To Build AI Literacy, Return to the Humanities
Why reading and writing skills matter more than ever in the age of AI
The following post from Marit MacArthur is a condensed version of her article that appeared in the journal, AI & Society, earlier this summer. The article and post make a strong argument for why human expertise is not likely to be devalued anytime soon, and why the process of writing (no matter what we call it) and related forms of critical reading and editing are still the most powerful tools for developing human expertise.
Marit J. MacArthur is a continuing lecturer in writing at the University of California, Davis. Her pedagogical research on generative AI and writing is informed by a decade of collaborative, interdisciplinary research in the digital humanities, applying an open-source LLM for speech recognition, with funding from the American Council of Learned Societies, the National Endowment for the Humanities, and the Social Sciences and Humanities Research Council (SSHRC) of Canada. She is a Principal Investigator for PAIRR (Peer and AI Review + Reflection), funded by the Learning Lab’s AI Grand Challenge program, with seven partner institutions in the California State University and California Community College systems. She is also a co-investigator with the NEH-funded Center for AI and Experimental Futures (CAIEF) at UC Davis.
Last year, “slop” made the short list for the Oxford Word of the Year, losing out to “brain rot.” Slop was defined as “art, writing, or other content generated using artificial intelligence, shared and distributed online in an indiscriminate or intrusive way, and characterized as being of low quality, inauthentic or inaccurate.” “Brain rot” (coined by Henry David Thoreau in 1854) is now defined as “the supposed deterioration of a person’s mental or intellectual state, especially viewed as the result of overconsumption of material (now particularly online content) considered to be trivial or unchallenging.”
Consumption of too much sloppy AI-generated text [which I purposefully do not refer to as “writing"] may rot the brain.
Built-in Hazards of AI-Generated Text
If we hope to leverage AI for the common good, yet fear widespread brain rot or cognitive overload by AI, should we feel reassured by the widespread apprehension of slop? No, because glaringly obvious slop is not the real challenge. The real challenge arises when AI-generated text seems correct and appropriate but is not—a sort of uncanny valley of seemingly authoritative text—which can fool novices or experts-in-training, and escape the attention of experts who, in haste, fail to notice or point out its flaws.
The current generation of professional experts, educated without access to all-purpose chatbots, may be adept at distinguishing accurate, appropriate AI-generated output from its close but inapt semblance —when, that is, they take the time to critically assess the output. But what about future generations? The answer will depend on how we educate them. Our task is to continue to train students in the traditional-but-not-outmoded literacies—reading and writing—and to cultivate future experts in the disciplines and professions. In carrying out this task, we must discourage overreliance on AI in reading and writing, because that would eventually render users unable to judge the quality of AI output, something recent research suggests may already be happening in professional fields.
If our educational systems can make a significant shift, to reinvest in the humanities and promote humanities-based education and training, we might feel more optimistic about today’s school children developing the ability to spot slop, and to distinguish appropriate text or code from a close but inapt semblance. This important educational task—re-emphasizing humanities-based education in the age of AI—is hiding in plain sight behind misleading terminology such as prompt engineering, which suggests that students need to study engineering, not writing, in order to collaborate with AI. Such misleading terminology is not merely a stylistic choice or a bothersome symptom of our technocentric culture. The stakes are higher than that.
The Need for Human Writers
The guidance I offer is not a naïve, self-interested attempt on behalf of the humanities to stem the flow of funding away from “us” toward STEM, which generative AI has accelerated. In “The Programming Era: The Art of Conversation Design from ELIZA to Alexa,” Christopher Grobe demonstrates the central role of “aesthetic experts” in transformative technological developments we normally attribute to computer science. In his research on the development of virtual agents, he interviewed many “writers … and performers” who wrote the scripts for virtual agents like Siri and Alexa and designed the ways that they converse with human interlocutors. Yet, Grobe points out, software developers often underestimate and misunderstand the skills and knowledge of writers and artists. Indeed, they habitually misconstrue writing as its polished, fluent end-product, rather than as the cognitive and rhetorical process it is. Many confidently assert that when a chatbot produces text or code, what it is doing is writing.
Writing, reading and understanding a text that represents disciplinary expertise are very challenging, time-intensive cognitive tasks. Critically assessing that text is also a very challenging, time-intensive cognitive task. Prompting a chatbot to generate slop is quick and easy. Prompting it to generate context-appropriate text, critically reading it to judge whether it has done so, and further editing either the prompt or the output to match one’s goals, is not.
So what happens when we ascribe the genius of generative AI solely to computer science, rather than highlighting the deep dependence of LLMs on “training data”—a misleading euphemism for a vast trove of expert human writing? For one thing, aesthetic experts fail to receive credit or compensation for their intellectual property. For another, we miss the importance of humanities-based education in adapting to generative AI—and thus may fail to equip college graduates with the skills that the present and future workforce demands.
