Reading with Language Models

Manuscript Information

This essay is part of a proposed collection of essays for Modern Fiction Studies on the topic of Cultural AI edited by Richard Jean So and Aarthi Vadde.

Abstract

Literary scholars’ justified opposition to their students’ use of language models (LMs) to avoid the reading, writing, and thinking crucial to the teaching of literature has obscured how LMs can advance literary scholarship. I argue that literary scholars using LMs have already begun to construct an implicit norm that gives greater license to read with LMs when they adopt what Louise Rosenblatt terms the efferent stance toward literary texts, but less license when they adopt what she terms the aesthetic stance. Rosenblatt’s distinction may help mediate the conflict between literary scholars who strongly oppose LMs and those who are curious to know what value they may have for literary studies by demarcating limits to LMs’ use within the discipline’s core practices of reading.

Introduction

Based on the evidence available online, literary scholars appear to be largely united in their opposition to language models (LMs) like ChatGPT. In this, surveys suggest that they are no different from the majority of Americans.See Michelle Faverio and Emma Kikuchi, “What the Data Says about Americans’ Views of Artificial Intelligence,” Artificial Intelligence in Pew Research Center, 2026.

There are many good reasons for opposition.See e.g., Timnit Gebru and Émile P. Torres, “The TESCREAL Bundle: Eugenics and the Promise of Utopia Through Artificial General Intelligence,” First Monday, ahead of print, April 2024, https://doi.org/10.5210/fm.v29i4.13636; Emily M. Bender and Alex Hanna, The AI Con: How to Fight Big Tech’s Hype and Create the Future We Want, First edition (HarperCollins, 2025); and Gael Varoquaux, Sasha Luccioni, and Meredith Whittaker, “Hype, Sustainability, and the Price of the Bigger-is-Better Paradigm in AI,” Proceedings of the 2025 ACM Conference on Fairness, Accountability, and Transparency (New York, NY, USA), FAccT ’25, June 2025, 61–75, https://doi.org/10.1145/3715275.3732006.

An important reason within literary studies is that instructors have observed their students using LMs to avoid the reading, writing, and thinking essential to the teaching of literature.See Beth McMurtrie, “The Reading Struggle Meets AI,” News in The Chronicle of Higher Education, https://www.chronicle.com/article/the-reading-struggle-meets-ai, 2025; and Nataliya Kosmyna, Eugene Hauptmann, Ye Tong Yuan, et al., Your Brain on ChatGPT: Accumulation of Cognitive Debt When Using an AI Assistant for Essay Writing Task, arXiv:2506.08872, arXiv, 2025, https://doi.org/10.48550/arXiv.2506.08872.

However, justified opposition to the ways some students have been using LMs obscures how these models can be used to advance literary scholarship. Used with caution by literary scholars in ways that I will describe, reading with LMs can also occasion reading, writing, and thinking.On LMs and the question of “reading,” see Melanie Mitchell and David C. Krakauer, “The Debate over Understanding in AI’s Large Language Models,” Proceedings of the National Academy of Sciences 120, no. 13 (2023): e2215907120, https://doi.org/10.1073/pnas.2215907120.

While most applications of LMs to texts of interest to literary studies thus far have occurred within digital humanities and adjacent fields, the use of LMs for literary scholarship is not limited to such work. In fact, their use can be so similar to the everyday work of literary studies that many scholars may not realize that they have already begun to use some of them.

Broadly, I will argue that scholars engaged in this work have begun to construct an implicit norm about how to read with LMs. Literary scholars grant greater license to read with LMs when they adopt what Rosenblatt terms the “efferent stance” toward a text, and less license when they adopt what she terms the “aesthetic stance.” A reading on the efferent end of Rosenblatt’s continuum “is centered predominantly on what is to be extracted and retained after the reading event,” whereas a reading on the aesthetic end “adopts an attitude of readiness to focus attention on what is being lived through during the reading event.”“The Transactional Theory of Reading and Writing,” in Theoretical Models and Processes of Literacy, 7th ed. (Routledge, 2018), 458.

This emerging norm has rarely been theorized in these terms.For a previous use of Rosenblatt’s distinction within digital humanities more generally, see Tom Liam Lynch, “Electrical Evocations: Computer Science, the Teaching of Literature, and the Future of English Education,” English Education 52, no. 1 (2019): 15–37, https://doi.org/10.58680/ee201930312.

While both stances are essential to the discipline’s “core practice” of close reading, literary studies also depends upon many other reading practices that receive less attention.John Guillory, On Close Reading (The University of Chicago Press, 2024), 3.

These include but are not limited to browsing, skimming, and scanning texts. Rosenblatt’s distinction may help mediate the conflict glossed above between literary scholars who strongly oppose the use of LMs within literary studies and those who are curious to know what value they may have for the field by demarcating limits to LMs’ use within the discipline’s core practices.

