Turns out metanarratives did not collapse—in form and function, they simply moved to the drop-down menu. As with document templates in Word or PowerPoint, metanarratives of the present moment are modular, preprogrammed by software. But they are also propagated through public discourse by market-researched content, promoted through search rankings by optimization strategies, and promulgated through the social media feeds blasted to our devices by algorithmic amplification. For a timely example, one need look no further than the discourse around generative artificial intelligence. It is by this point commonplace to observe that something called “AI” has infiltrated, to some degree, nearly all domains of cultural production, scientific research, and economic activity. So rapid and so thorough has been the implementation that “AI” has itself become commonplace, and particularly so in the global context of telecommunications. We may accept or ignore the operations of such tools as spam filters, autocomplete, and recommendation algorithms, and we may remark on them with irritation and bemusement. But the tools themselves are now so embedded in everyday life, literally at hand, that they go without saying—and often without naming. They are regarded as customizable feature sets rather than the machine learning applications that they are. What better marker of habituation than the dissolution of the charged synecdoche “AI” into an array of functions, processes, instruments, and brand names?Footnote ¹

Nonetheless, “AI” retains an affective potency that can be mobilized on demand, whether for risk assessments or investment opportunities. This dynamic too has become commonplace. The sensational release of a new model activates both amazement and anxiety as we collectively register the shock of what has become possible and what is likely to come.Footnote ² So too each high-profile error and injurious use (a chat application encourages monarchic assassination; a new image generation model goes predictably awry) animate dark fears of dystopian futures, as well as anger with corporations and institutions for having failed, again, to secure the public interest, to say nothing of basic humanity, which is perceived to be under threat from experimental nonhuman agents with evolutionary capabilities and the capacity for self-redefinition. Throughout, the AI model demos—as design fictions in every sense—do the work of conjuring up fantastic visions of profit and limitless growth.

Meanwhile, educational institutions respond by equal measures of pattern language, convening summits, commissioning white papers, charging ad hoc committees and task forces, and retaining executive consultants—themselves leading instigators of the very media cycles to which the institutions strain to respond.Footnote ³ That the output and outcomes of all this activity take the form of reports, visualizations, and slide decks—outputs that now can easily be machine-generated and by this point likely often are—is not so much symptom of a root cause but proof of concept of a continuous organizational strategy simultaneously anathema to the work of the university while scarily sanguine about the prospects for bringing the medieval institution into line with the data-drawn vectors of twenty-first-century capital.

We acknowledge, of course, that not all institutional responses are necessarily cynical or futile and that the possibilities for meaningful, affirmative, and even redemptive work have not entirely been foreclosed; if we thought they were, we would be writing a different kind of essay. Notwithstanding, what we have to offer is not a narrative of consolation, restoration, or repair. This is not another rallying call for the embrace of critique or resistance or retreat. Nor do we here enthusiastically advocate for the virtues of “leaning in”; the cultivation of new literacies; the opportunities afforded by, well, opportunity; or the intellectual rewards of interdisciplinary collaboration with data and computer scientist colleagues. Certainly, all of these are possible arguments, and all of them are necessary—indeed, we have each made them ourselves, repeatedly, and we will again in other contexts. But these arguments, themselves templated and, we suspect, less read than circulated, are not in our view adequate to the present moment for the profession's journal of record.

At issue is not only or perhaps even primarily “AI” as such. It is rather the all-too-common incapacity of educational institutions, of all types and at all levels, as well as their attendant professions, to be meaningfully responsive to the sociotechnical situation—that is to say, responsive in a manner other than denial, prohibition, capitalization, and capitulation. The rejoinder here might be that the deluge of summits and white papers suggests that perhaps from within the larger complex of partisan oversight, fiscal austerity, and brand maintenance an inspiring vision might emerge, however belatedly. One can always hope, of course, and even try to contribute to the effort. But there is a probability calculation to be made: Will the project devolve into an exercise in rationalization and containment, destined to be out-of-date before it even commences? Indeed, if the purpose is strictly managing the future, unsurprisingly, there are models for that. One can make allowances for, and even regard with some appreciation, the flurry of bans and blue books in the wake of the release of ChatGPT in late November 2022; educators are after all in the business of accreditation and were told, with some insistence, that the student essay was dead and the only way to preserve the academic mission was to cut the Wi-Fi. One also has to appreciate all the sacrificial labor that went into the educational town halls, tutorials, and guidance documents; for campuses (and bodies) still weathering the shocks and aftershocks of the pandemic, when no one has spare bandwidth for much else, a litany of well-intentioned “best practices” has been the default formula.