“Training Data” is Writing
Training data, for an LLM, is human expertise captured in writing, which must be continually developed and curated by humans. The term “training data” minimizes the tremendous effort and resources necessary for humans to develop expertise that is then fed to LLMs. As we know, the LLMs that underlie chatbots excel at engaging in and simulating humanlike dialogue because they were trained—using deep machine learning techniques that imitate the neural networks of the human brain— on a massive amount of text written by humans, notably from Common Crawl.
What’s deemphasized, by tech companies who want to keep “training data” free is that LLMs and all future AI tools and platforms will necessarily continue to rely on new human expertise, captured in writing or deployed as critical reading and editing skills, in every field of human endeavor (see for example, extensive hiring in so-called “data annotation"). Referring to human expertise captured in writing as “training data” obscures this fact—especially if we mistake AI output as writing, rather than probabilistic predictions of the likeliest string of tokens that the prompting human asks the chatbot to produce, based on the human-expertise-captured-in-writing that the chatbot was trained on.
The generation of AI output depends on prior human thinking and writing, and on a continued supply of human expertise—the right corpus of texts or code written or heavily edited, developed and/or curated by humans—as “training data.” The safe and appropriate application of that AI output to any task, in turn, depends on human expert “evaluation”, grounded in disciplinary expertise and rhetorical awareness.
Prompt “Engineering” Is Prompt Writing
And what would we gain by calling prompt engineering what it indisputably is—writing prompts for GPTs? Perhaps lower salaries, since “engineering” job titles are currently compensated at much higher rate than “writing” job titles. Nevertheless, Big Tech seems to recognize the economic importance of writing skills in collaborating with AI to some extent. OpenAI’s Logan Kilpatrick declared at the end of 2023 that “reading, writing, and speaking … [are] the 3 underlying skills” needed to succeed in “a world with AGI [artificial general intelligence].”
Prompting AI, however, is not the same thing as communicating with a human being—it is harder. Though the LLMs have ingested an enormous amount of specialized knowledge encoded in human-authored texts, they are, ultimately, computer programs. There are defined limits on the local rhetorical context they can assess, in part because they rely on language alone. The ability first to understand one’s immediate context, and then to precisely describe in words, is essential to writing effective prompts for AI, and then to evaluating the output and refining it—either by editing that output directly or further prompting AI to refine it.
Accordingly, I offer a redefinition of prompt engineering: Writing instructions for a probabilistic text-generating (or code, image, audio, etc.-generating) machine, detailing the immediate “rhetorical situation”—the purpose, audience(s), guidelines for the genre, and the real-world, local context—which the machine cannot access without human guidance, because the LLM is not embedded in the immediate physical, social, cultural, political world in the which the prompting human lives.
This definition, I hope, clarifies why disciplinary experts are adept at writing prompts that elicit relatively appropriate AI output, and at evaluating and refining that output, while novices struggle to do so. Novices often lack either a sophisticated understanding of their immediate context, genre, audience, etc., or lack the ability to describe these things clearly in words. This understanding of prompt engineering as prompt writing also has consequences.
Specialized chatbots might be trained—by human disciplinary experts who understand specialized rhetorical contexts—to prompt novices to articulate their immediate context, up to a point. This will be an easier and more successful process if we understand that prompting is writing, not engineering, and look to the humanities, not computer science, to train the future workforce to adapt to AI. If we fail to do so, many students will grow up to be unemployable. Others will become hazardously incompetent employees. They will apply prompts they cannot write and do not understand into AI platforms that in turn produce outputs they cannot evaluate. And they will apply these outputs to a wide range of tasks, with risks they cannot assess or even imagine.
Teaching Writing in the World of LLMs
For better or worse, generative AI has elevated critical reading and writing as the most important foundational skills to develop across and within disciplines in schools and universities. If we hope that future generations can grow up to critically assess and edit AI output from a position of disciplinary expertise, keeping in mind their distinct purpose, audience, genre and context—they will need to keep learning to write without a ghostwriter (or ghost coder) looking over their shoulders. To do that, AI must be limited to a tutorial role: to develop disciplinary expertise, students must go through the struggles and rewards of learning. And they must be encouraged to value their own intellectual property, so that they think twice before feeding their writing, or anyone else’s, to a chatbot hungry for more “training data”.
To learn more, please follow this open-access link to the full-length article, “Large Language Models and the Problem of Rhetorical Debt,” published in AI & Society in June 2025.
I get the point about prompt engineering being prompt writing, but I wonder about voice interfaces, which more and more people are using. It's all about communicating clearly, I suppose. Would love to know your thoughts.