I will make the case for this distinction through many examples that fruitfully apply LMs to literary texts, literary criticism, and paraliterary texts of interest to literary studies. By reviewing the tasks for which these models have been used, I show how they can be used persuasively despite their weaknesses. For example, hallucinations—LMs’ best-known weakness—cannot be eliminated; they are inherent to the autoregressive structures of these models.Mark Russinovich, Ahmed Salem, Santiago Zanella-Béguelin, and Yonatan Zunger, “The Price of Intelligence,” Commun. ACM, ahead of print, August 2025, https://doi.org/10.1145/3749447.

Hallucinations therefore pose a significant risk to the use of LMs for research. At the same time, researchers have used and invented techniques to mitigate hallucinations substantially. To claim that LMs are useless because hallucinations are inevitable is analogous to claiming that GPS turn-by-turn navigation is useless because it does not correctly guide drivers to their destination every time. In both cases, people figure out how to use imperfect technologies in ways that account for their shortcomings.

A lot of the research that has already applied LMs to literary texts and questions has been published and presented in venues unfamiliar to most literary scholars. Because of this, I review some of it here in a way that attempts to make its relevance to literary studies writ large clear. I discuss reading with LMs in several distinct contexts: How scholars have already applied LMs to literary texts and literary criticism for tasks including annotation, summarization, imitation, and information retrieval; how literary scholars have and will use vector search and retrieval augmented generation to complement their browsing, skimming, and scanning, especially of secondary sources; and how I have used LMs to classify and extrude data from paraliterary texts that contain unstructured information about authors, works, and their relative positions in literary canon. All of these uses of LMs occur closer to the efferent end of Rosenblatt’s continuum. However, information they reveal about the texts thus analyzed can also occasion readings closer to the aesthetic end.

My focus on how literary studies has and might read with LMs differs from many recent discussions about LMs and the humanities. Much of that work has focused on the relationship between the questions, knowledge, and methods of the humanities, and the creation, use, and effects of LMs in the world.e.g., Aarthi Vadde, “Inside and Outside the Language Machines,” PMLA 139, no. 3 (2024): 553–58, https://doi.org/10.1632/S0030812924000579; Ruha Benjamin, “The New Artificial Intelligentsia,” in Los Angeles Review of Books, https://lareviewofbooks.org/article/the-new-artificial-intelligentsia, 2024; Katherine Elkins, “A(I) University in Ruins: What Remains in a World with Large Language Models?” PMLA 139, no. 3 (2024): 559–65, https://doi.org/10.1632/S0030812924000543; Lauren Klein, Meredith Martin, André Brock, et al., Provocations from the Humanities for Generative AI Research, arXiv:2502.19190, arXiv, 2025, https://doi.org/10.48550/arXiv.2502.19190; Drew Hemment and Cody Kommers, Doing AI Differently: Rethinking the Foundations of AI via the Humanities (Zenodo, 2025); and Edwin Roland and Richard Jean So, Generative AI & Fictionality: How Novels Power Large Language Models, arXiv:2603.01220, arXiv, 2026, https://doi.org/10.48550/arXiv.2603.01220.

Others have focused on the interrelationships between university administrators, university workers, and the increasing (and increasingly worrisome) proportion of university budgets diverted to big tech, including but not limited to contracts for LMs.See e.g., Matthew Kirschenbaum and Rita Raley, “AI and the University as a Service,” PMLA 139, no. 3 (2024): 504–15, https://doi.org/10.1632/S003081292400052X; Annie McClanahan and Louise McCune, “Ed Tech,” in University Keywords, ed. Andy Hines, Critical University Studies (Johns Hopkins University Press, 2025); and Matt Seybold, “Against Technofeudal Education,” Substack Newsletter, in The American Vandal, 2025. For the view from the professioriate, see Artificial Intelligence and Academic Professions (American Association of University Professors, 2025). On AI and labor more generally, see Matteo Pasquinelli, The Eye of the Master: A Social History of Artificial Intelligence (Verso, 2023).

I do not focus on the field’s use of these models to add to the AI hype. Rather, I do so in opposition to a weak critique. Too many arguments against the use of LMs in literary studies proceed from the supposition that they are not or cannot be useful to the discipline because they have been harmful in the classroom. However true this argument may be of the classroom, it is demonstrably false with respect to research. Ethical arguments against LMs—how they disempower labor, reproduce biases including but not limited to racism and sexism, rely on copyrighted material without compensating its creators, accelerate surveillance by the state and by capital, etc.—are far stronger. Claims made for or against LMs on grounds of their usefulness demand serious engagement with both their capacities and disciplinary norms governing their use. While I cannot discuss the former here as the state of the art expires weekly, I will attempt to describe how I see the latter emerging.

