Rhetoric After Search: Composition in the Age of AI

A palimpsest: earlier writing partially erased beneath new text, a fitting image for how cultures layer one mode of communication over another.

When a society’s main way of communicating changes, the culture changes with it. The shift from oral storytelling to writing reorganized how people made and judged ideas. As Eric Havelock notes, orality tends to place ideas side by side (parataxis), while writing lets us order and rank them (hypotaxis). Later, the internet unsettled those hierarchies by putting everything in the same feed. Now generative AI pushes a new turn: instead of going out to find information, we ask a model to synthesize it for us in real time.

The shift from a largely static print culture (books, journals, newspapers) to the dynamic, hyperlink-laced world of the internet (posts, tweets, comments, videos, remixes) is instructive. If print stabilized hypotaxis—codified hierarchies of knowledge—the internet reintroduced powerful currents of parataxis, the flattening of ideas. Feeds place headlines, memes, and research side by side; comments appear co-present with reported stories; search results level institutions and hobby blogs into a single scroll. The effect isn’t a simple “reversion” to orality, but a hybrid: an always-on, text-heavy environment that nonetheless rewards immediacy, performance, and identity signals. We might call this the era of networked parataxis or feed culture. Authority did not vanish, but it was continuously jostled—ranked, re-ranked, and sometimes drowned—by the drumbeat of the new.

Now another shift is underway: from the internet as a place we go to an intelligence we bring to us. Generative AI reframes the web not as a destination but as a substrate for on-demand synthesis. Instead of clicking outward into a maze of links, we prompt, and the system composes a provisional text from learned patterns; a palimpsest of the internet, re-generated each time. In this sense, the interface transitions from navigation to conversation; from retrieval of artifacts to production of fresh, if probabilistic, prose.

What does this do to our rhetorical environment?

First, generative systems appear to restore hypotaxis, but of a different kind. Where the feed set items side by side (parataxis), AI models arrange them within a single, coherent utterance. Citations, definitions, warrants, and transitions arrive pre-braided, often with a competence that flatters the eye. Synthetic hypotaxis. Yet because the underlying process is statistical and unobserved, it risks performing coherence without guaranteeing evidence. The prose feels orderly; the epistemology may be wobbly. We are handed an essay when we might have needed a bibliography.

Second, generative AI re-centers dialogue as the controlling framework for knowledge work. Search terms give way to prompts, and prompts invite follow-ups, refinements, and counterfactuals. The standard unit of knowledge work becomes a conversation. This recovers something like the agility of oral exchange—call-and-response, iterative clarification—while living in a textual medium. In practice, this hybrid looks like scripted orality: improvisational yet instantly transcribed, searchable, editable, and archivable.

Third, the locus of authorship drifts. With the internet, we cited and linked; with AI models, we consult and compose. The user becomes a curator-designer, someone who specifies constraints, tones, examples, and audiences, while the model performs the heavy lifting of first-pass drafting and rephrasing. Our artifacts increasingly feel like bricolage: human intention wrapped around machine-generated scaffolds, tuned by promptcraft and revision.

Likely effects of the shift

Positive

Acceleration of synthesis. Students and researchers can pull together working overviews in minutes, explore counterpositions, and translate among registers or languages. This lowers the activation energy for inquiry and can widen participation.
Adaptive scaffolding. Models can perform as low-stakes tutors or writing partners, offering just-in-time explanations, outlines, and examples that match a learner’s current academic level.
Access workarounds. For people blocked by jargon, gatekept PDFs, or unfamiliar discourse conventions, generative AI can paraphrase, summarize, or simulate genres they need to enter.

Negative

Source erasure and credit drift. The move from links to syntheses obscures provenance. Without strong citation norms and tools, authority blurs and labor disappears into “the model.”
Confident misstatements. Synthetic hypotaxis can launder uncertainty; tidy paragraphs can mask speculative claims (or hallucinations) behind elegantly connective prose.
Homogenization of style. Fluency becomes formulaic. If everyone leans on the same engines, we risk a median voice—competent, placid, and forgettable—unless we deliberately cultivate voice.
Skill atrophy. If we outsource invention, arrangement, and revision too early or too often, we can lose the slow muscle of drafting, comparing sources, and building warrants from evidence.

