2. Acemoglu, D., Kong, D., & Ozdaglar, A. (2026). AI, Human Cognition and Knowledge Collapse.

3. Benedek, M., & Sziklai, B. R. (2025). Impact of AI Tools on Learning Outcomes: Decreasing Knowledge and Over-Reliance (arXiv:2510.16019). arXiv. https://doi.org/10.48550/arXiv.2510.16019

4. Budzyń, K., Romańczyk, M., Kitala, D., Kołodziej, P., Bugajski, M., Adami, H. O., Blom, J., Buszkiewicz, M., Halvorsen, N. G., Cesare, H., Romańczyk, T., Holme, Ø., Jarus, K., Fielding, S., Kunar, M. A., Pellise, M., Pilonis, N. D., Kamiński, M. F., Kalager, M., … Mori, Y. (2024). Endoscopist De-Skilling after Exposure to Artificial Intelligence in Colonoscopy: A Multicenter Observational Study. SSRN. https://doi.org/10.2139/ssrn.5070304

5. Cloud, A., Le, M., Chua, J., Betley, J., Sztyber-Betley, A., Mindermann, S., Hilton, J., Marks, S., & Evans, O. (2026). Language models transmit behavioural traits through hidden signals in data. Nature, 652(8110), 615–621. https://doi.org/10.1038/s41586-026-10319-8

6. Galdin, A., & Silbert, J. (2025). Making Talk Cheap: Generative AI and Labor Market Signaling.

7. Gerlich, M. (2025). AI Tools in Society: Impacts on Cognitive Offloading and the Future of Critical Thinking. Societies, 15(1), 6. https://doi.org/10.3390/soc15010006

8. Hao, Q., Xu, F., Li, Y., & Evans, J. (2026). Artificial intelligence tools expand scientists’ impact but contract science’s focus. Nature, 649(8099), 1237–1243. https://doi.org/10.1038/s41586-025-09922-y

9. Haverals, W., & Martin, M. (2025). Everyone prefers human writers, including AI (arXiv:2510.08831). arXiv. https://doi.org/10.48550/arXiv.2510.08831

10. Jelson, A., Manesh, D., Jang, A., Dunlap, D., & Lee, S. W. (2025). An Empirical Study to Understand How Students Use ChatGPT for Writing Essays (arXiv:2501.10551). arXiv. https://doi.org/10.48550/arXiv.2501.10551

11. Ju, H., & Aral, S. (2025). Collaborating with AI Agents: Field Experiments on Teamwork, Productivity, and Performance (Version 3). arXiv. https://doi.org/10.48550/ARXIV.2503.18238

12. Kosmyna, N., Hauptmann, E., Yuan, Y. T., & Situ, J. (2025). Your Brain on ChatGPT: Accumulation of Cognitive Debt when Using an AI Assistant for Essay Writing Tasks.

13. Laban, P., Schnabel, T., & Neville, J. (2026). LLMs Corrupt Your Documents When You Delegate (arXiv:2604.15597). arXiv. https://doi.org/10.48550/arXiv.2604.15597

14. Liu, G., Christian, B., Dumbalska, T., Bakker, M. A., & Dubey, R. (2026). AI Assistance Reduces Persistence and Hurts Independent Performance (arXiv:2604.04721). arXiv. https://doi.org/10.48550/arXiv.2604.04721

15. Maleki, N., Padmanabhan, B., & Dutta, K. (2024). AI Hallucinations: A Misnomer Worth Clarifying. 2024 IEEE Conference on Artificial Intelligence (CAI), 133–138. https://doi.org/10.1109/CAI59869.2024.00033

16. Moon, K., Kushlev, K., Bank, A., Lira, B., Viskontas, I., Johnson, D. R., Duckworth, A. L., & Green, A. E. (2026). The Creative Link Between Words and Ideas is Weakening in the AI Era.