Generating readings with LMs

While this essay discusses scholars using LMs to advance literary research, it does not advocate for LMs independently generating readings of literary texts. A common rejoinder to LM-generated text captures the prevailing view of this issue: “Why should I bother to read something that no one bothered to write?”There is little doubt that LM-generated research has already been submitted and may well be published, if it has not been already. For example, Weixin Liang, Yaohui Zhang, Zhengxuan Wu, et al., “Quantifying Large Language Model Usage in Scientific Papers,” Nature Human Behaviour 9, no. 12 (2025): 2599–609, https://doi.org/10.1038/s41562-025-02273-8 show that words disproportionately favored by LMs like pivotal and intricate began appearing much more often in scientific abstracts after the release of ChatGPT.

My aim in this section is to articulate the assumptions on which this question rests. A strong version of its argument must hold true in a hypothetical future where LM-generated literary scholarship is indistinguishable from the work of experts. Today, one can argue for the superiority of expert work on the merits. But it is more useful to contemplate why this judgment would persist even if LM-generated text were one day indistinguishable from expert work. That answer turns on reading and writing as embodied experiences.

A thought experiment will help to ground this point. Jorge Luis Borges’s character Pierre Menard wants to live his life such that he will “produce a number of pages which coincided—word for word and line for line—with those of Miguel de Cervantes.”“Pierre Menard, Author of the Quixote,” in Collected Fictions, trans. Andrew Hurley (Allen Lane The Penguin Press, 1999), 91.

In the judgment of the narrator of Borges’s story, “The Cervantes text and the Menard text are verbally identical, but the second is almost infinitely richer. (More ambiguous, his detractors will say—but ambiguity is richness).”“Pierre Menard, Author of the Quixote,” 94.

The same words, but not the same meaning. Menard’s seem “richer” because of his quixotism.

Now, imagine if a future LM could be trained on a scholar’s prior reading and writing such that it could generate a new reading of a new text that is identical to one that scholar had independently written. Even though these texts would be identical, most literary scholars would not regard them as being of equal value because they emerged from different contexts.

The scholar’s version would be considered more valuable because it testifies to the embodied experiences of specific texts encountered by specific readers at specific times. A version of this position has been central to feminist scholarship, Black studies, queer theory, disability studies, and many other fields. As Paula Moya put it recently, “Because a work of literature is only actualized in the process of being read, it can never be the same for all readers or even for the ‘same’ reader over time.”“Some Propositions on Close Reading,” Symploke 32, no. 1 (2024): 359.

And one reader’s embodied experience of one text at the moment of its reading—what Derek Attridge calls “the literary event”—cannot be modeled, either.The Singularity of Literature, Routledge Classics (Routledge, 2017), 84–85.

Writing after reading discloses aspects of that experience that the writer hopes to make meaningful to other readers. As I.A. Richards put it a century ago, “Criticism is the endeavour to discriminate between experiences.”Principles of Literary Criticism, International Library of Psychology, Philosophy, and Scientific Method (K. Paul, Trench, Trubner, & Co., ltd.; Harcourt, Brace & Co., inc, 1925) viii.

A century later, Lauren Klein et al. said, “Models make words, but people make meaning.”Provocations from the Humanities for Generative AI Research, 1.

Dan Sinykin and Johanna Winant emphasize embodiment in close reading by narrating how the method works with anaphoric emphasis on you: “…you start with someone else’s words; you see something in those words that means something to you…And now you might look up from the text because you want to show what’s happened to someone else. You want to explain something to another person: your own reader.”Close Reading for the Twenty-First Century, Skills for Scholars (Princeton University Press, 2025), 1.

N. Katherine Hayles has made the same point in the context of LMs: “Literary criticism…has always worked from one customary presupposition: that the texts it interrogates have been written by humans with language processed by human brains.”Bacteria to AI: Human Futures with Our Nonhuman Symbionts (The University of Chicago Press, 2025), 139.

What Hayles says here of literary texts applies equally to scholarship. LM-generated texts do not imply that embodied work of observation, interpretation, and communication. LM outputs mean differently without this context, just as Menard’s Quixote means differently than Cervantes’s.

Leif Weatherby might characterize this distinction as “remainder humanism.” It is a “remainder” in the sense that it defines as human the “ever shrinking area of things that ‘computers can’t do’.”Language Machines: Cultural AI and the End of Remainder Humanism, Posthumanities 74 (University of Minnesota Press, 2025), 37.

Weatherby is right that defining the human in the negative is a losing game. Recent studies have shown that the difference between human and machine performance on measurable aspects of close reading is surprisingly small: no statistically significant difference was observed in the grades assigned to essays about Old English poetry between Oxford University students and GPT-4;T. Revell, W. Yeadon, G. Cahilly-Bretzin, et al., “ChatGPT Versus Human Essayists: An Exploration of the Impact of Artificial Intelligence for Authorship and Academic Integrity in the Humanities,” International Journal for Educational Integrity 20, no. 1 (2024): 1–19, https://doi.org/10.1007/s40979-024-00161-8.