Neutral/ambivalent

New genres, new shibboleths. Prompts, system messages, and “prompt-sets” become shareable teaching artifacts; AI marginalia (notes explaining how output was shaped) may emerge as a norm. These could deepen transparency, or become ritual theater.
Assessment realignment. If first drafts are cheap, assessment shifts toward process evidence (versions, notes, prompts), oral defenses, and situated tasks. This can improve authenticity but demands more from instructors.
Attention economics. Conversation-first tools reduce tab-hopping, but they also reward rapid iteration. Some users will become more focused; others will live in an endless loop of “one more prompt.”
Institutional enclosure. Organizations will build bespoke models and walled knowledge bases. That can improve reliability for local use while narrowing horizons and reinforcing house orthodoxies.

So what do we call this era?

If the internet cultivated networked parataxis, generative AI installs a layer of synthetic hypotaxis, or structured language on demand. I’m partial to naming it consultative literacy (to stress the dialogic nature), or generative rhetoric (to mark how invention and arrangement are becoming collaborative). Whatever we call it, the practical task is the same: pair the speed and plasticity of AI with disciplined habits of citation, verification, and style. In other words, keep the conviviality of the feed and the rigor of the page, and teach writers to orchestrate both.

The culture will follow the mode. As we move from going out into the web to inviting the web to speak through us, our work becomes less about locating information and more about shaping it: specifying constraints, testing outputs, insisting on sources, and cultivating voice. That is both the promise and the peril of an age where every prompt yields a fresh, provisional world.

I hope AI doesn’t just become an elaborate advertising tool, like social media

For the last 50 years, the most significant technological advances have been in communication technology. We’ve developed incredible social media platforms, digital tools, and mobile technologies that have connected nearly everyone in the world in unprecedented ways. It is now easier than ever to communicate across vast distances, through diverse mediums, with anyone, anywhere.

This evolution in communication technology stands in contrast to the prior era of technological progress, which focused primarily on mechanical innovations. Before the Internet, breakthroughs were largely in the mechanical and physical domains—electricity, appliances, plumbing, transportation, and medical technologies. These advances transformed the physical and material conditions of life, whereas the digital and communication technologies of the last half-century have primarily transformed the social and symbolic: how we share, transmit, and receive information.

So where does AI fit within this framework? On the surface, AI seems to belong to the realm of communication technology, given its reliance on digital infrastructure and its ability to process, analyze, and generate information. However, AI is not just another faster, flashier app, nor is it simply an incremental improvement in how we connect or communicate. AI represents a qualitative shift—different not in degree but in kind. It’s not merely a tool to expand or refine existing communication technologies; it introduces entirely new capabilities: the ability to generate, interpret, and act on information autonomously, at scale, and with unprecedented sophistication.

This distinction raises an important question: What will we do with this power? Historically, the revolution in communication technology has largely been co-opted for one purpose: advertising. Social media, search engines, streaming platforms—all these breakthroughs have, at their core, been monetized by perfecting the art of targeting and persuading audiences to buy, click, or consume. The immense potential of communication technology has often been reduced to serving the interests of advertisers, prioritizing profit over other possible uses.

Will AI follow this same trajectory? Will its vast power be funneled into refining micro-targeted advertising and making marketing even more efficient? Or can its unique capabilities be directed toward broader, more meaningful purposes? AI has the potential to reshape education, healthcare, governance, and creative expression in ways that go far beyond commercial exploitation. But realizing this potential will depend on whether its deployment is driven by ethical considerations and the desire for collective good—or by the same profit motives that have shaped the digital landscape over the past half-century.

AI represents a turning point in the history of communication technology. It is not just a tool for transmitting or refining messages; it has the capacity to generate new knowledge, discover patterns we cannot see, and even challenge human creativity and decision-making. The real question, then, is whether we will seize this moment to redefine the role of communication technology in society or let it become yet another means to sell more products more effectively. The answer will determine whether AI becomes a transformative force for good or merely the next iteration of a decades-old advertising machine.