17. Nguyen, D. C., & Welch, C. (2025). Generative Artificial Intelligence in Qualitative Data Analysis: Analyzing—Or Just Chatting? Organizational Research Methods, 10944281251377154. https://doi.org/10.1177/10944281251377154

18. Resnik, D. B., & Hosseini, M. (2026). Hallucinated citations produced by generative artificial intelligence may constitute research misconduct when citations function as data in scholarly papers. Accountability in Research, 2645390. https://doi.org/10.1080/08989621.2026.2645390

19. Shaw, S. D., & Nave, G. (2026). Thinking—Fast, Slow, and Artificial: How AI is Reshaping Human Reasoning and the Rise of Cognitive Surrender.

20. Siler, K. (2026). The diffusion of large language models in published academic articles. Proceedings of the National Academy of Sciences, 123(22), e2605754123. https://doi.org/10.1073/pnas.2605754123

21. Sourati, Z., S. Ziabari, A., & Dehghani, M. (2026). The homogenizing effect of large language models on human expression and thought. Trends in Cognitive Sciences, S1364661326000033. https://doi.org/10.1016/j.tics.2026.01.003

22. Tan, M., Phalen, L., & Demszky, D. (2026). Marked Pedagogies: Examining Linguistic Biases in Personalized Automated Writing Feedback. Proceedings of the LAK26: 16th International Learning Analytics and Knowledge Conference, 568–577. https://doi.org/10.1145/3785022.3785113

23. Tomlinson, K., Jaffe, S., Wang, W., Counts, S., & Suri, S. (2025). Working with AI: Measuring the Occupational Implications of Generative AI (arXiv:2507.07935). arXiv. https://doi.org/10.48550/arXiv.2507.07935

When I think about this question, my mind always drifts to the stark distinction between my meals. During the work week, I eat my lunch at my desk, sorta watching YouTube while sending emails. It’s entirely dissatisfying. I loathe it. What makes it worse is that I eat horrifically fast. I’m usually done with my lunch within 5 minutes. (That’s actually a lie: I finish eating within 2 minutes.)

I contrast the way I eat my lunch with the way I eat my dinner. My family always eats together. Usually there are five of us. It’s drawn out. And boy do I enjoy it. We tell stories about our days, we share things that went well or poorly. I frequently share stories of my childhood (swimming with childhood friends Timmy, Dave, and Jeff) or college or just before the kiddos were born. I love story time. Kids love story time. I love listening to stories. I love telling them. Even when we have breakfast and lunch as a family, rarely do the kinds of stories get told like they do at dinner. My favorite part of the day is dinner.

I frequently wonder if there is a biological imperative to telling stories around dinner time, that is to say, in the evening. Dusk. It’s primordial, for danger lurks in darkness. Polly W. Wiessner has written about daytime versus evening conversations in “Embers of society: Firelight talk among the Ju/’hoansi Bushmen.” Here were her findings:

…day talk centers on economic matters and gossip to regulate social relations. Night activities steer away from tensions of the day to singing, dancing, religious ceremonies, and enthralling stories, often about known people. Such stories describe the workings of entire institutions in a small-scale society with little formal teaching. Night talk plays an important role in evoking higher orders of theory of mind via the imagination, conveying attributes of people in broad networks (virtual communities), and transmitting the “big picture” of cultural institutions that generate regularity of behavior, cooperation, and trust at the regional level.

Wiessner’s research indicates that, to some degree, humans developed a distinctive nighttime appetite for higher-order stories, that is, for the big picture imaginative theory of mind. For thousands of years, long enough for instinct to emerge, we humans talked about practical stuff during the day but imaginative stuff in the evening.

Contrast Wiessner’s findings with doomscrolling. People doomscroll the most in the evening. A lot of people say it’s their routine before bed. Sometimes for hours.

This leads me to a wacky theory: people are doomscrolling because they’re hungry for evening stories. They’re pining for long-form, elaborate storytelling.

But most online content is short form. Much of it, too, is complaining, outrage, or worries. In other words, most online content is daytime stuff that Wiessner writes about. The practical stuff. In fact, I think short form content can’t really ever be a story simply because it’s not long enough for a narrative or, dare I say, ritual.