GPT-4 approximated or outperformed literary scholars at correctly identifying poetic forms of unlabeled poems;Melanie Walsh, Anna Preus, and Maria Antoniak, Sonnet or Not, Bot? Poetry Evaluation for Large Models and Datasets, arXiv:2406.18906, arXiv, 2024, https://doi.org/10.48550/arXiv.2406.18906.

a small LM performed better than the average of human evaluators on college-level multiple-choice close reading questions;Peiqi Sui, Juan Diego Rodriguez, Philippe Laban, et al., KRISTEVA: Close Reading as a Novel Task for Benchmarking Interpretive Reasoning, arXiv:2505.09825, arXiv, 2025, https://doi.org/10.48550/arXiv.2505.09825.

and LM-generated interpretations of texts helped people answer close reading questions about those texts more accurately than they otherwise would have.Jiayin Zhi, Hoyt Long, Richard Jean So, and Mina Lee, What Does AI Do for Cultural Interpretation? A Randomized Experiment on Close Reading Poems with Exposure to AI Interpretation, 2026, https://doi.org/10.1145/3772318.3791727.

This goes beyond close reading quizzes to other supposedly distinctive human abilities. For example, one recent study found that LMs matched humans on questions designed to test theory of mind.James W. A. Strachan, Dalila Albergo, Giulia Borghini, et al., “Testing Theory of Mind in Large Language Models and Humans,” Nature Human Behaviour 8, no. 7 (2024): 1285–95, https://doi.org/10.1038/s41562-024-01882-z.

Even if LMs match or exceed measurable expert performance on such tasks, arguing from embodied experience means that context need not be evident in the text in order to make a meaningful difference.

However, arguing from embodied experience does cut against key insights from the conflict between phenomenology and structuralism. Jacques Derrida would have challenged the notion that the absence of embodied context from LM-generated text differentiates it from any other instance of writing because “a written sign carries with it a force that breaks with its context, that is, with the collectivity of presences organizing the moment of its inscription.”“Signature Event Context,” in Limited Inc (Northwestern University Press, 1988), 9.

For Derrida, what applies to writing also applies to “the entire field of what philosophy would call experience, even the experience of being,” challenging whether this distinction is any distinction at all.“Signature Event Context,” 9.

With respect to LMs, Hayles calls this “the null strategy,”Bacteria to AI, 147.

which she opposes because it relies on “the incorrect assumption that [LM-generated texts] display interiority and subjectivity.”Bacteria to AI, 144.

Widespread uptake of the rejoinder “Why should I bother to read something that no one bothered to write?” suggests that Derrida’s argument is being reconsidered in the context of LMs. This would be a striking turn of the dialectic, given that, as Ted Underwood has argued, contemporary LMs themselves represent “the empirical triumph of theory,” especially structuralism. As the philosopher Alva Noë recently put it, “Computers, however vitally important these may become as technological extensions of our work, never enter into human being.”The Entanglement: How Art and Philosophy Make Us What We Are (Princeton University Press, 2023), 160.

LMs cannot independently generate readings that can answer the question “Why should I bother to read something that no one bothered to write?” because no one bothered to write them. Embodied experiences of reading and writing appear to be prerequisites for such work, albeit ones that did not need to be named before it became possible that a reading could be created without them. However, LMs need not generate close readings independently to be useful to close readers.In some fields, LMs have been used to simulate feedback from peer reviewers and editors. In the sciences, see Weixin Liang, Yuhui Zhang, Hancheng Cao, et al., “Can Large Language Models Provide Useful Feedback on Research Papers? A Large-Scale Empirical Analysis,” NEJM AI 1, no. 8 (2024), https://doi.org/10.1056/AIoa2400196. In creative writing, see Katy Ilonka Gero, Tao Long, and Lydia B. Chilton, “Social Dynamics of AI Support in Creative Writing,” Proceedings of the 2023 CHI Conference on Human Factors in Computing Systems (New York, NY, USA), CHI ’23, April 2023, 1–15, https://doi.org/10.1145/3544548.3580782.

The remainder of this essay will show how LMs are already being used to create evidence that supports literary scholarship.

Reading with LMs

By now, researchers have already been reading literature with LMs for several years. For example, the National Endowment for the Humanities first funded the AI for Humanists project, which seeks to help position humanists “to make use of—and to critique” LMs, in 2021.“The AI for Humanists Project,” in AI for Humanists, http://www.bertforhumanists.org//, 2025.

However, much of this work has been published in venues far afield from the reading of most literary scholars. In this section, I describe some recent research that exemplifies why and how literary studies has and will read with LMs.For a review that covers the humanities and social sciences more broadly, see Andres Karjus, “Machine-Assisted Quantitizing Designs: Augmenting Humanities and Social Sciences with Artificial Intelligence,” Humanities and Social Sciences Communications 12, no. 1 (2025): 277, https://doi.org/10.1057/s41599-025-04503-w.

While these studies tend to foreground computational and statistical methods, I have prioritized discussing studies that have co-authors who are literary scholars, who are well aware of where and why their approaches break from traditional literary scholarship. I organize this brief discussion around some of the key tasks for which scholars have used LMs with literary texts.

