ChatGPT Could Prompt a Rennaisance in Writing Pedagogy
How to Teach Writing in a World with LLMs? Consider Why Writing is Taught in the First Place.
The Shape of the Problem
In late 2022, ChatGPT, and the technology of large language models, burst onto the scene overnight. In terms of the frantic process of cultural digestion that it set off, the only comparable phenomenon I have witnessed in my adult life was the advent of COVID-19. Nearly every arbiter of social consensus did not know what to make of it, but knew that something had to be made of it. Eight months later, our collective digestion of the technology is still well underway, though it has at least found a definite place in the cultural intestine.
The way the large language models function has been popularly demystified thanks to masterfully clarifying articles like Ted Chiang’s “Chat GPT is a Blurry JPEG of the Web” and we understand that it cannot meaningfully be called ‘intelligence.’ Countless articles have been written on the extent to which LLMs threaten a given job function (customer service reps or paralegals, for example), and the conclusion is typically the same: it will be used as a tool that increases efficiency, reduces total labor requirement, and forces upskilling… but will not erase core job functions in one blow. In this respect, self-driving vehicles are a greater threat to employment rates than LLMs are.
Where the goal is productivity, there will be a way to keep humans in the metaphorical driver’s seat (though perhaps not the literal one!) for the foreseeable future. There is, however, one domain of activity that LLMs are already fundamentally encroaching on education. In education, it is not the productive job function of the teachers that is encroached upon- but rather the processes by which students learn. Although large language models are incapable of creating any cutting-edge analytic or original creative text, they are essentially purpose-built to create simple but well-organized essays that cover well-trodden ground with established arguments, i.e. the kind essay which, when turned in to fulfill an educational writing assignment, will reliably earn a student a B+ from the 5th grade up to college graduation. This will lead to massive productivity gains in the student population, but unfortunately, students are not valued for their productive capacity. The contribution of high-school essays to the gross world product is zero. We hope, rather, that students practice and master certain processes- among which writing is included -and this is precisely what LLMs are poised to make superfluous.
For as long as humans have invented things, we have ceded the functions of our bodies and minds to our inventions. If we cede to LLMs the function of writing mindless marketing copy, technical descriptions, and customer inquiry response emails — it will ultimately be no greater tragedy than ceding long division to the calculator. But things look different in the education system. The practice of mentally dividing figures provides no real value to one’s inner life, and the concept of division can be learned rather quickly and then applied to more advanced mathematical reasoning. Basic writing skills work quite differently. Proficiency in the mental processes that underlay simple writing assignments like book summaries and five-paragraph essays provides immense value to one’s inner life and is acquired slowly and incrementally. Writing is a medium that augments and structures thought while leaving its basic dynamics unmolested. The ability to think in text is a critical learning outcome, even if a student’s productive capacity never exceeds that of an LLM.
My intention for this article is to illustrate the way that writing functions as a fundamental instrument of thought, and to accordingly indicate its immense educational value independent of its productive capacity. The question of how to protect language pedagogy from LLM-cheating and reliably “A.I.-proof” writing assignments follows naturally. A model of instruction that leans into the highly personal, introspective value of writing solves this problem better than any measures that seek to prevent LLM usage through detection or surveillance. LLMs can paraphrase cliche arguments and compile valid pieces of supporting evidence, but they cannot craft personal narratives that integrate understandings of class material and research into accounts of embodied experience. If this latter type of writing were the basis of writing education, it would not only solve the problem of cheating with LLMs, it would bring pedagogy into line with the real value of writing in the context of education.
The Media of Intuition
Intuition is a weird, slippery human faculty. It occurs in some region of the mind well below the level of conscious awareness where internal and external stimuli are silently processed, paths of association are traced, and bits of ideational content get spit out on the threshold of awareness. Intuition becomes thinking when we manipulate its inputs, discriminate among its products based on a set of parameters, and then translate an isolated product into some other medium (imagination, subvocalized speech, vocalized speech, calculation, writing, music, action, or image- to name a few).
Here’s a quick example of what that process looks like. A person asks me what I like about my job. How do I come up with an answer? I ‘manipulate inputs’, soliciting my intuition to bring some aspects of my job into the field of my awareness. Things start occurring to me, and I discriminate amongst them based on the simple parameter: is it an aspect of my job, and do I like it? “The commute” occurs to me, which I do not like and accordingly discard. “Eating lunch” occurs to me, which I do like but isn’t exactly an aspect of my job, so I discard it. “Working from home two days a week” occurs to me, which I do like and is truly an aspect of my job, so I isolate this and then find a way to translate intuition into speech. The same operation occurs in all instances of thinking.