The Central Tendencies of the Rhetoric of AI

As artificial intelligence increasingly generates the written and published text we consume, it’s worth considering the consequences on both individual and societal levels. On the micro level—the everyday use of AI in writing—I suspect the changes will be subtle but meaningful. Individual writing abilities are likely to improve, as AI tools act as an accessible public option for crafting coherent prose. Just as autocorrect has quietly raised the baseline for grammatical accuracy in text messages and online posts, AI tools will elevate the overall quality of written communication. AI will make polished, coherent writing accessible to more people, effectively raising the “floor” of writing ability.

On the macro level, however, the implications are more profound. To understand this, let’s consider three primary dimensions of rhetoric: syntax, vocabulary, and tropes. These dimensions encompass how sentences are structured (syntax), which words are chosen and how they’re used (vocabulary), and the creative use of rhetorical devices like metaphors or antithesis (tropes). Since AI operates by analyzing and replicating patterns in language datasets, its writing reflects the statistical tendencies of its training data. In other words, AI-generated text is governed by the same central tendencies—mean, median, and mode—that define any dataset.

Syntax: The Median Sentence

AI-generated syntax will likely gravitate toward a median level of complexity. Sentences will neither be overly elaborate nor starkly simplistic but will instead reflect the middle level of grammatical intricacy found in its training data. This tendency could lead to a homogenization of sentence structure, with AI producing text that feels competent but not particularly varied or daring in its syntax.

Vocabulary: The Modal Words

Vocabulary choices in AI writing are often dictated by the most common words and phrases in its dataset—the mode. This preference for the most frequent linguistic elements means AI text can sometimes feel generic or boilerplate, favoring safe, widely used terms over more distinctive or idiosyncratic language. While this might ensure accessibility, it also risks a flattening of linguistic diversity, where rarer or less conventional words are underused.

Tropes: The Mean Creativity

When it comes to rhetorical tropes, AI tends toward the mean—a sort of average level of creativity. It might generate metaphors or analogies that are effective but lack the originality or boldness that characterizes the most memorable human writing. The result is a tendency toward competent but predictable creativity, rather than the kind of transformative or disruptive innovation that pushes rhetorical boundaries.

Language as Dataset

If AI treats language as a dataset, it inevitably inherits the statistical biases and patterns inherent in that dataset. While central tendencies like mean, median, and mode are useful for operationalizing numerical datasets, their application to language introduces a new set of challenges. Syntax, vocabulary, and rhetorical tropes may become increasingly tethered to these statistical norms, creating a gravitational pull toward a homogenized style of writing.

This is not to suggest that all AI-generated text will be devoid of creativity or variety. Rather, the concern lies in how the ubiquity of AI writing might influence broader linguistic and rhetorical trends. Will the prevalence of AI-generated text subtly shift our expectations of what “good writing” looks like? Will it amplify certain linguistic conventions while marginalizing others? These are questions worth monitoring as AI continues to shape the ways we write, think, and communicate.

If language becomes tethered to the central tendencies of AI’s datasets, the consequences extend beyond mere stylistic homogenization. They touch on the very dynamism of human expression—the outliers, the deviations, the unexpected turns of phrase that make language vibrant and uniquely human. Monitoring these tendencies isn’t just about understanding AI’s capabilities; it’s about preserving the richness of language itself.

AI Hallucinates; Humans Experience the Mandela Effect

One consequence of artificial intelligence is how it has revealed to us differences and similarities between human cognition and computational processes. “The brain is a computer” is a pervasive metaphor in cognitive science, since we process, store, and retrieve information in ways that sometimes feel mechanical, even computerized. Yet AI and humans engage with the world and use language in fundamentally distinct ways. Among the most noted differences is the peculiar phenomenon of “hallucinations“, AI-generated text that contains false or misleading information presented as fact.

But humans hallucinate too. We just named it something else: the Mandela Effect. I think the Mandela Effect–not simple errors or lies–is the human cognate of AI hallucinations.