One such task is annotation. While different in form, annotation with LMs is identical in spirit to marking up a book with any systematic approach to marginalia, highlight colors, or page flags. The difference is that digital annotations are computationally tractable, so they can be used to count examples or retrieve text associated with one or more annotations. For instance, Andrew Piper and Sunyam Bagga use GPT-4 to annotate passages from many kinds of prose narratives to identify which predetermined characteristics, if any, a given passage possesses (e.g., does it contain specific markers of time?). They find that the model annotates passages in ways that tend to agree with human annotations of the same passages for the same features. Similarly, Catherine Yeh, Tara Menon, et al. use LMs to annotate where characters are geographically located at specific points in the narratives of novels, plays, and epic poems. They use the resulting data to visualize the relationship between time, space, and characters over an entire text, such that they can show which characters are at the Bennetts’ and which at Netherfield across narrative time in Pride and Prejudice. Their approach also demonstrates how LM annotations can be reviewed and corrected by researchers after generation, a research process usually referred to as “human-in-the-loop.” Haaris Mian, Melanie Subbiah, Sharon Marcus, et al. use LMs to operationalize and extend Alex Woloch’s account of character. Where previous such analyses primarily focused on computationally tractable elements (like direct mentions of a character’s name), they extend this model to more complex categories that they use LMs to annotate, including interiority, action, discussion by other characters, and discussion by the narrator for each character. Their model uses the sum of these factors as an index of how major or minor individual characters are.Haaris Mian, Melanie Subbiah, Sharon Marcus, Nora Shaalan, and Kathleen McKeown, Computational Representations of Character Significance in Novels, arXiv:2601.15508, arXiv, 2026, 3, https://doi.org/10.48550/arXiv.2601.15508.

In these examples and others like them, researchers use LMs to annotate passages from literary texts in order to extract data about structural or formal features that can be used for other analyses—a quintessentially efferent reading practice. As these examples also suggest, one standard approach to assessing the accuracy of LM annotations compares them to identical annotations created and validated by experts on a representative sample of the texts to be annotated. Such annotation tasks often find that LM annotation is faster and cheaper than human annotation, while producing results that evaluations suggest are as good or better than those produced by humans. Their speed and performance also makes it possible to do this kind of annotation at otherwise unimaginable scales.See Cody Kommers, Drew Hemment, Maria Antoniak, et al., Meaning Is Not A Metric: Using LLMs to Make Cultural Context Legible at Scale, arXiv:2505.23785, arXiv, 2025, https://doi.org/10.48550/arXiv.2505.23785.

While working at larger scales does not necessarily confer advantages for close reading (though it can), it clearly makes a difference for fields of literary study that attempt to make sense of larger bodies of texts, such as literary history, genre theory, and stylistics.

Summarization is another task for which researchers have applied LMs to literary texts. Cleanth Brooks’s heresy of paraphrase notwithstanding, summarizing texts can be useful for researchers who wish to identify passages that discuss similar themes or topics using dissimilar language. When literary works address subjects indirectly, metaphorically, or through omission, computational approaches like counting words can conceal similarities between passages. For example, as Toni Morrison emphasizes in Playing in the Dark, American literature’s pervasive “Africanist presence” is evinced through “significant and underscored omissions.”Playing in the Dark: Whiteness and the Literary Imagination, The William E. Massey, Sr. Lectures in the History of American Civilization 1990 (Harvard University Press, 1992), 6.

Building on such observations, Lucy Li et al. use LMs to summarize passages from fiction, asking the LM to “tell” rather than “show” what happens in each passage. They use the resulting summaries as inputs for topic modeling, a natural language processing technique that predates LMs for characterizing the topics that documents discuss based on the underlying distributions of words in a corpus.David M. Blei, “Probabilistic Topic Models,” Communications of the ACM 55, no. 4 (2012): 77, https://doi.org/10.1145/2133806.2133826.

Eschewing topic modeling, Andrew Piper and Sophie Wu have used LMs to annotate narrative topics in news and fiction directly, finding that LMs performed as well as humans with respect to identifying news topics, but outperformed humans when identifying topics in fiction. LM summaries help researchers identify topics, concepts, and patterns in their distributions.

Imitation of authorial style is another task related to but distinct from summarization. In literature classrooms, imitation has long been used as a pedagogical technique to help students better understand minutiae of authorial style. Gabi Kirilloff et al. use GPT-4 to generate 6,000 synthetic paragraphs in the styles of ten nineteenth-century authors ranging from Pauline E. Hopkins to Charles Dickens. They find that GPT-4’s imitations are generally easy to detect, capturing authors’ “themes without capturing literary style.”“‘Written in the Style of ’: ChatGPT and the Literary Canon,” Harvard Data Science Review 7, no. 4 (2025): 22, https://doi.org/10.1162/99608f92.6d5fb5ef.