My favorite part of college was hanging on Friday or Saturday night, usually after midnight. Somewheres between midnight and 3 am. I was part of a club called CALVIN and HOBBES (Creating A Lively, Valuable, Ingenious, New Habit Of Being at Bucknell and Enjoying Sobriety). We would just sit around in the living room of the club’s house, talking, chatting, storying. We’d talk about our love lives or classes or just the random drunk people walking by. We’d share our dreams post-college or fun stories from high school. This was 2002-2006; no one had phones on them. I always remember how the evening would end: we would talk ourselves into silence. We’d run out of stories. Breathless, sometimes exhausted. The evening would end with our stories. I always felt good going to bed, drained from socializing but filled with human connection.

I imagine how this plays out now, though. People just watch videos until they fall asleep, half formed thoughts never emerging from their lips because they consumed too many videos. Empty, without social connection but still drained from the practical concerns they were marinating in when compelled to consume content.

Instead of doomscrolling or being on screens, I think we need to spin some yarns. Face to face. At the end of her article, which was published back in 2014 and only truer now, Wiessner writes:

…artificial light and digital communication are invading the night worldwide, turning hours of darkness into economically productive time and overriding social time and story time. The day then ends with the flip of a switch, without taking the time to revisit, explore, ponder, and repair relationships, or let the issues of the day fade with the embers.

I know for a fact that people yearn for getting rid of their screens. But we need to replace the screens with something.

That something is people, the social intimacy of being with others, sharing our sagas, listening, pondering until the embers burn out the day’s close.

1) I value intellectually ambitious writing that pushes on an idea. I want the writing to explore in surprising crevices, explore ideas in non-cliché ways.

2) I want writing that makes me enjoy reading it.

3) I like interesting, unusual words and phrases. I prefer awkward intellectually challenging ideas over flat, boring prose that is mechanically correct.

4) I value writing that avoids academic jargon.

5) I have no patience for fluffy writing (AI or human) or bullshit (writing that sounds good but that once I think about it, means nothing).

6) I value deeper engagement with a fewer number of readings and ideas than a larger number of empty citations or quotes. I do not value citation chains at all (A citation chain is: AUTHOR, YEAR; AUTHOR, YEAR; AUTHOR, YEAR). Examples are always better. Engagement does not have to be agreement; in fact, I value disagreement if the opposing ideas are fairly represented. Part of education is vibrant disagreement.

7) I prefer direct quotes over paraphrasing.

8) I value writing with first person.

9) I do not need writing that comes to a nice, tight ending or “wraps” up ideas. Don’t seal ideas into boxes. I prefer complexity and messiness over stale “correct” endings.

10) I don’t police AI writing. I police good writing that meets the values I’ve laid out above.

Misunderstanding 1: GenAI software (such as LLMs) will automate qualitative analysis (QA) because it can automatically identify codes and themes.

• Correction to misunderstanding 1: While GenAI software can create descriptive tags (i.e., the lowest level of qualitative analysis), these tags are not reliable. These tags can in turn pollute a researcher’s perceptions of the qualitative data. GenAI software cannot generate codes (second level) or themes (highest level) because codes are inherently derived from the researchers’ interactions with the data. QA is highly idiosyncratic and not meant to be replicated.

Misunderstanding 2: GenAI software facilitates qualitative analysis by increasing the speed of analysis.

• Correction to misunderstanding 2: GenAI software short-circuits qualitative analysis by disrupting researchers’ process of familiarization. The automatic generation of tags means the qualitative researcher lacks epistemic familiarity with messy and complex data; the automatic generation of tags is thus a quantitative process. (In this sense, it’s no longer QA but it is analysis.) When tags are generated by the qualitative researcher (and not a machine) this process helps the researcher to learn about the inquiry under study, thereby embedding them in the data’s context. AI-driven software is not embedded in the context of the researcher’s questions or aims.

Misunderstanding 3: GenAI software can be used as a preliminary analytical tool.

• Correction to misunderstanding 3: GenAI software surfaces descriptive aspects of qualitative data. These tools are not analytical; they cannot make analytical or normative judgments. They are also not embedded in research questions and do not represent the researcher’s specific subject position. These tools are therefore most useful after a full analysis to help identify places where the researcher may have been biased (or just missed).