The Mandela Effect describes a collective misremembering of historical facts or cultural artifacts. Named after a collective false memory of Nelson Mandela dying in prison in the 1980s, the phenomenon is widespread and seemingly cross-cultural. One famous example is the belief that actor Sinbad starred in a 1990s movie called Shazaam. In reality, no such film exists. This memory is likely an amalgamation of associations: Sinbad’s name, his comedic persona, and a vague recollection of similar movies like Kazaam starring Shaquille O’Neal. In these moments, the brain substitutes related but incorrect details, resulting in a memory that feels vivid yet doesn’t correspond to reality.

AI hallucinations follow a similar pattern. Generative AI systems compose text by predicting words based on associations learned from vast datasets. But sometimes, those associations fail to align with facts. The result is text that sounds plausible and coherent but is demonstrably false.

We know that hallucinations are an inevitability of AI, given that it generates text via associations and probabilities. But observing AI hallucinate also illuminates the mystery of the Mandela Effect in humans, as well as showcases the limitations of associative reasoning. Humans and AI alike depend on patterns and connections to make sense of the world. But when those connections are incomplete or misaligned, the result can be a version of reality that is “almost correct but not quite.”

This raises questions about how we navigate a world increasingly mediated by AI-generated content. If we rely on tools that describe reality through association, we risk adopting a “Mandelian lens” on the world—one where things are close to the truth but subtly warped. Over time, these small inaccuracies could compound and accumulate, shifting our collective understanding of reality in ways we may not even notice.

The analogy underscores a broader caution about generative AI. While these systems can assist with countless tasks, they do so by replicating human-like cognitive shortcuts, including our susceptibility to error. Recognizing these parallels helps us remain critical users of AI and reminds us of the human fallibility it often mirrors. Perhaps the most important lesson is that AI can reveal to us shortcomings in our own cognition by resembling and exhibiting our own errors. Our perception of truth—whether biological or computational—is shaped not just by what is real, but by how we connect the dots.

Artificial Intelligence is Artificial

The rise of large language models (LLMs) like ChatGPT has led many to believe we’ve entered an era of artificial general intelligence (AGI). Their remarkable fluency with language—arguably one of the most defining markers of human intelligence—fuels this perception. Language is deeply intertwined with thought, shaping and reflecting the way we conceptualize the world. But we must not mistake linguistic proficiency for genuine understanding or consciousness.

LLMs operate through a process that is, at its core, vastly different from human cognition. Our thoughts originate from lived experience, encoded into language by our conscious minds. LLMs, on the other hand, process language in three distinct steps:

Text is translated into numerical data, where words and phrases are assigned numerical values based on probabilities.
These numbers are plotted within a vast multidimensional space, representing relationships between words.
The model uses these relationships to generate new numerical representations, which are then retranslated into human-readable text.

This process is an intricate and compute-intensive simulation of language use, not an emulation of human thought. It’s modeling all the way down. The magic of such models lies in their mathematical nature—computers excel at calculating probabilities and relationships at scales and efficiencies humans cannot match. But this magic comes at the cost of true understanding. LLMs grasp the syntax of language—the rules and patterns governing how words appear together—but they remain blind to semantics, the actual meaning derived from human experience.

Take the phrase “public park.” ChatGPT “knows” the term only because it has been trained on vast amounts of text where those two words frequently co-occur. The model assigns probabilities to their appearance in relation to other words, which helps it predict and generate coherent sentences. Humans, by contrast, understand “public park” semantically. We draw on lived experience—walking on grass, seeing children play, or reading a sign that designates a space as public. Our understanding is grounded in sensory and conceptual knowledge of the world, not in statistical associations.

Finally, fluffiness is quantified along a scale

This difference is critical. What humans and computers do may appear similar, but they are fundamentally distinct. LLMs imitate language use so convincingly that it can seem like they think as we do. But this is an illusion. Computers do not possess consciousness. They don’t experience the world through sensory input; they process language data, which is itself encoded information. From input to output, their entire function involves encoding, decoding, and re-encoding data, without ever bridging the gap to experiential understanding.