For example, the synthetic passages use nouns and determiners much more often than do the authors the LM was assigned to imitate, so much so that the researchers can predict whether a passage was written by the author or the LM more than 95% of the time. However, as in the classroom, degrees of failure in imitation also reveal aspects of authors’ styles. They show that Mark Twain is an exception to their general conclusion: GPT-4 imitates Twain much better than the other authors in their sample.

Information retrieval is a fourth task for which LMs have been used with literary texts, and one that I will discuss further in the next section. Where the previous studies I have cited in this section apply LMs directly (or, in the case of imitation, indirectly) to literary texts, Katherine Thai and Mohit Iyyer apply LMs to both literary fiction and literary criticism simultaneously. Specifically, they provide an LM with the full text of a work of prose fiction as well as a work of literary criticism that features at least one direct quote from that work of fiction. For the experiment, one direct quote has been blanked out, though the surrounding critical context remains. The LM is then asked to determine which quote from the provided fiction has been blanked out in the criticism. The authors find that Google’s LM Gemini outperforms experts at correctly identifying which quote has been removed. This is an exemplary information retrieval task because the goal is to find a specific passage in the fiction (the “needle”) that best fits the context provided by both the criticism and the fiction itself (the “haystack”).On needles in haystacks, see also Sil Hamilton, Rebecca M. M. Hicke, Matthew Wilkens, and David Mimno, Too Long, Didn’t Model: Decomposing LLM Long-Context Understanding With Novels, arXiv:2505.14925, arXiv, 2025, https://doi.org/10.48550/arXiv.2505.14925.

The authors argue that this demonstrates LMs’ capacity to assist with what they term literary evidence retrieval: Given a source text and a critical context, an LM can do as well as (or better than) experts at identifying omitted evidence.

Although there are other studies that I could discuss (as well as nits that could be picked with these studies), these four tasks give a good sense of how and why scholars have used LMs to read literary texts and literary scholarship directly. These descriptions also suggest some of the ways in which the assumptions underlying this work differ from those of most literary scholarship. For example, much of this work uses LMs to identify and evaluate many examples with shared properties, rather than focusing on how a few examples illuminate larger wholes. Though such aggregative approaches are most strongly associated with computational literary studies, they are also directly applicable to other fields such as literary history, genre studies, and stylistics, among others.See e.g., Oleg Sobchuk and Artjoms Šeļa, “Computational Thematics: Comparing Algorithms for Clustering the Genres of Literary Fiction,” Humanities and Social Sciences Communications 11, no. 1 (2024): 1–12, https://doi.org/10.1057/s41599-024-02933-6.

All of these examples also directly or indirectly acknowledge two limitations of these approaches. The first is that all of these tasks artificially reduce the complexity of individual literary texts, though they also do so for limited purposes. This reduction is characteristic of efferent reading. The second is that LMs’ stochastic natures mean that, even if their outputs are correct in the vast majority of cases, they will be wrong in some cases, and it is not possible to predict why or how they will go wrong when they go wrong. Such errors are intrinsic to the autoregressive structures of these models; they can be reduced, but it is not clear that they can be eliminated.See Gary Marcus, Taming Silicon Valley: How We Can Ensure That AI Works for Us (The MIT Press, 2024).

This is usually cited as the best argument against using LMs for research. However, critiques that stop there miss two points that the research reviewed above demonstrates. First, researchers can account for that degree of error using techniques like human-in-the-loop verification, ensemble methods, and statistical approaches to quantify their uncertainty. Second, the relevant comparator here is not perfect accuracy in evaluating a small number of cases, but expert accuracy in evaluating a large number of cases.

Browsing, scanning, skimming with LMs

As both Pierre Bayard and Amy Hungerford have suggested, literary scholars must solve the problem of having too much to read by browsing, skimming, scanning, and sometimes skipping texts.The distinction: “Skimming is defined as getting the main idea or gist of a selection quickly and scanning as a high speed search for the answer to a specific question or the location of a specific fact” (Martha J. Maxwell, “Skimming and Scanning Improvement: The Needs, Assumptions and Knowledge Base,” Journal of Reading Behavior 5, no. 1 (1972): 48, https://doi.org/10.1080/10862967209547021).

These judgments are made using efferent reading practices, not aesthetic ones. To be sure, much literary scholarship itself rewards Rosenblatt’s aesthetic stance. But it would be impossible for a scholar to do all of their professional reading in that mode. Despite increases in publication and the expansion of admissible evidence, literary studies’ aggregate research time has shrunk as the academic precariat has grown.For the latest figures, see Glenn Colby, “Data Snapshot: Tenure and Contingency in US Higher Education, Fall 2023,” Academe Magazine 111, no. 2 (2025).

Below, I demonstrate how scholars have already begun to complement these techniques for reviewing secondary literature with others enabled by LMs, specifically vector search and retrieval augmented generation (RAG).

