The orbital argumentation of AI writing

Introduction

I’m endlessly fascinated by large language models (LLMs). They’re the engines of GenAI tools. They’re exhilarating. On the back end, they execute interesting mathematics, using gradient descent. On the front end, they’re a bit like a crystal ball: users aren’t entirely sure what they will get, whether it’s something useful or some Lovecraftian experience. Or an error. The crystal ball metaphor seems to work well because we interpret the sentences they produce, imbuing them with meaning. To paraphrase Nietzsche, they’re a bit like an abyss: if you use them long enough, the abyss stares back in strange ways.

That doesn’t mean that they should be used in classrooms where the point is learning. It doesn’t mean we should be investing billions into what might be the Beta Max of AI technologies. (For those of you not old enough to remember Beta Max, it couldn’t record as long and was more expensive than VHS tapes. It lost the “format wars” of the 1980s.) In a lot of ways, I loathe the hype around these technologies more than the technologies themselves. I detest, too, the theft of copyrighted material. I’m pretty sure my first book, Update Culture, was scraped by the AI bots. But these are all very practical problems. There is a conceptual problem lurking in the background.

The problem is in how a GenAI machine makes an argument. Perhaps, my faithful readers, you think that I’m going to argue AI chatbots and other machines don’t argue. Nope. That’s not my argument here, although I’ve argued that in previous essays. My argument: AI writing has an orbital form of argumentation. It’s elliptical argumentation, one that has the feel of someone who didn’t do the reading but did manage to skim it, while maybe memorizing several important passages. When LLMs write in response to a query/prompt, they often orbit around the actual query. When I use the word orbital here, I mean something like circling around an object without touching it but is caught in the gravity. AI writing, almost always the summaries, mentions “insights” without stating what the insights are; Gemini’s podcast creator is particularly guilty of doing this. When I read an AI written summary, it’s an uncanny valley. A picture of a picture. Grainier than it should be. Not wrong but blurry. It feels correct if you don’t read it too carefully. If you read it firmly, you’ll notice all the valleys. Sentences don’t actually connect. If you don’t let the machine bully you, if you do more than skim the output of the LLM, you notice all the cracks. I’m convinced this is why the AI bots write so effusively. They produce so much text in order to provide the user something that could be interpreted as useful. It’s the same way a modern fortune teller uses their crystal ball. Say enough orbital things and you render something useful for the fortune seeker.

The Map Problem

I genuinely think this is a technical problem. As I’ve argued before, LLMs don’t understand meaning but triangulate it. They don’t understand words as much as calculate their meaning. GenAI machines don’t understand what they produce in the ways that humans understand those words. They place vectors into the madness of high dimensional space. That said, LLMs are intelligent in this sense: they produce signs (words, images, charts) that have some type of facility with language. Let me elaborate a bit more.

I will call it the map problem of GenAI. The map problem helps explain what I mean by orbital argumentation. The map problem got a lot of attention in the summer of 2025 because ChatGPT is downright awful at making them. Below are two examples that I had GPT5 create during the week of August 18^th, 2025.

Verbatim instructions: “draw a map of illinois with the cities labeled.”

Verbatim instructions: “just create a map of the usa with states and capitals for a presentation”

The first map, of Illinois, is so astonishingly odd that it makes me wonder if the software is functioning correctly. The most obvious observation is that the map is a rectangle. Illinois, to be clear, is not a rectangle; it is rectangle-ish. And that’s important here. My second observation is that the longitude and latitude are fairly correct (I had to double check this but the numbers are pretty-mostly-fairly close) with the location of the cities more or less correct, located somewhat accurately if we imagined Illinois to be a rectangle. These observations bring me around to the orbital nature of the diagram: the map seems incomplete, devoid of an epistemic center while being intelligent about something. It has some type of facility about the task I asked it to do.

