There are many approaches to “AI readiness” in higher ed. Mine, a 2000-level English course, begins with the second L in large language models, language. Accordingly, and with an eye to state legislators scrutinizing syllabi for racy texts, we begin with Aristotle.
Most AI 101 college courses begin with the technology. Stanford’s CS324 (“a new course on understanding and developing large language models”) opens by defining a language model as a probability distribution over sequences of tokens, which are words or pieces of words. “The mouse ate the cheese” is more probable than “the cheese ate the mouse,” and both are far more probable than “mouse the the cheese ate.” Students learn about entropy and Claude Shannon’s 1951 theory.1
Ed-tech vendors instruct teachers to tell students that a chatbot is a sophisticated autocomplete that does not think, feel, or know. Writing Program courses either assign a book with a unicorn, rainbow, and robot on the cover or long lists of readings that hurt my head. All of these courses assume that students understand how language works and therefore all they need is to have their hand held while they learn a technology. (Arguments abound whether it is “like a calculator” etc.) Anyone watching the news sees that the technology is changing every few months, with lower hallucination rates and higher rogue agent rates. I wanted to teach durable skills about what an LLM is.
What does a young person born around 2006 know about language as a corpus? A lot less than people born before the internet. Many college students learned to read by a now discredited method called three-cueing, told to figure out unfamiliar words by looking at a picture and the first letter, guessing what made sense. Some have read many whole books but the news says they are reading less. According to a 2024–25 survey, high school English teachers have been assigning an average of 2.81 books a year, down from four in 2008–09, and on the 2024 national assessment, twelfth-grade reading scores were the lowest in the test’s history.2
More importantly for a college-level AI course, most of today’s students have never held a dictionary in their hands, a nice fat one with the little thumb cutouts to get to the first letter of the word you’re looking up. Merriam-Webster’s free website went up in 1996, a decade before they were born. It’s unclear who is still buying physical dictionaries but sales are way, way down. The point is that for students, language is largely digital.
What does it mean to hold a substantial chunk of the language that you use in your hands? It means you understand that it is a corpus and that it is yours to use. Most 20 year olds have not been taught to understand and use language as a tool. Young people are taught to use language as a vehicle for expressing needs and feelings, via heart maps, first-grade Small Moments, and personal stories. Teachers tell students to write honestly and from the heart.
So I created my class for college students who have mostly encountered language in a fugitive sense, in passing, on social media, as a vehicle for expressing internal states. They have always had autocomplete. Their language has been shaped by what Adam Aleksic calls algospeak. They are fluent but they have no real account of how language works. Saying that chatbots “just predict the next word” means less than it means to someone who has held a dictionary. If we lived in an era when college students still bought books at a bookstore, I would have assigned a big one. I would say look, you own language, it doesn’t have to own you.
Aristotle’s Rhetoric and the Poetics are manuals for using that universe of words. Aristotle explains how to choose a word for a particular hearer and how to arrange words and events so that they have an effect. Once a student grasps how dictionaries and Aristotle approach language, it’s much easier to explain how LLMs are trained on massive amounts of human writing referred to as the pretraining corpus. LLMs divide words into tokens, which are words or pieces of words; given a prompt or set of tokens, a model assigns a probability to every possible next token, and the full set of those probabilities is its distribution. It draws a token from that distribution and adds it to the text, and so on, predicting the next token from the longer text.
If you skip the first step, you are skipping a central point: LLMs should be compared to other older ways of organizing and understanding language. Seasoned faculty who have grown up, like me, with well thumbed dictionaries still on shelves, may not understand how important this grounding is.
Aristotle was among the first to divide language into parts, beginning with sounds and syllables and moving up to words and sentences. He sorts words into the ordinary ones a community uses and strange ones that strangers use. He defines the probable as what typically happens. He defines plot and argument by what a listener expects to hear. Most importantly, he assumes his reader knows nothing. He cares about language at a basic level and communication as something performed.
In the Poetics, Aristotle warns that words wear out. Style is a matter of using a good mixture of current words, which are already used up, and strange ones. He warns that compound words should be used sparingly. He warns against using rare words improperly. Centuries before autocorrect he understands that the wrong word can be hilarious. He says that the occasional unfamiliar word dignifies language because people admire what is distant. Dropping a little française into language makes you dignified until everyone starts doing it. In short, the supply of words is not fixed. Stale metaphor and cliché were problems even in Aristotle’s day.
LLMs, which are trained on everything ever written, have new, old, current, strange, stale, cliché, all in the distribution, weighted by how often people used it, with enough fine tuning that a Victorian metaphor wouldn’t be used in generating a business letter. Aristotle provokes these discussions. Students who learned as children that voice is the heart and soul and emotion can now better see that voice is a set of choices made for a hearer, and that unlike an LLM, they know who is being addressed and what that person already expects.
The Rhetoric is about language as a tool for persuasion, which you can think of as social next-word prediction. Metrical prose is unpersuasive because it seems fabricated, Aristotle says. Predictability is a defect. Hearers resent detected artifice. Much of the Rhetoric also seems like something an LLM could have written: lists and templates with arguments for both sides of a case. It’s a good model for teaching alternatives to next token prediction.