But shouldn't we—and here our enunciative “we” expands to the whole of the MLA, past and present—be able to do better? No more than a cursory scan of the contents of this journal since its inception in 1884 is needed to substantiate the basic point: our profession has all along been preparing its members to think and work through structural transformations in language practices. For all the disparate methods and materials, for all the agonistic disputes, there has been a coherent project on the back end: we're thrown into this game called language, so let's figure out where it comes from, what it can do, what can be done with it, and how it might be changing. Mentions of actual technologies (typewriter, teletype, film, word processor) are sporadic up to a historical tipping point, and often simply with respect to office communications, but it is nonetheless also clear that, over the centuries, language and literature both have been understood to be subject to external forces that are, if not precisely technological, then at least material. Just as media environments have evolved, so too have scholarly and pedagogic practices, and we are now, as we have continuously been, exhorted to keep up, or to hold the line, to stabilize in relation to inherent instability.

Even so, for all the implicit and explicit worries over time that the profession and its objects are perhaps a little too susceptible to change—that the academy has lost its historical moorings or, now, that market share, mind share, and cultural influence are eroding—the prima facie assumption has been that the thing itself, language, was secure. Fierce contestations of practical uses and metaphysical questions might have suggested different rules for the game being played, but the symbolic ground seemed to remain constant, and intact. That symbolic ground, made up of human-constructed sign systems, can no longer be presumed a priori, and empirical study would confirm the argument. Such has always been the general condition of language in computational environments, of course: we “don't know what our writing does,” because when we type, the flickering signifiers that appear on the screen mask a cascading series of translations from character encoding down through levels of programming languages and assembly code, at the root of which is machine code's manipulation of electrical voltages. Input and output signals, in other words, have been undergirding language as such for many decades of this journal's publication.

However opaque, and however inaccessible to human perception, these processes of translation, numerical representation, compilation, and execution seem now, in the wake of almost-dizzying advances in the domain of natural language processing (NLP), to be quite simple, and quaintly so. We may not have known what our writing did, either technically or theoretically, but in actual practice the scene of writing was unambiguous: press the keys, and the words you typed were (generally) the words you got. The advent of large language models (LLMs) has radically transformed this technolinguistic situation, full stop.Footnote ⁴ More precisely, the implications of the already-widespread and now-accelerating implementation of LLMs are both epistemic and epistemological. And, more concretely, the break has been initiated by the entire end-to-end apparatus of machine learning for language processing—from the initial formation of massive unlabeled training datasets and subsequent tokenization and word embeddings, through complex model architectures (GPT, BERT) and model training, up to an array of sampling techniques (top-k, top-p) and postprocessing steps (filtering, personalization). Thus the scene of writing in this historical moment: prompt a model, and the words you see have been probabilistically generated based on the extrapolation of linguistic patterns from training data composed of sequences of tokens that have been converted into numerical representations.

“Prompting a model” is, appropriately, both a specific activity and an abstraction, just as a “model” is itself both an actual instrument and an epistemic thing. Prompting—calling a model to respond, “chameleon-like,” to user input, as OpenAI explained when it partially released GPT-2 in February 2019 (“Better Language Models”)—is one of a series of cultural techniques for conditioning language model behavior, in addition to fine-tuning, metrics, and benchmarks. But the phrase “prompt a model” also functions as shorthand or a simplified representation or perhaps again a synecdoche for the practice of engaging an increasingly extensive ensemble of applications and interfaces built on top of, or otherwise integrated with, an equally extensive ensemble of language models, large and small, proprietary and open. Put more plainly, prompting a model already means a good deal more than typing words into a text-entry window. So-called prompt engineering, an inexact science to be sure, takes as its core mission the design and development of methods by which to instruct, cue, or otherwise initiate the process of language generation so as to result in “good” output. What constitutes the good, and how it is identified and evaluated, are matters to which we return below (and here we revert to the authorial “we”).