I feel most in touch with my faculty of intuition when I have closed my eyes and am waiting to fall asleep. With almost zero conscious deliberation I begin reviewing my day, get lost in the occasional imagistic fantasy, remember things I should have done, worry about inconsequential conversations, and so on. Occasionally, I will become self-aware and consider the things I had just been thinking about. In these moments I am almost always surprised by what I find. I will discover, for example, that I have spent the last 10 minutes crafting a fictive narrative of a coworker’s life or reconstructing the general atmosphere of a bar I was at the week before. In this space, I shift between different media of thought, like subvocalized speech, imagination, calculation, etc, without much deliberation.
There are certain rumblings of the spirit that are difficult to hear except in prolonged swims in the river of intuition: maybe I am becoming dissatisfied with my career, maybe I have been an inattentive friend lately, maybe I should apologize for that thing I did. However, it is fantastically difficult to come to drive these rumblings to a meaningful conclusion in the borderless void of the mind. I sometimes try to logically order several silent thoughts in hopes of arriving at a particular insight. Almost always, the result is that I lose sight of all the pieces, make significant detours whose relevance I immediately forget, begin reliving a faded memory, and then find myself considering what I want to have for dinner.
This is the limit of thought accomplished without an external medium. It provides the building blocks of all complex thinking but does not offer them as durable objects. It’s hard to hold them down for long enough to evaluate them thoroughly, and almost impossible to hold them up next to each other.
The two basic media which remedy this shortcoming of intuition are speech and writing. Thoughts expressed in speech remain basically ephemeral. Though it is easier to forget something merely thought than it is to forget something thought and spoken, both fade into the recesses of the mind in time. However, the spoken thought possesses a syntactic imperative that the purely ideational thought lacks. It demands to be connected to another spoken thought, to be built on, replied to, confirmed, or refuted.
Two intuitive thoughts can immediately follow each other without syntactic bindings. If my thoughts were magically transcribed as I was walking my dog this morning, the following excerpt would result: “ (1.) There is a narrative in Curb Your Enthusiasm where Larry David convinces his therapist to steal his wife’s therapist’s purse so that Larry David can intervene and look like a hero. He does this because his wife’s therapist is encouraging a divorce, and he believes he can stop this if his wife’s therapist thinks positively of him. (2.) Butter and noodles taste surprisingly well together. (3.) There are so many moments in 20th-century film and literature where psychiatrists have a dominating Kafkaesque power over their patients. (4.) Tonight I’d like to make pasta with oil and lemon, and then finish it with butter.”
This is a completely natural flow of intuitive thought, but if I were to speak those exact words in that exact order to a coworker, they would think I was completely unhinged, as the thoughts fail to meet the syntactic demands of speech. Thoughts 2 and 4 are closely associated, as are 1 and 3. An interesting connection could be drawn between either pair, but on their own, they lack the integration necessary to form a coherent utterance.
If I were to describe the scene from Curb Your Enthusiasm in conversation, I might begin to reflect on how the power dynamics between Larry David and the institution of psychiatry are an interesting reversal of those characteristic of the 20th Century, and productively connect it to One Flew Over the Cuckoo’s Nest. I might then also determine what I’d like to eat for dinner in light of my sudden appreciation for the taste of buttered noodles. While speaking, I would be aware that the utterance has an audience other than myself, and that this places demands of coherence, brevity, and accessibility on it. This forces a real structure on the thought that allows it to integrate levels of complexity that I could never maintain in the flow of intuition.
Most speech events take place in conversations, and as such thinking strongly bifurcates into passive and active modes: listening and evaluating, vs. concluding and speaking. Intuitive thought, in contrast, never really settles into either of these modes and just shapelessly flows. When you enter the passive mode in a conversation, you are integrating your thoughts with those of your interlocutor, which are a product of a set of experiences, assumptions, and dispositions you do not yourself possess. This allows for even more complexity to be integrated into verbal thought.
Occasionally, when I’m having a great conversation, I become rather anxious about all the valuable tangents that are being surrendered to the flow of time. While listening to a friend speak, I will identify three ideas I want to respond to. By the time he’s done talking, I will have forgotten the first, and as I’m responding to the third, I forget the second. Thinking is often analogized to cartography. Along those lines, thinking in a conversation can feel like making a map of a city in real-time, after passing through 10 four-way intersections only once. Blank space hovers over all the turns you didn’t take, and there is a palpable sense of incompleteness. Speech allows thought to chart courses, but not entire territories.