To extend the analogy: a language model understands the geography of words in the way a GPS system represents the geography of the world. A GPS system can map distances, show boundaries, and indicate terrain, but it is not the world itself. It’s a tool, a representation—useful, precise, but fundamentally distinct from the lived reality it depicts. To say AI is intelligent in the way humans are is like saying Google Maps has traveled the world and been everywhere on its virtual globe; this is sort of true, in the sense that a decentralized convoy of Google cars equipped with cameras have indeed crawled the earth collecting visual data for Google street view, but is not literally true.

As we marvel at the capabilities of LLMs, we must remain clear-eyed about their limitations. Their proficiency with language reflects the sophistication of their statistical models, not the emergence of thought. Understanding this distinction is crucial as we navigate an era where AI tools increasingly shape our communication and decision-making.

The surprising power of n=2

We are enmeshed in data every day. It shapes our decisions, informs our perspectives, and drives much of modern life. Often, we wish for more data; rarely do we wish for less.

Yet there are moments when all we have is a single datapoint. And what can we do with just one? One datapoint offers almost nothing. It is isolated, contextless, and inert—a fragment of information without relationship or meaning. One datapoint might as well be no datapoint.

But two datapoints? That’s transformative. Moving from one to two is not just an incremental improvement; it is a fundamental shift. Your dataset has doubled in size, a 100% increase. More importantly, with two datapoints, you can begin to make connections. You can compare and combine, correlate and coordinate.

From Isolation to Interaction

Consider the possibilities unlocked by having two datapoints rather than one. A single name—first or last—is practically useless; it cannot identify a person. But a full name—two datapoints—suddenly carries weight. It situates someone in a specific context, distinguishing them from others and enabling meaningful identification.

The same holds true for testimony. A single witness to a crime might not provide enough perspective to reconstruct what happened. Their account could be unreliable, incomplete, or subjective. But with two witnesses, we gain a second perspective. Their testimonies can corroborate or contradict each other, offering a deeper understanding of the event.

Or think about computation. A solitary binary digit—0 or 1—cannot do much. It is a static state. But introduce a second binary digit, and the world changes. With two bits, you unlock four possible combinations (00, 01, 10, 11), the foundation of all logical computation. Every computer, no matter how powerful, builds its intricate systems of thought from this basic doubling.

The Exponential Power of Pairing

Why is the shift from one to two so significant? It is not simply the doubling of data, but the transition from isolation to interaction. A single datapoint cannot create relationships, patterns, or meaning. It is static. Two datapoints, however, introduce dynamics. They allow for comparison and combination, for movement between states, for a framework within which meaning can emerge.

This leap—from one to two—is the smallest step toward creating systems of knowledge. Science relies on comparisons to establish causality. A single experimental result is meaningless without a control group to measure it against. Literature and language depend on dualities—protagonist and antagonist, question and answer, speaker and audience. Even human vision is based on the comparison of binocular inputs, it is our two eyes that allow us to see depth.

AI and the Power of Two

The transformative power of n=2 is most recently demonstrated in the operation of generative AI. At its core, generative AI depends on the interaction of two distinct but interdependent datasets: the training data and the user’s prompt. The training data serves as the foundation—a vast repository of language patterns, structures, and examples amassed from diverse sources. This data alone, however, is inert; it is an immense collection of information without activation or direction. Similarly, a prompt—a fragment of input text provided by a user—is meaningless without context. It is a solitary datapoint, incapable of producing anything on its own.

When these two datasets combine, however, the true power of AI is unlocked. The training data provides a rich, multidimensional context, while the prompt activates specific pathways within that context, directing the AI to generate meaningful output. This dynamic interaction transforms static data into a creative process. Much like the leap from one to two datapoints, the relationship between the training data and the prompt enables the emergence of patterns, coherence, and utility. Without the prompt, the AI remains silent; without the training data, the prompt is purposeless. Together, they form a system capable of producing complex and contextually relevant language.

This relationship between training data and prompts underscores the profound significance of pairing, the power of n=2. The interaction between these two elements mirrors a broader principle: meaning arises not from isolation, but from connection. Just as two witnesses can construct a fuller account of an event, and two binary digits can enable computation, the union of training data and prompts enables AI to simulate human-like language and reasoning, creating systems that are both dynamic and generative. The leap from one to two here is not just a quantitative doubling—it is a qualitative transformation that makes the impossible possible.

