks, those positions are under direct threat of automation.

On the other hand, AI is potentially de-proletarianizing when used by marginal actors to push their way through or around cultural enclosures. Obviously, this requires the existence of many models and approaches, transparency and access to multiple views, interoperability, and a patronage model that favors autonomy and plurality. But these tools are enormously powerful, particularly for people who are not trained to operate crisply in highly routinized environments, who are unschooled (wholly or partially), who are autodidacts, who develop ideas about established knowledge areas from outside perspectives, who have entrepreneurial ambitions, whose minds or behaviors are eccentric. There are a great many people out in the world who have rich experiences and brilliant ideas that are borne out of situations that generally exclude them from participation in industrial culture. And since AI can assist in research, brainstorming, prototyping and testing, the sudden proliferation of participants (particularly those who have little to lose by taking chances) has enormous potential.

We forget, from our stable knowledge institutions and settled epistemologies, about how genius works and how knowledge is created. But as we stand at the twilight of a passing system and at the dawn of a new one, we must consider the possibility of a new Renaissance. In an age where certain kinds of busywork (particularly the kind of busywork that is done on a computer) is likely to be automated, we must push towards creative practices beyond those typically done on a computer. This means creating practical workspaces—workshops, studios, seminars—and pushing for speculative practices oriented towards discovery—experiments, poetics, explorations. Historically, these are areas that welcome outsiders, tinkerers, artists, entrepreneurs, hustlers, and hackers. It’s worth remembering that Gutenberg was not a publisher (not even a printer), but a jeweler trying to bail himself out of economic trouble. Never mind the fact that Ben Franklin, Abraham Lincoln, Jane Goodall, Maya Angelou, Malcolm X, Albert Einstein, Jorge Luis Borges, Mark Twain, Buckminster Fuller, Vincent Van Gogh, Frank Lloyd Wright, and many of the other autodidacts we revere would all be considered failures in their fields, but for the fact that they defined those fields.

This returns us to the question of significance. Many scholars lament the decline of reading, noting its connection to the cultivation of empathy, its relationship to interiority, the development of free expression, its role in deepening concentration, and the fruits of humanistic sensibilities. But what if our cognition is less a product of a particular medium of expression and more a consequence of a framework that enabled agency? What if the Gutenberg revolution unleashed a drive for significance whose chief mode of expression was in the outputs associated with participation in the ecology that contained it? And what if, as one writing system reaches its point of conclusion and a new paradigm takes root, we are on the cusp of a new wave of creative achievement? I argue that as the long era of the book comes to a close, we see evidence of its waning potential. I do not mean to say that “books” will stop being written and read or that they will cease to be significant, but only that the next wave of great human achievements will not retread the past patterns of success that are formally coded into the highly formulaic approaches that are mere echoes of past achievement. If we remain bound to the professionalized routines of the past, buried in incessant busywork, the emulation of competence and pantomimes of creativity, our achievements will be mere shadows cast by those who came before us.

On the other hand, those who see themselves at the threshold of a new era have the potential to create things of significance, both great and small. Empathy comes from the real desire to listen to needs and respond to them with creativity. Interiority comes from the sense that one’s own thoughts are of consequence, and the logical expectation that the passions of others is invested with a similar expectation of consequence. Free expression comes from the feeling that one has something worth offering. Concentration comes from the realization that focus is worth the effort. And the cultivation of well-rounded Humanism ultimately derives from the feeling that our lives, in a world of billions, still matter. In other words, unleashing ideas is significant. In an age when writing offered people from far flung places the opportunity to participate in the life of a lively emerging culture, the act of reading and writing was a dominant path to significance. In an age where complex ideas can be shared through multi-modal artifacts of the new media environment, we have new opportunities for significance. And while this new environment is likely to be awash in all manner of unimaginative slop, from this frantic array of noise, we will see an increasing proliferation signals, incredible creative efforts, and groundbreaking achievements.

Works Cited

Adorno, Theodor, and Max Horkheimer. “The Culture Industry: Enlightenment as Mass Deception.” Excerpt of “The Culture Industry: Enlightenment as Mass Deception” from The Dialectic of Enlightenment, transcribed by A. Blunden [1998] 2005. https://www.marxists.org/reference/archive/adorno/1944/culture-industry.htm

Augustine. The Confessions of Saint Augustine. Trans. E.B. Pusey. 1909. https://www.gutenberg.org/files/3296/3296-h/3296-h.htm

Baudrillard, Jean. The System of Objects. Trans. J. Benedict. New York: Verso, 1996.

Benjamin, Walter. “The Work of Art in the Age of Mechanical Reproduction.” Trans. Harry Zohn. Marxists.org. 1998 [1936]. https://www.marxists.org/reference/subject/philosophy/works/ge/benjamin.htm

Bush, Vannevar. “As We May Think.” The Atlantic Monthly. (July 1945). https://web.mit.edu/sts.035/www/PDFs/think.pdf

De Certeau, Michel. The Practice of Everyday Life. Trans. Steven Rendall. Los Angeles: U of California Press, 1984.

Deleuze, Gilles, and Felix Guattari. A Thousand Plateaus. Trans. Brian Massumi. Minneapolis: University of Minnesota, 1987.

Derrida, Jacques. Of Grammatology. Trans. Gayatri Chakravorty Spivak. Baltimore & London: Johns Hopkins University Press, 1976.