The second map is overwhelmingly embarrassing. I decided to have ChatGPT5 produce it after I read that GenAI couldn’t create a map of the USA with the states or capitals labeled. I was so intrigued that I gave it a try myself. As you can see, the labels are downright absurd. But peer at the map a bit. Stare into the crystal ball. Let the abyss take you in for a moment. If we don’t think too hard, and think about it loosely, the map itself, without any labels is pretty good (notice, that Vermont and New Hampshire appear to be one state, but honestly, is that really a mistake?…zing). The Illinois map is egregious with echoes of being correct. This second map has boundaries that are better than the first.

Let’s think about the maps together: what’s so weird: the second map has a rather well contoured map of Illinois: but I made these maps back-to-back. There was no actual connection between the Illinois map and the USA map. The machine started over. Very peculiar no? It’s like it forgot or it just doesn’t understand what it’s doing. Just like a writer using three colons in a row. It’s peculiar because a writer should know better.

(Fun side note here: one of the best books I’ve ever read about technology was Lewis Mumford’s Technics and Civilization. It has so many good points about technology, but its style has always stuck with me: Mumford uses multiple colons in the same sentence: it’s jarring while being poetic. I’ve never encountered such writing before.)

You might be thinking, after rolling your eyes at my overuse of colons, “Okay, John, what does this have to do with writing? I read your blog for writing, not to read about you playing with pictures.” Fair enough. My point here, as it relates to writing: the images reveal the cracks in AI written arguments. Our brains aren’t able to pick up glaring deficiencies with just written words. With images, we can pick out the mechanical things that don’t look correct.

The map problem indicates visually that same thing is likely happening with written words. AI-written sentences might seem fine on “first skim.” But as any good English major will tell you, close reading is required if you’re going to think about the writing. Readers need to be firm.

Herein lies the tricky problem of argumentation in general: readers, the audience if you will, need to think about an argument in order to understand it. An argument is never one-sided; writers can’t just produce an argument knowing exactly how it will be received. The same goes for readers; as readers, we must be practiced at reading carefully and slowly. Pausing to reflect. Actively take notes. Because if we don’t do this, we end up not being able to critique the writing.

A brief defense of orbital argumentation

There are some good things about orbital augmentation, though, as any person stuck in the corporate world will tell you. Orbital augmentation is excellent for bureaucratic writing. AI writing is good for documents that will never be read carefully. Or read at all. AI writing is good at bureaucratic argumentation. I mean this as a compliment, in a genuine sense.

I’m not being sarcastic. If you know me personally, you know that I’m not a sarcastic person. I’ve never been sarcastic. I mean what I say, a product of growing up around loud, obnoxious Philadelphians perhaps. When the Eagles won the Superbowl in 2017, my dad had a shrine to Nick Foles, complete with candles. I don’t think my dad ever made a sarcastic comment. I take after my dad in many ways.

What do I mean by bureaucratic augmentation? The first thing that comes to my mind is required, transactional writing. Think here of reports that are never truly read in the sense of being thought through carefully. Bureaucratic writing is supposed to be vague, orbiting around the point. Bureaucratic writing is often just meant as a record. Think here of meeting minutes. No one truly wants the details of every meeting recorded, often for exculpatory reasons.

Of course, AI writing doesn’t have a monopoly on orbital argumentation. Plenty of adults use the technique, sometimes intentionally, sometime accidentally. Students do it when they haven’t done the reading. My friend Graham in college was amazing at writing about books he’d never read. He’d pick out passages, stitch them together with vague, orbital claims, and usually get a good grade. He used to smoke outside the library in the evening. When passing him, I’d ask him what he was doing. “Thinking,” he’d say, inhaling, “about how to bullshit my way through this paper.”

Human stakes

Orbital argumentation shows what AI writing is missing. LLMs have no stakes in the writing. When a person truly, genuinely argues, they have a position at stake in the argument. There can, of course, be good and bad arguments, with many humans not fully understanding what or why they are arguing. Humans, particularly adult humans, almost never make an argument without a stake in it. We even have a term for it when they don’t: playing devil’s advocate. Or we say they’re arguing in bad faith. Yet, we can’t even say that GenAI is arguing in good or bad faith. GenAI has no stakes in whether Illinois is a rectangle. I do. ChatGPT can’t even be said to be indifferent about what it writes. GenAI remains silent about the stakes of human argumentation.