Building Toward Complexity

Two is not the end point; it is the beginning. Once we have two datapoints, we can imagine three, then four, and so on, building increasingly complex systems. But we should not overlook the profound importance of the leap from one to two. It is the first and most crucial step toward understanding—toward the ability to identify patterns, make connections, and draw conclusions.

N=2 is the minimum threshold for meaning, the simplest structure capable of supporting complexity. From two datapoints, entire worlds of logic, creativity, and understanding can unfold.

Science is a process of elimination

Science affirms truth not by direct assertion but by negation. If someone were to ask a scientist, “How are you?” the response, in scientific terms, could not simply be “good.” A more fitting answer would be “not bad” or something equivalent. This distinction highlights a fundamental characteristic of the scientific method: it avoids direct affirmation in favor of ruling out alternatives. Science is not in the business of making unequivocal positive statements about reality but instead progresses by systematically eliminating what is not true.

This framework resembles the process of elimination in a multiple-choice test. For example, when scientists seek to answer a complex question such as “What causes cancer?” they rarely pinpoint a singular, definitive cause from the outset. Instead, they proceed by excluding possibilities—narrowing the field of potential answers by identifying what doesn’t cause cancer. Over time, these negations lead to an indirect approximation of truth.

In rhetorical terms, this mode of expression aligns with litotes, a figure of speech characterized by understatement. Litotes operates by asserting something indirectly, through the negation of its opposite. For instance, saying “not bad” rather than “good” captures a nuanced, precise meaning. Similarly, the scientific method uses this rhetorical approach to articulate findings, allowing for a careful and measured representation of truth that avoids overstatement.

The Rhetoric of Science

The rhetoric of science, a field dedicated to studying how scientists communicate and persuade, reveals that this litotic approach pervades the language of scientific inquiry. Scientists primarily communicate through their studies, and these studies often present findings in a litotic manner. Rather than offering unequivocal proof, scientific studies partially affirm hypotheses by negating competing explanations. In this way, scientific discourse functions less as a mechanism for declaring truths and more as a process of reducing uncertainty.

For example, scientific studies do not state definitively, “This is the cause,” but instead provide evidence that rejects—or fails to reject—the null hypothesis. This distinction underscores the probabilistic nature of scientific claims: science rarely deals in absolutes. Instead, it evaluates competing possibilities, gradually narrowing the scope of uncertainty by eliminating incorrect answers.

Science as Litotic Language

Contrary to popular perceptions of science as a taxonomic system that definitively names and categorizes truths, the language of science is fundamentally litotic. It constructs meaning by naming what something is not, rather than what it is. Each scientific study contributes a single datapoint that refines understanding by rejecting potential errors. Through this iterative process, science approaches truth indirectly, never declaring certainty but instead offering probabilities.

This litotic mode of expression reflects a broader reality: in science, certainty is a moving target, perpetually replaced by degrees of confidence. By articulating truths through negation, science not only mirrors the logic of litotes but also exemplifies its rhetorical precision. In doing so, it avoids the pitfalls of overstatement while offering a uniquely rigorous path to knowledge.

Human versus Artificial Intelligence

From my appearance on NPR’s The Academic Minute

Humans think in words, AI in numbers. The revolutionary Large Language Model ChatGPT works like a round of Family Feud: it answers our questions with only the likeliest responses, as determined by probability distributions. Is this “intelligence”? How should we understand truth in a world where words are assigned numbers, like the points in a Family Feud survey? 

We often think of science as taxonomic, but it’s not really. Scientific classification is negative and imperfect; it names by ruling out. Science says a mammal is an animal that doesn’t lay eggs. But what about the platypus?

In rhetoric this is known as litotes, a rhetorical device in which something is affirmed by negating its opposite, like if you ask how I’m doing and I respond, “not bad.” Paradoxically, this rhetorical approach can offer greater accuracy while granting less detail. 

Science is litotical. Frequently accurate but insufficiently detailed, scientific studies are limited to two types of negative findings; they either reject or fail to reject a hypothesis. This kind of knowledge is useful in a laboratory, but the real world has platypuses. Truth in the real world is more than the difference of everything it’s not.

