Literacy, Language Models, Cultural Capital

I’m on sabbatical, but this post goes out with an affirmation of solidarity with everyone getting down to fall teaching in this low dishonest decade. It’s a bad time for teaching. On top of social and political crisis, there’s AI. Or rather, as part of said crisis, there’s AI. It is exhaustingly present everywhere you turn, even when you have retreated to your sabbatical bunker—perhaps especially there, if you are in your sabbatical bunker, as I am, to study the institutions of genre fiction. I am studying how, for a century, Anglophone culture and society developed a system of (apparently) formula-bound text production which, whatever its limitations, many people have turned to for its seeming copiousness and its openness even to those with few literary and economic resources. Meanwhile, in the past decade, Anglophone culture and society have developed a system of (apparently) formula-bound text production which, whatever its limitations, people are turning to in order to fill the internet with garbage text while filling the atmosphere with carbon dioxide and filling the pockets of some astonishingly awful people with more money than you can imagine. So here’s my idea: like the development of genre fiction, “AI” is the product and beneficiary of a significant transformation of literacy—that is, the production and use of writing—and its role in the distribution of social rewards.

In the US, genre fiction’s basic condition of possibility was the achievement of near-universal literacy (thanks to compulsory elementary education) amidst extreme class division. By the early 1920s, large segments of the urban working class had the time, the money, and the ability to enjoy large volumes of cheap fiction; the dispersed and ethnically fractured cultural geography of that class was, I believe, crucial to the establishment of genre specialization in both readers and writers. Genre niches offered possibilities for belonging to their readerships, and they facilitated subcultural success, or at least steady work, for writers who, whether by disposition or due to sexism, anti-Semitism, xenophobia, and classism, had slim chances of success in the literary mainstream.1 There wasn’t necessarily anything heroic about this; consider the stew of xenophobia and ressentiment animating H.P. Lovecraft’s cosmic horror fiction, to say nothing of Lovecraft’s far more boring imitators or their equivalents in other subgenres. But they all emerged when the resources of literacy—writing as well as reading—were more widely distributed than ever before. Pulp genres satisfied cultural and economic needs for writing produced in quantity and sorted in niches. It is a characteristic formation of the short century (say, 1880–1970) of what Wendy Griswold calls “reading culture,” when most people read for information and entertainment.

I think “AI” tells us something about what is happening as reading culture continues to unwind, and the resources of literacy become more scarcely distributed once again.2 The “AI” of the Large Language Models is, as most people appreciate, really text-generation. Insofar as Large Language Models have successfully insinuated themselves into so many social scenes, it is because they can fill the need for text in those scenes. AI boosters want us to believe this is because their products’ “intelligence” is at the same “level” as that of the white-collar employees they are promising to replace; text, they want us to believe, is just a medium for “skills.” But on the contrary, LLMs thrive where the text is the point. LLMs cannot care about what the texts they emit say.3 People turn to them when they cannot make themselves care enough either, but they still have something at stake in how the text looks.

Everyone is familiar with the superficial polish of LLMified text. This is visible in LLM-generated student papers, of course, but equally in LLM-generated contributions to open source projects, clotted with verbosely articulate comments and superfluous trivial code; it goes also for all those spurious scientific papers flooding the arXiv. Circulating scientific preprints used to work precisely because the “smell test”—combining reputational and linguistic signals with scientific ones—was good enough to keep the flow of papers at a reasonable and useful level. I can’t even imagine the flood of blandly nonsensical memos and reports that must be flying over the intranets of the big corporations. The great, baleful discovery of the GPT-makers was that training a deep neural net over a big enough text corpus (the web circa 2022 plus a few tens of thousands of pirated books) could create a plausible generator not just of grammatical texts in English but of the white-collar sociolect and its associated genres.

This lect is what is perceivable as the LLM’s polished writing. The smoothness of its surfaces comes from an apparently apt manipulation of all the conventions of standard written English, not simply standard grammar and orthography but rhetorical conventions of paragraphing, sequencing and (fake-)citing. In the last century the skill of using those conventions was instilled, with great and increasing difficulty, by the high-school English and college gen-ed curricula. That is why it is a class-bound sociolect, or, in other words, a form of cultural capital that distinguishes the beneficiaries of a good high-school or post-secondary education from the rest of a mass-educated society. (It is significant, incidentally, that the workforce for the fine-tuning by “reinforcement learning” has been located in places like Kenya, that is, former British colonies where exploitable workers with English educations were easily available. It was not the first time that transformations of the meaning of English schooling have been launched from colonial situations.)