Expert writers understand that their argument has stakes with a reader. As a result, their writing is shaped by audiences that will read their documents while they aim to persuade those readers. Expert writers, too, consider genre. They consider the ways genres cull audiences into positions to make judgments of their writing. Human writers can’t predict what will always hit. Because writing for readers always involves audiences that cannot be fully known or fully predicted. Audiences are capricious. Fickle. Difficult. Smart. Orbital. They are the heartbreaking delight of human writing.

I hope Annette, Collin, and Justin read this blog post. If they like the post, I’ll smile. If they comment, I’m eager to read their comments. Giddy even. I picture Katherine and Michael reading this essay, perhaps together, maybe they’ll talk about it in the evening over dinner. Maybe Eleanor and Ted will argue over it. I imagine these people as I write.

Human writers desire to be read. We wish to be read widely. Maybe Cara will restack this essay. Maybe Anuj or Trey will share it on LinkedIn. I hope I can write a blog post good enough that I’ll go viral. I have vanity associated with my writing. AI writing just doesn’t have the knack for wanting to be read. Human writers crave attention. Yearn for it. Pine for it.

I want to be paid for my writing, which means I aspire to write something someone would be willing to pay for. I need to, therefore, compose. To craft. I can’t willy-nilly use punctuation. To write well is to leave readers with a gut punch.

Good writers are exceptional at seducing readers to keep reading. Good writers keep their readers moving through a text. They choose unusual words or syntactic constructions or punctuation. Sometimes they miss the mark, getting a bit too weird.

When an essay I write falls flat, I teeter. When a human pours themselves into their writing (I mean truly writes with the hopes of being read widely) and no one reads us, it’s heartbreaking. AI chatbots are never heartbroken. They just march along, writing more. But human writers don’t always want to try again. We teeter, we fall.

Flourishes

I sense, when I read good writing, that a human wanted the writing to flourish. In good writing, I can see the craft. Where the human has made decisions about where to write the clauses or where to position the subject in interesting, maybe even flamboyant, ways. To get there requires a premise: interpretation is not an equation. When we (human beings) write well, there is typically a flourish to it. There are facts and accuracies but usually we add something of ourselves to our writing. The Trace^TM, e.g., a Philadelphian who lives in Illinois. A human leaves traces, hopefully indelible ones. I can’t help but think that the very idea of a static summary, one without flourishes, is a very computer sciency way to think about summarizing. Sometimes I wonder if ChatGPT is revenge from long graduated computer science majors being made to write essays in their English classes. They wanted to know the correct interpretation, so they invented a technology that measured the distances between words.

Essays cannot be distilled in such a way. I know this because of Graham. In my dreams, Graham remains standing outside the library, smoking his cigarette. Graham was a decidedly 90s dude, always wearing flannel with oversized jeans. Scruffy. We had a lot of philosophy classes together. To be fair to him, he did a lot of the readings. But he was a performance artist when it came to last minute writing. I once saw him standing outside the library at 10 pm and then at 8 am the next day. He hadn’t left the library. I asked if I could read his paper. He obliged. It was good, though full of vague claims out of left field, some of which made no sense. A bit orbital.

But there was delight to the weirdness of his essays that I recall fondly. It was performance art. Nay. It was performance writing. Like he was walking a tightrope with words. Sometimes, he made breathtakingly good observations, replete with uneven syntax. His essays were delightful, if a bit wafty, riding on hot air.

If you give me delightfully weird writing, I’ll take a bit of hot air. But I don’t want orbital arguments if they’re boring, blanched tripe. I desperately want to read something by AI that floors me. But I haven’t read it yet. I also desperately want the same thing from humans. But it’s easier knowing there is someone with stakes in the writing they’re producing.

It’s heartbreaking that future writers are learning orbital argumentation without learning to prod their readers into a sense of flourishing. Or to write so hard the writing launches itself into the audience, a gut punch, bombarding the reader, a supernova.