Unlike the research described in the previous section, vector search and RAG will seem similar to the ordinary research practices of most literary scholars today, especially for literature reviews. Specifically, they complement keyword search. Because of this connection, it is worth briefly considering how keyword search recently changed scholarly practices of browsing, skimming, and scanning, and how the discipline responded at that time. In 1995, Yahoo! Search came online; scholarly concern about keyword search followed soon thereafter.See Scott Stebelman, “Cybercheating: Dishonesty Goes Digital.” American Libraries 29, no. 8 (1998): 48–51 and Lisa Renard, “Cut and Paste 101: Plagiarism and the Net.” Educational Leadership 57, no. 4 (1999): 38.

Suddenly, students could find texts to plagiarize with ease. Over recent decades, this has gotten easier. Yet warnings about keyword search abated as researchers began to use this tool for their own work.See Christine L. Borgman, Scholarship in the Digital Age: Information, Infrastructure, and the Internet (MIT Press, 2007) and Hannah Frydman, “In Defense of the Search Bar,” The American Historical Review 130, no. 2 (2025): 714–35, https://doi.org/10.1093/ahr/rhaf010.

Few scholars today oppose keyword search with the vigor that some did in the 1990s. In a 2001 essay reflecting on the impact of keyword search on literary studies, David S. Miall lamented that full texts online “offer only a partial and inadequate solution to the needs of a literary scholar; even full-text searching provides access only to words, not to concepts.”“The Library Versus the Internet: Literary Studies Under Siege?” PMLA 116, no. 5 (2001): 1407, https://www.jstor.org/stable/463544.

Vector search and RAG can be combined to search texts in ways more closely aligned with Miall’s conceptual search than keyword search. Though it may seem like a remote possibility now, what happened with keyword search after the 1990s may repeat with vector search and RAG, especially as they are currently being incorporated into electronic resources that scholars already depend on like EBSCO and Primo.

Reading paraliterary texts with LMs

The preceding two sections discussed using LMs to read literary texts and literary scholarship. Both involved the evaluation of LM outputs by experts, whether through comparison to validated data, or through iterative refinement of vector search and RAG results. However, irrespective of the validation processes put in place, some literary scholars likely remain skeptical about using these models to gather information directly from literary texts themselves because of the literary qualities like irony and ambiguity that make those texts interesting to study in the first place. In this section, I will show how LMs can be used to study literature without applying them to literary texts. I will do so by briefly discussing some of my ongoing research on the literary canon, which uses LMs in different ways and on different kinds of texts than those discussed thus far.

Although the literary canon “never appears as as a complete and uncontested list in any particular time and place,” there exist some good proxies for that list, including comprehensive literature anthologies like those published by Norton as well as reference works like The MLA International Bibliography.John Guillory, Cultural Capital: The Problem of Literary Canon Formation, First edition, enlarged (The University of Chicago Press, 2023), 30.

See Erik Fredner and J. D. Porter, “Counting on The Norton Anthology of American Literature,” PMLA 139, no. 1 (2024): 50–65, https://doi.org/10.1632/S0030812923001189 as well as Erik Fredner and Mark Algee-Hewitt, “The MLA International Bibliography’s History of English-Language Literary Studies, 1982-2023,” DH2024 (Arlington, VA), August 2024.

In pursuit of other such lists, I have extended this line of research to the quiz show Jeopardy.

Jeopardy should be of interest to literary scholars because it is one of the few places in American popular culture that routinely engages with literary history. About one in five questions references literature. Unlike anthologies or the Bibliography—both of which approach canonicity from a scholarly perspective—Jeopardy provides evidence about both popular and scholarly understandings of the canon.For other efforts to quantify the popularity-prestige distinction with respect to canonicity, see J. D. Porter, “Popularity/Prestige,” Pamphlets of the Stanford Literary Lab, no. 17 (September 2018) as well as Jean Barré, Jean-Baptiste Camps, and Thierry Poibeau, “Operationalizing Canonicity: A Quantitative Study of French 19th and 20th Century Literature,” Journal of Cultural Analytics 8, no. 3 (2023), https://doi.org/10.22148/001c.88113.

Although about one in five Jeopardy questions references literature, they are not all equally difficult. Easy questions will likely be known by many viewers, whereas difficult questions may only be known by experts. Jeopardy quantifies difficulty through gameplay, such as the different dollar values writers assign to each question. As a result, over the past forty years, Jeopardy clue writers have made thousands of specific wagers about which literary authors and works Americans would be more or less likely to know. Aggregating these wagers provides new insights into the structure of the literary canon. But it remained inaccessible to scholars because Jeopardy questions are not structured for this analysis as written. However, I have used LMs to extract and restructure that latent information for analysis.

The results of this research have already shown difficult questions are more likely to be associated with authors and works better known to literary scholars than to the general public, whereas easy questions are more likely to be associated with literature for young people and popular literature, especially authors and texts with film and television adaptations.Erik Fredner, “The Literary Canon on Jeopardy!, 1984-2024,” DH2025 (Universidade NOVA de Lisboa), July 2025, https://doi.org/10.5281/zenodo.19494801.