Writing as a Medium of Thought
Writing liberates thought from time and thereby enables the charting of whole territories. Once translated into text, a thought will stand in the same place in the same form, for as long as you need it to. In writing, one can patiently unfold the structure of a thought’s implications without worrying that it might slip away. This opens up several possibilities for thought that are not accessible in other media. Writing is a radically effective facilitator of thoughts with a large number of immutable components (a shopping list), thoughts that require rigorous syntax (a mathematical proof), thoughts that involve highly determined subtleties of meaning in multiple semiotic fields (a poem), and thoughts which require precise collaboration between multiple parties (a legal code).
Writing allows for a very productive application of both intuitive and verbal modes of thought. Writing places the same syntactic guardrails on thought that speech does: a written sentence, like an utterance, asks to be connected to another. The immutability of a written sentence allows for an even more precise logic of combination. The connection between two written thoughts can be more easily identified, and therefore more easily evaluated and refuted. When speaking, you determine a connected utterance following one intuitive impulse; when writing, you can roll the dice of intuition an infinite number of times when developing a connecting thought.
Writing’s utility as a mode of thought touches almost all areas of life. Its utility is so great that it is a functional necessity for complex fiction, legal codes, environmental impact reports, reporting, drafting policy, conducting business communication, etc. All of these things are written products, obviously, but that’s not what I’m pointing out. There is an audio-book version of David Foster Wallace’s Infinite Jest, but the ideas and the structure of the book were both developed and finalized in writing. If you think of Infinite Jest, or any other book, less as a cultural product than as a remarkably coherent 500,000-word thought, the power of writing as a medium of thought comes clearly into focus. There is no other medium in which it would be possible to generate and arrange 500,000 words into a thought that is not only coherent, but also footnoted, internally hyperlinked, and extensively self-referential… to say nothing of its aesthetic value.
Writing’s instrumentality in creating complex works of fiction is not the best way to understand its utility as a medium of thought, as it still frames writing principally as a mode of production. A better way to understand writing as a medium of thought is to consider the kind of thinking that occurs when one writes a love letter. When I sit down to begin writing a love letter I do so with two simple presuppositions: that I love the addressee, and that there are some specific reasons why, and/or some unique qualities to that love. I identify one of those reasons why, or one quality of that love, and make it the subject of the letter. For example, I choose to write about the feeling that our love was an inevitable, somehow fated thing. I think of all the pieces of evidence for this conclusion: the circumstances under which we met, the intensity of feeling from the outset, the thoughtless fluidity of our interactions, the synchronous details of our biographies, all the unlikely events that had to happen to place us in front of each other. As these things occur to me they are immediately translated from vague intuitions into concrete truths. Each detail I write down reminds me of another, and soon the list proliferates. I inscribe each recorded detail with a reflection on its significance to me. By the time I finish, a slew of once-shapeless facts and feelings have been elevated into components of one profoundly meaningful truth.
When a love letter is signed, what has occurred? An act of production, to be sure. Where there was once a pen and paper, there is now a gift and a testament. But even if this gift and testament were to vanish from the earth the same moment it was signed, its value as a medium of thought would be preserved. In writing the letter, I have imbued my feelings and experiences with concrete meaning, clarified my orientation to my partner and the world, organized my memories, and strengthened the basis of my love. Writing the letter has allowed me to engage in a kind of thinking that is tremendously valuable in its own right.
Writing is a highly effective medium of thought whenever subtle or easily overlooked implications of a primary thought are sought after. Even when such implications aren’t sought after, writing has a way of bringing them to mind. Writing emails at work often leads me to realize something valuable about the subject of the email that hadn’t occurred to me before I started writing. I work for a third-party logistics company, and customers often will request status updates on their shipments over email. I will set out writing a purely descriptive answer, and then get stuck on the last sentence, which might be, for example: “the shipment has been delayed by two days and will now arrive on Friday via Turkish Airlines.” I sit and stare at the phrase while I think about what to write next. Various half-articulated thoughts pop up in my intuitive flow: “Turkish Airlines SFO warehouse is at 632 W Field Road,” “They have a cargo handling contract with Swiss World Cargo,” “My supervisor is Turkish,” “I have never eaten Baklava because I am allergic to pistachios.” The customer wouldn’t be interested in any of this, so the thoughts flicker and fade rapidly.
