Guest Forum: Nicholas Birns “On Generative AI, Prompt Chaining, and Resisting Epistemological Indifference”

Taken as a whole, ChatGPT’s generalizations about scholarship are both too hyperbolic and too haphazard to command real authority. Clearly, ChatGPT cannot draw nuanced inferences about research expertise from its training data. Instead it favors laudatory generalities. In this context it is worth repeating that ChatGPT is not a database or a continuously updated crowd-sourced resource like Wikipedia. Nor (see Shah and Bender 2022) does it operate like conventional search engines that, at their best, provide links to sources that are algorithmically predicted to represent the most relevant, popular, and/or reliable sources on the internet. As Shah and Bender explain, the statistical modeling of training data privileges the quantity of patterns in data, with no direct means of signaling the quality of those patterns. Far from providing any thoroughgoing percipience to its users, when ChatGPT becomes a “go to” resource for information, what gets sacrificed is the plurality and serendipity that more time-intensive and unpredictable research processes tend to yield.

In so doing, systems like ChatGPT suppress the “friction” that is integral to strong habits of information retrieval (Allison and DeRewal 2024). My own experiments suggest that is certainly true in the field of literary criticism. Research in the humanities should yield a sense of multiple possibilities that may, between themselves, be irreconcilable and yet testify, in their range of conjecture, toward a heuristic interpretive field. By contrast, ChatGPT promotes a consensus that smooths out the truly generative contradictions. Moreover, with its tendencies to privilege the dead and out-of-date, ChatGPT, at least as far as literary criticism goes, is more likely to provide stale, tendentious, and misleading ideas than an undergraduate’s random pass through a Google Scholar search or—dare I say it—a look at the library shelves. A carapace of statistical flattening prevails over the messier and more vital plurality of information that is actually, or at least potentially, at our disposal. The presumed “ontology” of ChatGPT turns out to be no ontology at all.

“Writing” Without Writing

Given this inescapable, if inevitably disappointing reality, how are we to frame and mediate the effect that ChatGPT and its correlates will doubtless exert on humanities disciplines such as literary criticism? One immediate ramification, I suggest, is to concertedly prioritize acts of primary reading and writing. The experience of querying ChatGPT for information reminded me of an undergraduate professor of mine who advised her students to set aside secondary criticism. “If you have time to read anything else but the work you are analyzing, read another primary work of the same author or period.” My professor was not disparaging the value of secondary criticism or being deliberately formalist. She was instead assuming that undergraduates would probably not know how to find the most cogent criticism in the library. And she believed that secondary work, even if well-done, could distort our perception of the primary work. She was telling us to wait to make any foray into the secondary work until we had learned to fully engage with the primary work. Although I did not always obey her injunction, I well understood the logic behind it. A student’s most meaningful thinking about any object of analysis should, to the extent possible, not begin by being derivative of some other thinking. Eventually, one learns the habit of reading secondary criticism (as one does everything else) critically. But those habits take time to cultivate.

By contrast, ChatGPT crystallizes a kind of potted, half-outdated version of the conventional wisdom. In this way it makes palpable what Lionel Trilling (1950: 206) called “the hum and buzz of implication.” Trilling was referring to what was “in” or accepted: playing with ChatGPT on matters of literary criticism—and much else, I suspect—makes one aware that said hum and buzz are better off implied than realized. To be sure, people are encouraged to go to these mechanisms to save time. But this effort to minimize effort may end up costing far more than what any student or researcher should be willing to pay. In her evaluation of NotebookLM as a tool for research, Tiffany DeRewal (2025) shows how such products sidestep the cognitive and experiential trajectories of process through which so much genuine learning and research occurs. As with most generative AI applications, the most adept users are likely to be experienced scholars, fully and arduously trained in the art of research and analysis. How will society sustain such practices and produce the next generation of scholars if higher education loses the war on “friction”?

The advent of LLM-based text generators is not just an opportunity to defend literary thinking but also to understand what literary thinking is and does. Though the field has its own issues with trends and herd mentalities, literary criticism has historically had little patience for the parroting of simplified consensus positions—much less the goal of conducting “research” at the touch of a button. In my current work in progress, my experiences with ChatGPT have led me to thinking about how literary criticism, going back to Aristotle, is in the realm of production—in other words, of action. When critics actively analyze the production of others, they undertake a mode of production themselves. Without being purely functional or utilitarian, words are deeds to literary critics. If, in Kathy Eden’s words (1986: 34), literature is part of “the practical or active life,” literary works can practically operate in the realm of causation. But merely to observe the difference between the literary critic and the “stochastic parrot” (Bender et al. 2021) is not to sufficiently contend with the challenge of LLMs. That these systems use words as their medium makes any assessment of post-generative AI textuality a tangled one.