It is difficult to avoid further mention of ChatGPT, although we would prefer it if we could because the term, perhaps as a consequence of insistent repetition, has come to mean nothing, something, and everything, all at once. For all its semantic ambiguity and imprecision, the term AI does at least summon a speculative, ordering imagination that sweeps the ethicopolitical into the domain of science and informs discussions of bias, fairness, mental health, trust, safety, open systems, a data commons, and public resources. In contrast, the term ChatGPT, precisely because of its common use as shorthand for the broader sociotechnical condition we have been sketching, has the effect of short-circuiting thought itself—it is referential and self-referential, all at once. Practically, ChatGPT refers to a chat application built on top of an LLM (the Generative Pretrained Transformer), the technocultural authority of which derives in part from its accessibility and ease of use. A chatbot's end-user interface is indeed more accommodating than OpenAI's “playground,” which allows users to control hyperparameters such as temperature, frequency penalty, and max tokens (even using language such as this risks losing a mass audience; hence the appeal of an interface with only a basic messaging window and familiar menu options). But there already have been, and will be, many more chat applications and application programming interfaces integrated with, or otherwise supported by, many more LLMs, all of which will close off access in the name of granting access. Indeed, ChatGPT's position at the top of the proverbial leaderboard has already started to give way to other models and applications, and there are a number of scenarios in which it simply disappears.Footnote ⁵ Nonetheless, the term ChatGPT continues to index a host of objects, agents, activities, behaviors, and attitudes that extend well beyond the strictly technical domain.

We return then to the question of how the profession can best position itself within this milieu. On the surface, the new scene of writing—that is, language processing—may seem to be composed simply of permutations of familiar activities, processes, and concepts: symbolic representation, semantic relationships, corpus construction, translation, indexicality, the evaluation of output, authorship, and generic forms. In this version of the class, there is at least still a text. So too with word embedding, the computational representation of words with numbers—more precisely, the representation of words in a continuous vector space by assigning them numerical values, or coordinates—which is fundamental to computational text analysis as it has been historically leveraged within the digital humanities.

What is less familiar about the ecosystem of LLMs, and not often remarked in a context such as this, is subword tokenization, one of the crucial NLP techniques. Subwords are as they sound: not words but rather parts of words that bear no necessary relation to the linguistic units identified and categorized by human sciences. They can be morphemes, but not necessarily or exclusively. A good example is th: statistical analysis of any English-language corpus would probably capture the frequency of this pairing and identify it as a subword token, by means of Byte Pair Encoding. Digraphs are not of course unrelated to writing as it has historically been understood—after all, morphological rather than linguistic principles can inform the typographic element of the ligature. But the ligature is only about appearance, and it presents no challenge to the operative principles of language. Subword tokens, in contrast, result from a process of segmenting words within a corpus according to statistical or rules-based criteria. These segmented tokens are then used to construct what are called vocabularies (that is, collections of parts of words) that help a model learn the statistical representation of text data. And here the names of tokenization algorithms and libraries such as WordPiece and SentencePiece are particularly suggestive. Perhaps linguistic protocols have been eclipsed by arithmetic means. Perhaps the leading sciences and technologies are no longer operated and mediated by language as such. And perhaps we truly are at the beginning of something new.

At the 2024 MLA convention, “Reading Generative AI: Theory, Data, Critique” was one of three pilot sessions convened to test a new format: professional development seminars “meant to catalyze scholarship on research topics that cross areas of study, disciplines, methods, and perspectives” (“MLA Convention Seminars”). After submitting abstracts the preceding spring, a dozen participants were selected. They had proposed work on a wide variety of questions and problems related to the topic at hand and represented a range of institutions, approaches, and career stages. Attention then turned toward individual and collective preparation for the seminar, which included many months of drafting and peer review. Once in a room together in Philadelphia, we engaged in a spirited three-hour discussion, one whose breadth, depth, and vibrancy provided the impetus for the cluster of essays that appear here.