But this form of cultural capital has plainly been dropping in social value, well before LLMs appeared on the scene. In 1993 John Guillory diagnosed the decline in literature’s status as cultural capital as the underlying motor of the “canon wars” as they then appeared. That arguments about identity representation in literature are still with us is one indicator that the decline and its dilemmas did not go away. Likewise, the long-run diminishment of the college English major and the degradation of schoolteaching as a profession—thanks to unremitting hostility to public schooling on the part of political and economic elites, long predating the latest phase of barbarism in the US—are both symptom and cause of the dwindling role of advanced literacy in social reproduction.

Even so, reading and writing retain significant social honor. It’s still thought virtuous and healthy to be a reader—Usha Vance agrees!—and journalists are still eager to bewail signs of a crisis of literacy: surveys that reveal that only a limited population of people say they have read a book, trends to spend ever more time lost in gadgets with screens, declining test scores, etc. These things—and especially the really shocking fifteen-year downtrend in junior-high-school level reading skills—are indeed major problems, but that they appear as a scandal reveals the degree to which our culture holds some kinds of literacy in esteem.

I think LLMs offer to bridge the gap between the continuing honor accorded to white-collar literacy and the dwindling capacity of the educational apparatus to reproduce it. In general, LLMs bridge this gap with garbage, empty words, bullshit, confabulations, hollow generalizations. But, precisely because advanced literacy is in retreat, the garbage appears more acceptable to the people who attempt to purvey it and to the people who receive it. And not just them: the whole society is tempted to embrace the endless circulation of crap, just as the ruling oligarchs already have. The anomie arising from our seemingly intractable social crisis obviously doesn’t help either. Think of all the organizations that run on flows of texts: not just schools and news media, but pretty much all bureaucracies, from local government to giant corporations. Though it is hard to imagine these organizations can all run indefinitely by pushing machine texts around, in the short term many people may find it tempting to tell themselves that there isn’t much difference between the document they would have had to write five years ago and the text a chatbot can produce on command.

But, to return to the smaller scale of the English classroom: that classroom is always the site of a struggle to get students to stretch their capacity to read and write, overcoming the frustration and shame that comes from their failures on the way. The teacher tries to motivate it by offering the rewards of imagination or pleasure or knowledge—or even social distinction!—but the credibility of the teacher’s promises depends on collective belief in what that teacher purveys. Now, the struggle is getting harder, the rewards diminishing; to more and more students, frustrated, bored, distracted, and fearful, it must all seem like bullshit anyway.


(Edited, September 21, 2026: fixed a few typos that bothered me on re-reading. I don’t have this habitus for nothing.)


  1. Black writers were largely—though not entirely (Gertrude Schalk, Nella Larsen)—excluded from the pulp magazines, and segregated America left less social room, I think, for the multiplication of genre niches for black writers, as opposed to the more culturally aspirant production of the Harlem Renaissance. Still, as Brooks Hefner has shown, parallel genres developed in the fiction printed in black newspapers, even if specialist publication venues did not. ↩︎

  2. The term “AI literacy” is apparently now in common circulation with the meaning “knowing how to use AI to do whatever it is people do with it”; some AI critics also use the term “critical AI literacy,” meaning “knowing what’s wrong with AI, according to us.” But I want to keep my focus on literacy in the sense of writing and reading human language in socially sanctioned ways. ↩︎

  3. In an illuminating essay, the philosopher Keith Frankish argues that large language models should be understood as playing a game, “the chat game”:

    The chat game is a one-player game. The player receives textual inputs from an unknown source, and their task is to produce textual responses that are cooperative by human conversational standards, given the context. Cooperativeness here might be summarized by the four maxims that compose the Cooperative Principle articulated by the British philosopher of language Paul Grice [entailing presumptions of informativeness, truth, relevance, and perspicuousness]…LLMs are playing this game and their responses are motivated solely by a desire to play it or by instrumental desires that subserve this desire…. And that’s the only illocutionary act LLMs perform. They don’t assert, suggest, advise, warn, apologize, question, or do any of the other things we do in producing meaningful utterances.

    The point here is to distinguish the chat game from other kinds of communicative acts without having to commit to any “deep” idea about the specialness of human intention. ↩︎