Misunderstanding 4: GenAI software augments researchers’ qualitative analysis.

• Correction to misunderstanding 4: QA necessarily draws on researchers’ subject position for epistemic understanding. GenAI models culture not individuals (by definition). These models are not a reflection of a researcher’s subject position but a fusion of average viewpoints that undercut QA. The researcher's experiences in performing qualitative analysis are neither transferable to a model nor reducible to objective averages.

Misunderstanding 5: GenAI software can help make the generation of a qualitative codebook more efficient.

• Correction to misunderstanding 5: The codebook cannot be automated for it is the site of epistemic learning. The codebook is inherently inefficient. (That’s where we learn fun and interesting stuff!)

Misunderstanding 6: GenAI software can function as a rater in terms of interrater reliability.

• Correction to misunderstanding 6: Interrater reliability is more than the transaction of the rating. Interrater reliability must involve human beings; it must involve training them. Interrater reliability is significantly more complex than coming to agreement; in the training for ratings, human raters come to new analysis as well as synthesized knowledge between raters (this prevents qualitative analysis from being solipsistic). Most interrater reliability use cases overlook this epistemic knowledge building.

Thanks for reading! Happy to answer questions.

When I began writing this essay, I originally titled it, “The need for good writing.” Now, I hadn’t actually written this essay yet. All I had was the title. It’s a weird habit of mine, titling things before I’ve written them. I felt the word “need” wasn’t good enough. I thus googled, “need synonym.” After clicking on the Merriam Webster link, I encountered “desideratum.”

Now that’s a glorious word. On the surface it means highly needed or desired. There is a famous poem, as I would find, by Max Ehrmann called “Desiderata”. I ended up looking up the etymology on the Oxford English Dictionary (OED). The OED told me, “Desiderate typically occurs about 0.02 times per million words in modern written English.” To understand how rare that is, let’s calculate that in terms of people on earth. That would be about 160 people around the entire globe. It’s so outlandishly rare that we might consider it aberrant. No AI tool would ever give you that word. A human would barely give you that word.

“Desideratum” has roots in astronomy, meaning to look at the stars. A more accurate translation, though, is closer to a longing for the stars. Here is the OED entry:

“Latin dēsīderāt-, participial stem of dēsīderāre to miss, long for, desire, < dē- (de- prefix 1a, 1b) + a radical also found in con-sīderāre, perhaps connected with sīdus, sīder- star, constellation; but the sense-history is unknown: compare consider v.”

Desideratum is a need, a consideration, a longing, a desire all rolled into one. That’s the word I’m searching for.

Well, dear reader, you are probably curious why I didn’t choose desideratum for inclusion in my title. Good question. My second title was, indeed, “The desideratum of good writing.” I desperately wanted to title the essay with this marvelous word. But if I was going to share this essay online, not enough people would have any idea what that word meant. There would not be enough resonance with readers.

The choice between desideratum or resonance for this blog post was a frustrating decision because it risked being unread. The most devastating thing for writing: to be unread. The most devastating thing for writers: to lack an audience. There are real consequences to tweaking a single word.

Writing is filled with dangerous choices. Dangerous because writers must make choices with real consequences. There is a brutality to good writing. These kinds of consequences don’t really exist within classroom settings in which teachers are the only true audience. For real writing, an audience must willingly drink in the writer’s craft.

The morass

This brings me to an emerging problem related to writing: we’re stuck with a morass of fancy writing, reports and texts that are verbose because writers have made no choices, taken no real risks. It’s not so much that the AI era ushered in bad writing as much as it accelerated the acceptance of bad writing. Everyone can now sound like a fancy academic writer, which itself is a problem. I’m convinced that AI writing is primarily trained on some zombie fusion of academic writing and Silicon Valley hype. Part of my belief is that the internet just isn’t a good enough place for these companies to train their models. What is good enough? All of those academic papers that are hosted on various sites (arXiv and a bunch of other academic paper repositories).

I often disabuse friends and family that the fancy, long sentences of academic prose are valuable. Academic prose often succeeds despite its writing, not because of it. Professional editors have known this for some time. Donald Holden, a deceased editor in the New York publishing scene, wrote of this issue in 1979. I quote this at length because it’s just a devastating take.