Some of the authors most frequently referenced in the most difficult clues include Gogol, Pynchon, Petrarch, and Strindberg whereas some of the easiest are Aesop, Schulz, Collins, Rowling, and Tolkien. However, Jeopardy does not reproduce the popularity-prestige continuum solely along the axis of difficulty. A small number of the most canonical authors like Shakespeare and Cervantes appear at every level of difficulty. However, the relative difficulty of authors’ or texts’ questions tends to reproduce the relative valuation that characterizes canonicity itself.

While there is much more to be said about these results elsewhere, my present purpose is to explain how LMs enabled this research by classifying and extruding structured data from Jeopardy clues. This work differs from other research discussed thus far by using LMs to read paraliterary texts in order to study a literary topic. This distinction matters because it avoids the problem of attempting to read literary texts themselves, which are often studied precisely because of their ability to resist encapsulation.

Conclusion

Researchers have already been reading literature, literary criticism, and paraliterary texts with LMs in ways that should be of interest to literary studies writ large for years. I have argued that an emergent norm across the many uses of LMs discussed here is a greater license to read with LMs for what Rosenblatt calls efferent reading (“what is to be extracted and retained after the reading event”), but lesser license for aesthetic reading (“what is being lived through during the reading event”). LMs cannot independently generate close readings that fulfill literary studies’ expectations without embodied experiences of reading and writing. However, reading with LMs can complement and accelerate the discipline’s equally necessary but less often theorized efferent reading.

I wish to conclude by suggesting some of the ways that the uses of LMs described here seem likely to impact literary studies in the short term. Whether and how literary studies will continue to read with LMs depends on material factors, the most important of which may be cost. In 2026, capital expenditure on LM infrastructure by US big tech companies is expected to exceed two percent of US gross domestic product. If it does, this will make it the second most expensive capital project in US history: far greater than the space race, slightly greater than the railroad build-out of the 1850s, and exceeded only by the Louisiana Purchase.Meghan Bobrowsky, Drew An-Pham, and Alana Pipe, “Big Tech’s AI Push Is Costing a Lot More Than the Moon Landing,” Wall Street Journal, February 2026.

Investors will demand that these extraordinary costs be recouped with interest. However, China’s “six tigers”—LM companies that make models competitive with the state of the art—have prioritized releasing open-weight models that can be freely downloaded and run on hardware of one’s choosing, thereby driving down costs. To compete, both Google and OpenAI have released open-weight models. This competition matters because open-weight models are adequate for many of the research tasks described here, and have other advantages for researchers—chief among them, reproducibility. Proprietary LMs like ChatGPT and Claude may never be cheaper than they now are, which could preclude their use for some kinds of research in the future. At the same time, open-weight models are sufficient for much of the work described here.

Of the applications discussed in this essay, those likely to have the most widespread impact on literary studies are vector search and RAG. Both of these are being incorporated into tools like EBSCO and Primo, which provide the search interfaces for many university library catalogs, as well as search engines like Google. Vector search and RAG will make missing relevant research less likely, such as a monograph that discusses a topic that is related to but would not be identified by a particular keyword search. Tools like Primo Research Assistant also make it possible to generate annotated bibliographies using a library’s holdings. While the latter represents a paradigm shift for library patrons from reviewing a list of search results to reading a chatbot-like interface, I suggested earlier that these kinds of changes may eventually seem as unremarkable as keyword search now seems.

Many of the data-driven research questions discussed above differ from those studied by most literary scholars. However, not every question in literary studies is best approached through close reading. Questions of the sort asked by fields like literary history, genre theory, and stylistics require gathering information from across many different texts, tasks at which these models excel. Scholars studying individual authors with large oeuvres face similar problems. Computational approaches to these problems of scale have historically been limited to the small group of literary scholars who are either computer programmers themselves or who collaborate with programmers. For scholars whose barriers to conducting computational research were technical rather than theoretical, the recent emergence of LM coding agents, the most famous of which is currently Claude Code, suggests that a lack of coding knowledge may no longer be a significant barrier to generating code that meets their needs. That said, doing so responsibly is an entirely different question, and one that falls outside of the bounds of this essay.

Finally, the Jeopardy example demonstrates that even a priori objections to the use of LMs to read literary texts directly does not exhaust their usefulness to literary studies. LMs can read paraliterary texts and extrude latent information about literary history. Obvious candidates for similar work include identifying literary allusions in periodicals, on social media, and in film, television, or podcast transcripts.

Through all of these examples, I have tried to show that the argument against LMs on grounds of uselessness for literary studies is false with respect to scholarship. The better arguments are ethical. Despite intrinsic problems like hallucination, emerging practices make the use of LMs more reliable. Furthermore, the relevant criterion for assessing their reliability is not perfect accuracy, but human accuracy. Under such conditions, LMs can complement literary studies’ reading, though that does not mean that they must.