AI is similar, for now at least. AI doesn’t name, it affirms by negation. ChatGPT sees the world as a multiple-choice question, and it responds through a process of elimination. Humans, meanwhile, fill in the blanks. We confront uncertainty not by calculating probabilities but by consulting wisdom. 

Each word generated by an AI represents the rejection of an alternative; artificial intelligence is fueled by probability rather than possibility. That’s a new world. Before it’s here, we should remember that human intelligence isn’t confined to the artificial horizon between rejecting and failing to reject hypotheses, and that our wisdom is deeper than its syntax.

Automated Writing Technology: Generative AI, ChatGPT

In grad school (2015-2020) I focused my dissertation research on what I termed “automated writing technologies.” I was fascinated by such technology because it represented a nexus of the linguistic and the numeric. These computer programs allowed me to explore my fascination with both alphabetic and quantitative languages. Specifically, my project tested the ability of computers to evaluate human-generated text and asked whether automated writing evaluation could, at the time, provide effective formative feedback on the rhetorical dimension of prose writing at the level of metaphor, irony, humor, and analogy rather than the more rote aspects of grammatical correctness. My dissertation findings, while limited, were not promising.

Large Language Models (LLM) were around during this time, but quite rudimentary. They were about as good as the predictive texting on our smartphones, limited to accurately predicting one or two words at a time. In 2017, however, researchers from Google published Attention Is All You Need, a revolutionary paper about the Transformer architecture that serves as the foundation of generative AI and which represents the “T” in GPT.

By the time I finished grad school in 2020, OpenAI’s second generation LLM, GPT-2, was available to the public. Now, less than three years later, we have OpenAI’s GPT-3 and ChatGPT, as well as other proprietary AI text generators from Google, IBM, Facebook, and more. This area of technology has advanced rapidly, and I am fortunate to have graduated when I did to ride this wave.

While I am not a computer scientist, and don’t pretend to be one despite my attempts to read and understand some of the primary papers from this field, I bring a background in rhetoric and writing assessment to the study of generative AI generally and Large Language Models specifically. My experience in writing pedagogy/assessment and the history of rhetoric prepares me to contribute to the scholastic conversations about these breakthrough technologies in productive ways.

I urge other humanities and liberal arts scholars to join this conversation. I think many of us in the liberal arts have much to contribute to decisions about how to incorporate generative AI into educational and professional institutions, which we need to start thinking about sooner than later. I also think we should make an effort to try to learn some of the more technical aspects of LLMs in order to be conversant with the programmers and companies designing them.

To that end, I will be updating this site mostly about future research regarding LLMs, GPT, and generative AI more broadly, considering both theoretical questions (does AI use rhetoric in the same way we do?) and practical applications (how can we use this technology to better teach writing and thinking?)

I recently gave a talk to my university about what LLM technology is and how we might work with rather than against it in our classrooms. I also published an op-ed on Thursday 23 February 2023 in the Dallas Morning News about the same topic, how generative AI represents an opportunity to re-think and re-define how we teach writing. Finally, I gave an interview to The Texas Standard, a Texas public radio station, on why I’m not as anxious about generative AI’s impact on higher education as some others in my field. The segment aired around 10.30am Wednesday morning on 22 February 2023.

Stay tuned for more updates.

The Hypothetical: Between Classroom Lecture and Discussion

A perennial debate among teachers seasoned and new concerns how much lecture versus discussion to feature in class. The debate is especially relevant to teachers of liberal arts subjects, since the content of such courses is not always conducive to rote learning techniques. Liberal arts subjects often require the completion of previously assigned reading, and even if enough students read to engage in fruitful discussion there remains the risk of devolving into debate rather than dialogue. To complicate matters further, lectures and class discussions are frequently, and falsely, pitted against one another, viewed as binary opposites and sometimes filtered through a political lens which codes the former as conservative and the latter as progressive. Consider the image of a professor lecturing to a class of students dutifully taking notes versus a classroom of circled desks with students and teacher alike engaging in a dynamic conversation.