Aristotle makes the point in both books that there is “pooled opinion.” Words mean what the people understand words to mean; people aim for truth and what resembles truth. Pretraining on human text assumes that people mostly attain the truth when they speak, by which is meant people mean what they say. The English word “stochastic” comes from the Greek word for skill at aiming. Emily Bender’s famous label “stochastic parrot,” for LLMs, gets at this in a negative sense. Aristotle pools opinion and trusts the pool. A pretraining corpus likewise pools opinion.
Is peripeteia, Aristotle’s word for reversal the same as improbability? If pity and fear arise when events happen “contrary to expectation,” is the point of an LLM never to cause a user fear and pity? Homer, Aristotle says, taught poets to tell falsehoods by paralogism, a false inference. When one thing follows another, people who see the later thing think the earlier thing exists too. You can spend endless time on metaphor.3 There are fun lists of terms for class projects, to see if an LLM will “perform” them: enthymeme, blame, taking an oath, anger, pity, ethos, pathos, shame. It is pretty hard to get an LLM to shame a user.
Bottom line, I think teaching LLMs to anyone under 25 should start with language rather than starting with the technology. Students need an account of where language “is,” to understand there is a reservoir (to use a new metaphor), that nobody owns it but everyone draws from it. You can own a specific arrangement. Aristotle’s account of style and plot is basically a theory of selection. It’s remarkably easy to move from Aristotle to an LLM and back, to compare and contrast.
Once students see language in an Aristotelian way, nobody talks about what a chatbot thinks or feels. A few weeks in they appreciate the LLM’s definition of a dictionary as an alphabetized projection of a vector space where the words are positions in semantic geometry, where each word’s definition is its cosine similarities to all the others, and the axis for retrieval is alphabetical order orthogonal to every axis of meaning.
Perhaps more people will see this is the right way to do it. Anthropic’s student course is organized around next-token prediction, knowledge, working memory, steerability. A student can pass this class and still have no account of language.
It’s interesting to me that Anthropic goes to Aristotle for virtue instead of language. Claude’s Constitution calls for good values, knowledge of the world, and wisdom to turn values and knowledge into action. Readers swiftly heard the echoes of virtue ethics, with phronesis and parrhesia on the list of desired traits. Critics responded that Aristotle’s virtues come from habituation and Claude’s training was just a nice to-do list for which it was rewarded. The same document also tells the model to keep the length of a response proportionate to the request, to adapt its tone to the person and the setting, to skip stock phrases and filler, and not to flatter. Book 3 of the Rhetoric covers all of it. It’s the better path.


Partial reading list (which also includes the Entropy/Compression video and Weaver’s Recent Contributions)
Shannon, Claude E. “A Mathematical Theory of Communication.” Bell System Technical Journal 27, no. 3 (July 1948)
Quinn, Arthur, and Barney R. Quinn. Figures of Speech: 60 Ways to Turn a Phrase. London: Routledge (Hermagoras Press), 1995.
Buchanan-Brown, John, et al. Le Mot Juste: A Dictionary of Classical and Foreign Words and Phrases. Vintage Book, 1981. ISBN 0-394-74690-2.
Aristotle. Poetics.
Aristotle. On Rhetoric: A Theory of Civic Discourse. W. Rhys Roberts translation, public domain — http://classics.mit.edu/Aristotle/rhetoric.html
Disney, Walt. Inter-office memo to Don Graham, 23 December 1935. 8 pp. https://filmic-light.blogspot.com/2010/07/don-graham-how-to-train-animator.html
Lodigiani, Cento. “The Illusion of Life.” 2014. https://archive.org/details/TheIllusionOfLife
Coats, Emma. “The 22 Rules of Storytelling.” 2011.
Shaw, George Bernard. Pygmalion. 1912
Disney 12 principles of Animation https://www.nyfa.edu/student-resources/12-principles-of-animation/
Vaswani, et al. “Attention Is All You Need.” https://arxiv.org/abs/1706.03762
Scarry, Elaine. The Body in Pain: The Making and Unmaking of the World. Oxford University Press, 1985 ISBN 978-0195049961 (paperback, 1987). Assigned: chapters 1 and 4.
Plato. Seventh Letter. http://classics.mit.edu/Plato/seventh_letter.html
Anthropic. Claude’s Constitution. January 2026. https://www.anthropic.com/constitution
Anthropic. “Claude’s Character.” 8 June 2024. https://www.anthropic.com/research/claude-character
I never ever criticize my students; this is data about the generation broadly.
Poetics 21, the Butcher translation: “Metaphor is the application of an alien name by transference either from genus to species, or from species to genus, or from species to species, or by analogy, that is, proportion. Thus from genus to species, as: “There lies my ship”; for lying at anchor is a species of lying. From species to genus, as: “Verily ten thousand noble deeds hath Odysseus wrought”; for ten thousand is a species of large number, and is here used for a large number generally. From species to species, as: “With blade of bronze drew away the life,” and “Cleft the water with the vessel of unyielding bronze.” … Sometimes too we qualify the metaphor by adding the term to which the proper word is relative. Thus the cup is to Dionysus as the shield to Ares. The cup may, therefore, be called “the shield of Dionysus,” and the shield “the cup of Ares.”


This was riveting, and is still resonating. Also, what a delight to find “Figures of Speech,” a favorite book, in the footnotes.
You can hand someone an instrument and a playing fundamentals book, but if they don't know anything about music, they will never be a musician.