Like most of my friends in book publishing I groan when I get a thick package from a university.

Inside the envelope, I know there’s a manuscript from some professor who’s bucking for a raise, a promotion, tenure — a manuscript that serves no one but the professor himself and some faculty committee.

I know that the book will be written in the illiterate scholarly jargon that publishers (and cynical graduate students) call “dissertationese.” Because we try to give every author a fair hearing, my staff may spend hours putting the manuscript through its paces. But in the end, I know that almost certainly the book will be rejected.

Publishers know that most professors are bad writers. The professor’s grammar, spelling and punctuation are usually passable, but his prose is apt to be pretentious, unclear and chaotic. And like most people who write badly, he does it on purpose. As George Orwell points out in “Politics and the English Language,” bad prose doesn’t happen by accident: It’s a conscious attempt to deceive.

“Dissertationese,” the standard lingo of doctoral dissertations, is the language that professors use to disguise self-contempt with pomposity. The scholar, who often spends his professional life counting the cats in Zanzibar, secretly knows that what he has to say is unimportant. So he inflates his diction, convolutes his sentences, and builds a complex scholarly structure (with lots of footnotes) to make the content sound important.

Faculty committees — who grade him on how well he plays the academic game — may be impressed, but publishers are not.

The publisher also knows that bad writing usually hides (or tries to hide) intellectual laziness. Every experienced editor can tell you the real purpose of those turgid, ponderous sentences: They’re designed to conceal the fact that the writer hasn’t taken the time to figure out precisely what he wants to say. And why is the professor’s “story line” so chaotic and bewildering? Not because his ideas are so lofty, but because he’s dodged the writer’s toughest intellectual job — to link his facts in a tight, logical chain.

Holden describes this in a rather vicious way: there is a cloak of obfuscation to academic discourse. And the corporate world seems to be leaning in to this cloak, having learned it from GenAI tools, tools that learned it from academic papers (or books essentially stolen by AI corporations). The adoption of this inflated language is particularly pernicious in corporate environments. The vague, ambiguous writing of GenAI excuses bad management thinking (including goal-setting and organization) by shifting burdensome tasks to workers, thereby excusing managers from culpability. Workslop has become the name for it.

There is a more damning problem that Holden gets at with respect to bad writing: bad writing doesn’t have anything important to say. It doesn’t make people interested in reading it. In Holden’s case, the style of prose is so impenetrable that readers (outside of a tiny select few) won’t take any interest.

Bad writing conceals how little is being said. Although I like the word desideratum more, that word obfuscates my meaning, hides meaning more than I want. Bad writing is boring drivel precisely because it has nothing to say. AI writing is empty connective tissue, filled with ands or buts or ifs without function because it says so little. The ideas in AI writing aren’t fake but rather fluff. But humans have been writing fluff for years. CEOs love fluffy reports that sound like they say something. The only difference is that CEOs can now use ChatGPT to write it for them.

On the other hand. Good writing gets a reader excited precisely because it has something to say, even if it’s examining a remarkably obscure word. Good writing makes the audience learn something, sometimes, perhaps oftentimes, against readers’ will. Great writing is a thermonuclear experience: you laugh out loud or throw the book into the air out of sheer excitement.

The erosion

I suppose this brings me to why I dislike AI writing: it doesn’t push. Even after people have iterated with AI writing tools, thereby adding something of themselves to the text or manuscript, it misses something: intellectual ambition. Their ambition has been eroded or corroded by the machine’s template. AI tools use predictive statistics to create their output. They create sentences by putting together words and phrases that are likely to have been put together before. They don’t savor words or speculate about trifling (or onerous) choices. Machines don’t pine to be read.

If you read AI writing long enough, you see the din of the templates, so overused… once human constructions but no longer (“not x but y” or phrases like “ever evolving landscape”). That’s why everyone dislikes AI writing. All of these constructions were once human but used sparingly. AI tools overwrite to cover the hollowness of their ideas. We dislike AI writing because it’s merely templates of writing. Empty templates, filled in by echoes of the past, echoes learned by machines that write using calculation. A writing machine never attempts to move an audience into a thermonuclear experience. Great writers ignite readers with their words.