Of course this was not the only conversation about AI at the convention: indeed, there were enough dedicated panels and individual papers that it would not have been possible to attend them all. But we recap these details about the professional development seminar because neither its mode nor its deployment at the convention is incidental to the broad sketch we offer in the preceding paragraphs. The three initial topics—generative AI, prison literature, and the medical humanities—all suggest new directions for research, teaching, and even activism. The seminar format itself is an experiment occasioned in part by the new realities of the MLA convention: with the displacement of scant hiring activity to remote platforms, as well as the staggered recovery from the pandemic and the reduction in travel support for nearly all its constituents, the convention seems clearly to be seeking new ways to position itself as the chassis for the profession(s).

This journal is another such chassis. In the context of a previous Theories and Methodologies section on distant reading and computational approaches to literary studies (May 2017), a language model would have been understood as a statistical profile of a text or corpus of texts. This description holds true for LLMs as well, but the scales are vastly different; GPT-4, for example, reportedly consists of 1.76 trillion parameters (that is, variables that determine the model's behavior and capabilities). What this indicates is not yet another novel computational application or approach but rather a general condition of language and life (or, again, an episteme). If some of the public enthusiasm around digital humanities has cooled in recent years, it is perhaps because it has been a victim of its own success, at least insofar as the kinds of “theories and methodologies” previously singled out for special attention are increasingly regarded as altogether ordinary, if not normative, within the disciplines and subdisciplines they serve. By contrast, whatever uses LLMs may prove to have within domain-specific literary and historical research will be well downstream of the far more basic challenges posed by the technology and its market uptake, yielding institutional configurations that may hold out little promise of exceptional status for even the most digitally forward humanists.

Some sense of urgency about the emerging situation is evident in the fact that just two months after the public release of ChatGPT, the MLA moved to partner with a peer organization to convene the MLA-CCCC Joint Task Force on Writing and AI. The charge included “[t]aking stock of the current state of the issue and identifying implications for teachers, students, organizations, and scholars.”Footnote ⁶ Since January 2023, the task force has released two (soon to be three) public working papers, offered multiple webinars, and curated a selection of instructional materials. All this activity is attuned to the exigencies of the moment, and yet it also—and precisely not paradoxically—conforms deeply to scripted organizational norms. At the same time, the task force is not a committee or standing entity of either organization, which means that its working papers (note the provisional and amorphous form), while sponsored and reviewed by the two executive boards, do not carry the weight of official policy. Thus, the agility demanded by the moment also begets a deferral of agency and responsibility.

The task force itself, we wish to be clear, has done vital and necessary work (one of us, Matthew Kirschenbaum, is a member); but like a seminar or a journal feature, the mode is symptomatic. However critical or otherwise aligned with the profession's values and expressed commitments, the responses to AI largely subsist within a bounded and belated repertoire, a circumstance from which, we hope we have made clear, we do not seek to exclude ourselves. There has been an outpouring of enterprising activity across the institution of the university, all of it (one imagines) capturable by the metrics of performance review. That there is actual and symbolic profit to be made (in whatever small measures of reputational currency, honoraria, and grant monies) is not incidental or too unbecoming to acknowledge. But transactions on the level of the individual are magnified in larger disciplinary economies as we all conceive new curricula and assignments, new certificates and degree programs, all straining to meet the moment by demonstrating “leadership” once again.

With that as a substrate, we now relate, in highly stylized terms, the principal disciplinary operations functioning in correlation with LLMs and generative AI. We adopt as our own rhetorical device the model, figuratively speaking—in fact a suite of figurative models fine-tuned (as it were) from our disciplinary foundation models. The foundation model itself, we should say, is the true heir to the metanarrative, being the de facto prescriptive, closely guarded, and contested core from which commercial downstream models are spawned for individuated tasks and applications. Here then we put forward some imperfectly rendered, inevitably nondiscrete, and appropriately idiosyncratic models of models, presenting in each instance first the foundation model and then, after the arrow, its fine-tuned elements, followed by a short gloss. Whether, how, and to what extent they may prove sustainable through the institutional transformations we outline in the conclusion of this essay must remain an open question.