Eventually, something more relevant comes to mind: “Turkish Airlines allows you to store a shipment in their warehouse for 24 hours before they charge storage at a rate of 85 cents per kilo per day.” I follow the implications of this thought in a similar manner and eventually discover that I need to propose a new trucking plan including weekend pickup if we are to avoid storage charges. I propose a few alternate trucking methods, all of which are more expensive than the original one, but will result in net savings in light of the storage fees. Writing a brief descriptive report on a delay prompted me to consider all the implications of this delay, and propose solutions to the problematic ones. When you commit a thought to writing, you are able to scan through a litany of associated thoughts without risking losing track of the thought that prompted the scanning. This allows you to fully unfold the implications of a thought in a way that is very difficult to do intuitively.
Writing as Process and Product
Writing will always possess a dual nature as a scaffold of thought, valuable as an instrument, and as a final product, valuable for some purpose. There is writing that leans further in one direction than the other. Lecture notes written for oneself serve more as an instrument of thought, but that scaffolding becomes useless if it isn’t arranged with function in mind. The 1847th clickbaity article written about the symptoms of diabetes serves primarily as a piece of content, valuable in so far as it attracts clicks and eyes… but it also bears traces of one author’s hasty scramble to organize their bits of understanding into a monetizable object. A novel is a testament of an author’s thought process and is also an artifact that can be distributed and consumed. The Charter of the United Nations is a result of sustained collaboration and compromise between a large number of parties and is also a functional document that codifies the structure and purpose of an organization.
To develop a skill is to develop one’s ability to think under certain circumstances, according to a particular logic, in a certain medium. This is true no matter the skill: tennis, carpentry, chess, filmmaking, wrestling, pipefitting, architecture, surgery, and music are all thought expressed under certain circumstances, according to a particular logic, in a certain medium. Aptitude in those activities is the fluency, accuracy, and breadth of thinking within the relevant constraints. To be skilled at any discipline is, in a meaningful way, to be skilled at thinking. Writing, as a discipline, places only the most abstract constraint on thought: that it represents itself concretely on the margins of space and time. Skill at writing is a very close correlate to skill at abstract thought and the operations of observation, intuition, and organization.
It should come as no surprise, then, that writing is such a crucial technology for thinking through situations, life stories, feelings, problems, and shopping lists. The importance of writing, for the majority of people, is not so much consisting in its value as a mode of production, but rather in its value as a medium of thought. Very few people need writing to produce novels or plays or legal decisions. However, for almost everybody, it is or can be a critical tool for structuring identity and memory, crafting arguments, and ordering priorities and understandings. If there is a reason to include writing in educational curricula at the K-12 up to the bachelor’s level, it is precisely to develop students’ proficiency in using writing as a medium of thought.
High school and college essays are not expected to make groundbreaking contributions to their discipline or display significant aesthetic value. They are, with rare exceptions, written, graded, and then never read again by a student, teacher, or any other party. Their value as a pedagogical tool, however, is immense. They lead the writer to adopt a searching and critical approach to class material and then to organize information, observations, and conclusions about that material into a cohesive, presentable thought. They are effectively the only pedagogical tool that actually prompts and rewards deep thinking about learned material. The value of writing, in these educational contexts, consists almost entirely in its function as a medium of thought- the product, on the other hand, is effectively worthless. Only in pedagogy is writing’s process-function superior to its product-function to this extent.
Large Language Models and The Future of Pedagogy
The advent of large language models is going to seriously challenge the role of essays in the education system. LLMs can produce high school and college level essays at an acceptable standard of quality without showing obvious traces of their origin. They can be trained on any given student’s actual writing style and asked to replicate it in the text they generate. They can scrape all the information on a given subject off the internet, paraphrase common arguments about that subject, include and organize the evidence that commonly appears in connection to that argument, and produce a paper that is coherent and defensible at virtually any length. This results in a product essentially matching that which students are asked to produce in grading rubrics. LLMs allow the product of writing to stand as a false index of the process of writing.
Unless significant changes are made to the way writing projects are assigned and evaluated, the use of LLMs to complete academic assignments will be widespread. Many students consider writing assignments to be burdensome and unrewarding, just as they consider the education system to be either an imposition or a means to a credential. Educators may have been able to ignore this fact in the past, but they are now being forced to confront it. In a production of The New York Times production The Daily titled “Suspicion, Cheating and Bans: A.I. Hits America’s Schools” interviews with students and teachers show that this understanding is setting in rapidly. The shortcut is simply too tempting and too undetectable to not become endemic, as things stand.