Good writing resonates

How do we get to good writing, regardless of whether it’s AI generated or not? First, we must disregard whether the writing is AI or not. These models learn from human writing. We don’t like AI writing because it’s bad. We just need to say that and move on. If you’re a teacher, don’t dwell on whether students used AI. No need to police students or corporations or social media. Dwell on your own sense of good writing. Dwell upon good writing.

Second, reject fluff in the writing. We must extirpate (rip it out by the root) vacuous writing. We stop caring whether it’s AI fluff or human bullshit. We just say, “Good writing is good writing. I don’t accept vacuous ideas from anyone anymore.” Let’s demand a stop to ponderous sentences, as Holden advised. We demand a stop to writing that obfuscates, to writing that is mere performance. Demand ambitious writing.

Third, good writing is in the eye of the reader. Which means that good writers must consult with potential readers. Good writers seek out readers. There is a reason why editors exist. They aren’t just copyeditors or gatekeepers. Editors help writers develop. They offer biting, caustic, and necessary advice. Too many writers I know bristle at hard feedback. I yearn to be annihilated by editors. My friend Aaron made me cut the first section of this essay in half. He was correct. Annihilate my writing so that I might become a better writer.

Fourth, good writing isn’t private. It demands the writer to consider others. In turn, this aspect of good writing means that writing is so much more complicated than thinking. Good writers consider audiences and genre expectations, for example. All of this means that good writers aren’t solipsistic. They don’t compose in the attic, locked away from others. This means you have to think about how much readers are willing to give you, as a writer. Whether they’ll tolerate academic puffery or whether they’ll get impatient, ready to close the page or log-off.

And finally, the most difficult thing of all: good writing must have something to say. One criticism I hear from AI advocates is that most writing isn’t good. They’re right. Most books aren’t good, most articles are forgettable, most blog posts have an inevitable obsolescence. Perhaps that is the very nature of writing, though. Good writing doesn’t necessarily need to stand the test of time.

Good writing must instead resonate with readers, reverberating with audiences, rattling around in readers’ heads. It needs to make them think. Shake them. Now at this point, I could quote you a bunch of citations about resonance theory (Hartmut Rosa’s Resonance and Amanda Lotz’s After Mass Media are two good starting places). But you, my reader, have slogged your way through this essay. My one and only goal is to make you consider good writing as a reader. For good writing requires readers to long for more. The resonance of good writing is the desideratum of good reading.

Jinx is a game when two people say the same word at the same time. After they say the same word, the first person to say “jinx” wins. Then, at least in the version I play, the loser can’t talk until the winner frees them. My kids and I were playing jinx recently. The younger one, who is in kindergarten, learned the rules wrong: in this version, jinx was where people say the same word kinda close to one another. I said, “look at the ducks on the pond” and then my kiddo said that phrase about 10 seconds or so later. I got jinxed by said kindergartener.

That was the game of jinx in the kindergartener’s mind. Something was being learned about words being spoken very, very, very close together but it wasn’t accurate to the way that most of us understand the game of jinx. I explained it accurately, dear reader, if you’re wondering. I cleared up the miscommunication over frozen waffles at 6 am. My kindergartener likes to eat them frozen, not toasted. Never fear, for we can all now play jinx correctly.

What exactly happened here though?

My child learned something, and I think that something is quite elaborate: matching patterns doesn’t always mean something, even if the pattern is non-random. If we were educational theorists, we might say the child is in the preoperational stage (hello Piaget!) with an understanding of the broad concept (people saying the same words close in time) but not realizing the central rule of jinx (saying the word at the same time). That’s quite clever.

Small, nascent humans do this all the time because they are quick to identify non-random patterns. Adult humans do this, too. We as human beings are quite ready to pick up on patterns even when there isn’t one; really marinate on all the times when something has happened to you but twice in a row. You immediately generalize it to a pattern. Happens all the time with my kids, myself, students, my colleagues, my family. Humans are pattern matching savants, often to their own detriment. They can often spot the disruption in patterns too.