In contrast to the earlier effects of the internet, generative AI is not just a computerized information technology, but also a technology that renders legible text in the style of human-wrought representation. For example, with only minimal instructions from a human user (perhaps by means of prompt chaining), generative AI can produce something that readers could recognize as a novel. I set aside whether anyone would knowingly wish to read such a novel (though we know that plenty of people are right now duped into buying them on sites like Amazon which allow them to be passed off as the work of favorite author; see Jones [2025]). For now, I am more interested in recalling Chiang’s point about the machine incapacity for metacognition; ChatGPT’s capacity to “write” a “novel” points to a fundamental incapacity. I don’t mean that one could not prompt ChatGPT to generate a novel in which, say, a generative AI system reflects on generative AI or shares its “thoughts” on any other topic. But while generative AI is perfectly capable of going “meta” in this performative way, it cannot take account cognitively of its own representative action.

To make this point clearer, consider the work of Alexander Manshel (2020), who has argued that, relative to the fictional depictions of earlier media such as radio and television, the fictional depiction of the internet and mobile devices in comparable early twenty-first century fiction is “uncannily rare.” “Though digital technology is now instrumental in the research, writing, publication, marketing, purchasing, and even reading of contemporary fiction, it remains largely absent from literary fiction itself” (53). What the argument overlooks, however, is the possibility that the novelists of this era have consciously chosen how to represent this technology, through a deliberate reduction in degree. That is, a human author might set her novel in, say, 2007 but consciously underrepresent the presence of the internet at that time for strictly novelistic reasons. Yet, such a novel would reveal more interpretive knowledge about the internet, than if one instructed ChatGPT to produce a novel of so many words, set in 2007, which depicted the modes of online use available at that time, even though the latter text would almost certainly discuss the internet quite explicitly. Put differently, LLMs are unable to really “write” a novel in the way that human authors do but their ability give the impression that they can explodes the differential wall between literary representation and depiction of technology.

Whereas radio, television, and the internet diverted some literate people from the act of reading fiction (and perhaps reading anything), the marketing and implementation of text generating chatbots pose the danger of diverting many people who communicate in text from the act of writing as I conceive it (see also Kitzinger 2024). In this sense generative AI is not simply a media advance of the kind that interests Manshel. It is also an instance of what Dustin Edwards (2025) calls “digital damage.” Edwards shows that the building of centers for data processing and mining inflicts significant, tangible harm to land and water resources. I believe that the chatbotization of writing could inflict parallel damage on the cognitive and imaginative resources of young people today and future generations. Some may quibble about the extent of the damage and they may quibble about the meaning of “writing.” Nevertheless, a mechanism with the apparent ability to generate textual content that rhetorically resembles human writing but which, in reality, emerges from a wholly or largely automated process will, at the very least, require critics (among others who study writing) to think much more about the advent of legible bodies of textual content borne of cognitive absence and epistemological indifference.

It is still too early to know how far the adoption of generative AI will take us given that there are already signs of a backlash. I believe that one reason for the anxiety about generative AI undermining the literature classroom is that, as a society, we have not, for some time, valued literature or literary criticism nearly enough. In particular, we do not see literary criticism as an act of production. The most optimal way to counter epistemological indifference, I contend, is not just by knowing, but by doing; by taking action.

In the literature classroom that action begins with reading. That many people have come to believe that reading a summary of a literary work is somehow sufficient seems implicitly to mark a lack of understanding of what it means to read a literary work—questions that literary critics may need to take up anew. What critics such as Terry Eagleton (1983) marked forty years ago as hostility to theory has seemed to degenerate into a full-blown hostility to reading books that predates chatbots by at least a decade. Cryptocurrency magnate Sam Bankman-Fried infamously commented that he never read anything longer than a blog post of a few paragraphs (see Roberts [2022]). His comment was edgy to be sure, but, sadly, hardly a complete outlier. Now the developers of commercialized chatbots seek to “solve” the problem of “writing” about what one has never actually read through the technically mediated recycling of consensus-driven textual fodder.