Foucault, Michel. The Order of Things: An Archaeology of the Human Sciences. New York: Vintage Books, 1994

Flusser, Vilem. “The Gesture of Writing.” Flusser Studies 8 (May 2009): https://www.flusserstudies.net/archive/flusser-studies-08-may-2009

Geoghegan, Bernard. Code: From Information Theory to French Theory. Durhan, North Carolina: Duke University Press, 2024.

Hall, Stuart. “Encoding and Decoding in the television discourse.” Discussion Paper. University of Birmingham, Birmingham, 1973. https://epapers.bham.ac.uk/id/eprint/2962/

Halpern, Orit, and Robert Mitchell. The Smartness Mandate. Cambridge: MIT Press, 2023.

Heidegger, Martin. Being and Time. Trans. Joan Stambaugh. Albany: State University of New York Press, 1996.

—. “The Question Concerning Technology.” Trans. William Lovitt. Basic Writings. Editor David Farrell Krell. San Francisco: Harper Collins, 1993. 311-41.

Lyotard, Jean-François. The Postmodern Condition: A Report on Knowledge. Trans. Bennington and Massumi, Manchester: Manchester University Press, 1984.

Nelson, Theodor Holm. “A File Structure for The Complex, The Changing and the Indeterminate.” ACM National Conference. 1965. https://monoskop.org/images/archive/7/73/20180929161639%21Nelson_Ted_1965_A_File_Structure_for_the_Complex_the_Changing_and_the_Indeterminate.pdf

Nietzsche, Friedrich. Thus Spoke Zarathustra. Trans. Walter Kaufmann. New York: Random House, 1954.

Sartre, Jean-Paul. Being and Nothingness. Trans. Hazel E. Barnes. New York: Washington Square Press, 1956.

Stiegler, Bernard. “The Disaffected Individual in Process of Psychic and Collective Disindividuation.” Ars Industrialis. trans. Patrick Crogan and Daniel Ross. August 2006. https://arsindustrialis.org/disaffected-individual-process-psychic-and-collective-disindividuation Tokenizer. https://platform.openai.com/tokenizer

Vaswani, Ashish, et al (2017). “Attention is All You Need” (PDF). 31st Conference on Neural Information Processing Systems (NIPS 2017), 4-9 December 2017, Long Beach, CA, USA. https://proceedings.neurips.cc/paper/2017/file/3f5ee243547dee91fbd053c1c4a845aa-Paper.pdf

Footnotes

1. Stiegler writes about this question of individuation in relation to media: But in the hyperindustrial political economy, value must be completely calculable; which is to say, it is condemned to become valueless—such is its nihilism. The problem is that it is the consumer who not only is devalued (for he is evaluated, for example, by the calculation of his “life time value”) but who equally is devalorized—or, more precisely, he is disindividuated. In such a society—which liquidates desire, desire which is, however, energy, libidinal energy—value is what annihilates itself and, with it, those who, evaluating it, are themselves evaluated. This is why it is society as such which appears finally to its members, themselves devalorized (and melancholic), as being without value—and this is also why society fantasizes its “values” that much more noisily and ostentatiously, “values” which are only deceptions, compensatory discourses, and consolations. Such is the lot of a society which no longer loves itself. (“Disaffected”) ↩

2. We could create an entire bibliography of 20th Century thinkers describing the trajectory. From Adorno and Horkeimer’s critique of culture industries transforming the field of human expression into commodified industrial products to Foucault’s archaeology of epistemic systems, from Baudrillard’s quirky incursions into spectacular media to Lyotard’s observation about the collapse of metanarratives, from Jameson’s critique of late capitalism to Derrida’s reflections on language itself, the process of culture’s transformation from a kind of social architecture to surface (which is a core assumption of Marxist orthodoxy, though I don’t think it is fair to reduce postmodernism to Marxism). Personally, I think the relationship between base and superstructure is not linear, a point which is understood by a range of post-Marxist scholars (whether we are talking about those who are self-described “neo-Marxists,” the many “antimodernist” critics, or those who are accused by many right-wing critics of “cultural Marxism”). ↩

3. In an attempt to better represent the complex web of approaches and the many criss-crossing efforts, I asked ChatGPT to access project documentation for major Closed, Open weight, and Open source models at the time of this writing: The closed models included OpenAI (GPT-4, GPT-4o, GPT-5), Google DeepMind (Gemini Ultra, Pro, Nano), Anthropic (Claude Opus, Sonnet, Haiku), Amazon (Titan), xAI (Grok), and Baidu (ERNIE). The open-weight models included Meta (Llama 2, Llama 3), Mistral AI (Mistral 7B, Mixtral 8x7B, Mixtral 8x22B), Google (Gemma), Alibaba (Qwen), 01.AI (Yi), TII (Falcon), Microsoft (Phi-2 and Phi-3), Stability AI (StableLM), and DeepSeek. Among the projects commonly regarded as open source are BigScience (BLOOM), the Allen Institute for AI (OLMo), EleutherAI (GPT-Neo, GPT-J, GPT-NeoX), Google (T5 and FLAN-T5), and RWKV. I subsequently asked ChatGPT to critique my classification, to which it replied: “Most modern AI isn’t truly open source. It’s either closed, or open-weight with significant restrictions.” ↩