I believe there are more or less appropriate times and subjects for either lectures or discussions, and I don’t buy that either has a particularly salient political character; after all, even the teacher discussing texts in a circle still assigns grades at the end of the term. But there are challenges to both lectures and class discussions that I think frustrate new teachers in particular. Lectures aren’t especially engaging unless done well, which comes with time and practice. On the other hand, discussions can intimidate students to the point of non-participation, especially if the topic of discussion is particularly controversial. Also, the skill of facilitating and nourishing discussions is underrated and quite challenging, another pedagogical virtue that comes with time and practice. Additionally, I find neither mode particularly effective for beginning a class. Opening with a lecture (especially in the early morning) risks students nodding off to sleep, as their engagement with lectured material is entirely determined by their individual commitment. Conversely, jumping into a discussion right away often fails to take off, as any teacher can attest, since students can be reluctant to break the ice.

Enter the hypothetical. The hypothetical is exactly as it sounds: a contrived example or problem related to the day’s concepts is posed to students who must reason through its various frictions. Incorporating hypotheticals into liberal arts classes is particularly effective, as it occupies a middle ground between lecture and discussion, combines the best elements of both, and requires no prior reading. Like a lecture, a hypothetical allows the instructor discretion in guiding students more or less forcefully towards the concepts intended for learning. And like a discussion, it engages students’ creative and critical thinking muscles, prompting them to respond and participate. Hypotheticals are a great icebreaker to boot; completing the reading is not required to participate in the deliberation of the hypothetical, as its contrived parameters present enough content for participation. It also doesn’t risk intimidating students from offering controversial opinions since its hypothetical nature provides a comfortable space for intellectual exploration.

Here is one hypothetical I offer students in one of my first-year composition courses:

Here’s what typically happens when I pose it. The majority of students almost immediately answer “no,” which allows me to oppose the class consensus to encourage deeper thinking. For instance, I point out that only a 1% chance of a wrongful conviction is quite good, especially considering current estimates put the US wrongful conviction rate somewhere between 2-10%. I then ask which kind of evidence would persuade them to convict, to which they usually respond “eyewitness testimony,” “DNA evidence,” and “video footage.” All of which, I suggest, is likely less than 99% accurate. So what gives, I press.

I then switch to another hypothetical, this time involving blood pressure medication, to which the entire class always answers “yes.” (As a side note, a great technique is to juxtapose two structurally-similar hypotheticals that nevertheless induce students to opposite conclusions; the cognitive dissonance is generative.)

“So what’s the difference?” I ask. Students usually reason many of the differences pretty well, in my experience: The stakes are different for sending an innocent person to prison compared to leaving blood pressure untreated; the former involves deciding someone else’s fate and the latter your own; and there’s a matter of “trust” when it comes to doctors that feels distinct from the consequentialism of the judicial system.

The larger point I make is this: The American judicial system (in theory, at least) does not consider the probability of committing a crime as admissible evidence, knowledge of guilt. (Thank God Minority Report is just a movie.) Even if there’s only a 1% chance, you are presumed innocent until proven guilty. In American medical science, however, the probability of a drug’s effectiveness, inferred from carefully controlled experimental trials, is considered knowledge. In fact, inferring the effectiveness of medical treatments from samples is the only way to know anything in medicine at all; it is impossible to “witness” or testify to, at least rigorously and systematically, the evidence of medical efficacy. These hypotheticals therefore elegantly demonstrate the difference between two types of information: observational knowledge, obtained and verified through experience, and probabilistic belief, inferred through studies of manipulated samples.

In the larger context of the course, we are discussing the nature of knowledge, how it is we can say we know something. The goal of this class period is to demonstrate how knowledge is context-dependent; what counts as knowledge in a medical trial is not what counts as knowledge in a court of law. Does that mean knowledge is whatever we say it is? No, it means that certain domains (medicine, law, science) have developed their own rules that govern how knowledge in that domain is understood and counted. Awareness that there are differences in how we derive knowledge is a concept I discuss in my first-year composition courses because I believe the idea is essential to understand before reading and writing academic research at the college level (your mileage may vary). This hypothetical helps students to intuit the primary takeaway of the lesson, the constructed nature of knowledge, without me lecturing at them about it.