Removing the essay from pedagogy would be disastrous. The development of proficiency in writing, as a medium of thought and expression, is one of the single most important basic educational outcomes. There are whole dimensions of creative and analytic thought which will be outside of students’ capability if they do not have the opportunity and imperative to reach for them in writing. I would encourage anyone who has access to a writing assignment completed by an 11-year-old to consider it carefully, as this is, in a rather conservative estimation, the age at which it would become impossible to restrict a child’s access to tools like ChatGPT. Imagine then what the effect on your inner life would be if your ability to think in writing were not to improve from that point. The effect of LLMs in education will not be like that of the calculator, which effectively augments one’s ability to reason mathematically by reducing the time and effort required by basic operations. Neither will it be like that of spelling and grammar correction tools, which do the same for orthographic and semantic operations. LLMs can substitute for the very cognitive functions education systems seek to develop. A.I-proofing the pedagogical essay should be a top priority for educators at every grade level in every country.
The most draconian approach to A.I. proofing the essay would be to proctor writing assignments. In my mind, this is a non-starter. The value proposition of writing is that it allows you to think slowly, methodically, independently, and unimpeded by the temporal constraints characteristic of other media of thinking. Proctored essays might be a good way to indirectly gauge a student’s writing ability, but they are no way to develop it. Tools designed to identify text written by a LLM are already starting to spring up, and will surely continue to be developed and improved. My guess is that they will always remain one step behind LLMs, that the likelihood of punishing an earnest student for a false positive will always be too great, and that they will ultimately fail to serve the education system productively.
The change required to A.I.-proof the pedagogical essay is not a change to the method of administering writing assignments, but rather to the method of teaching and evaluating writing. A pedagogy that teaches and evaluates student writing based on a privileging of writing’s value as a medium of thought, rather than as a mode of production, would not only effectively A.I.-proof writing assignments- it would dramatically improve the practical value of writing instruction for students.
Writing is currently taught primarily as if its value were as a mode of production: specifically, as a mode of producing graduate research. Beginning at the elementary level, students are instructed to progressively incorporate the rhetorical, analytic, and structural standards of academic writing into their own. Accordingly, the central milestones in writing instruction include paragraph structure, introduction-body-conclusion structure, expanding page counts, adoption of impersonal language, peer-review compliance, etc. Some of these standards are useful for using writing as a medium of thought, some are considerably less so. This writing curriculum is an excellent way to prepare students to eventually produce graduate research papers but is an inefficient and indirect way of teaching them how to use writing as a tool to augment their thinking and make sense of their world. The vast majority of K-12 students will never undertake graduate research- so why is this the standard their education is geared to?
A pedagogy that dials in on writing’s universal value as a medium of thought would serve students much better, and would also effectively A.I.-proof writing assignments. An LLM can easily produce a four-page essay that paraphrases a cliche argument about the significance of some historical moment. However, an LLM could not produce a first-person narrative account of the experience of learning about that same material, in which the author recounts the way he developed his understanding of the material and the questions it raised for him, how he sought to answer those questions through research, the personal experiences it made him recall, the comments made by other students during class discussion, etc. The second kind of writing can be coached to contain arguments, evidence, and organization just as rigorous as the first- all while being more useful, natural, and enjoyable for the student.
Machines can scrape and synthesize existing text, but they will never be able to perceive the world and share in the subjective, embodied experience of it. Machines cannot produce text that charts the connections between class material and the vast intersubjectivity in which that material is encountered. LLMs can only regurgitate and synthesize courses already charted. If LLMs are to have a positive effect on our writing and writing pedagogy, it will be because they nudge us into uncharted waters of thought. The Sun Also Rises by Ernest Hemmingway is a classic of American literature- hundreds of thousands of essays have been written about it. Robert Cohn and Jake Barnes, two of its central characters, have had their personalities compared tens of thousands of times in such essays. These are thoroughly charted waters, and if students are asked to write an essay comparing the two figures, they will be able to ask ChatGPT to do so. If they are diligent students and choose to write the essay themselves, they will come to an unoriginal conclusion that has been made much more convincingly by someone else, and it will have limited bearing on their life. However, if they are asked to compare Robert Cohn to one of their family members (for example), they will be charting completely new waters, inaccessible to LLMs. The resulting essay would be of great personal relevance and will have prompted the students to undertake a more thoughtful reading and character analysis. It will also train them in the kind of writing and thinking that, regardless of one’s course through life, can be used to enrich one’s lived experience.
With machines slated to dominate the production of generic and impersonal text for eternity, it is time that we as breathing subjects shift towards the idiosyncratic and personal. The imperative to do so is nowhere as clear as in education, where under the current system text generated by LLMs is perfectly substitutable for that which students are asked to create. Even in a world without LLMs, there would be a clear benefit to prioritizing the teaching of introspective, embodied writing. We don’t need to compete with machines, we just need to embrace the truly human.