To elaborate on my point take the line of random characters below:

In quote 1, I literally just bashed my fingers over the keyboard. I later added in the space between every five letters. Those spacings make it look like there was some sort of pattern there. How about the second example below. What do you think? Is there a pattern?

Here there appears to be a little more of pattern, right? The letter “f” is pretty prominent. Sorta appears regularly. Maybe a pattern? Eh…there was a rhythm to my motions: I slid my finger, unevenly, over the keyboard, which results in a non-random pattern-like recurrence. A pattern-like recurrence is when there is believable recurrence but no underlying meaningful pattern.

This “pattern recurrence” is why AI writing is so difficult to challenge for not being very good: the words that LLMs manufactures are definitely NOT random. They make coherent statistical sentences to the human mind that seeks patterns, a mind that yearns for patterns. LLMs are non-random.

And, again, as I keep arguing throughout this blog, it’s quite impressive that LLMs can harness language as much as they do without having any idea what written words mean. LLMs don’t know what a dog is. They operationalize the letters D-O-G, identifying corresponding words in the training data. They have no idea what the word “jinx” means but they have associations of that word with other words in the training data. That’s pretty neat. As someone who has read a lot of philosophy and post-modern jargon, I find it fascinating. I just don’t like, though, that somehow this ideology has hijacked our media, political, economic, and educational institutions. Anyways. LLMs produce non-random patterns of writing that don’t have any correspondence to reality. They’re written words, all the way down.

Here is where we encounter the first of three pickles: we as human beings want to see patterns even when they’re meaningless. For patterns can still happen even if they have no meaning. Let me show you how LLMs might do this with a sentence I just wrote. In the previous paragraph, I finished a sentence with this clause: “social, political, economic, and educational institutions.” That sounds like I’m making a very grand claim when I’m really saying, “everything.” And I was. But what I’m really doing here, as a writer, is being totally unspecific. Contrast this with my highly idiosyncratic story about jinx (including those frozen waffles; reader, I don’t get it either but humans are marvelous creatures). LLMs, if they wrote that sentence, write “social, political, economic, and educational” because the training data just says so. In order to make those sentences become non-random patterns, LLMs write with innumerable vague, median words, strung together, hoping human users ruminate on them, thereby imbuing them with meaning. When I read phrases like “dynamic interplay” or words such as “insights” from LLMs or human beings, the profligate vagueness of those words makes a reader squint, wondering about meaning where there isn’t any acumen.

This brings me to the second pickle: the acceptance of non-random writing patterns, whether it’s from LLMs or from people. A shrewd realization of novice writers is when they finally realize they must largely cede control of their writing to readers. Like letting go of a child on their first day of school. For readers imbue words with meaning.

The long-term consequences of accepting non-random writing patterns is that you become dulled to them. You start ignoring writing in general, seeing the detritus everywhere, acclimating yourself to the tedium. This kind of thing is actually what I’m truly fearful of. A world filled with bureaucratic writing that sounds good but doesn’t have any meaning. Like a corporate memo sanded down by lawyers and PR specialists until it’s more vapid than a ghost. And, because people are so adjusted to the jetsam, they start to think words don’t matter, that word choice doesn’t matter, that style doesn’t matter, that good writing is irrelevant. Writing itself becomes opprobrium. Regardless of whether it’s AI writing or human.

But do you see how I used those words “detritus” and “jetsam” in the previous paragraph? I meant them. Detritus isn’t garbage; it’s not a synonym for trash. It’s actually much closer to fertilizer, an organic compound. For words, even those generated by AI machines, are a type of organic sloughed off from human beings. And jetsam is not flotsam; this website describes these two terms well:

“Flotsam and jetsam are terms that describe two types of marine debris associated with vessels. Flotsam is defined as debris in the water that was not deliberately thrown overboard, often as a result from a shipwreck or accident. Jetsam describes debris that was deliberately thrown overboard by a crew of a ship in distress, most often to lighten the ship’s load.

Under maritime law the distinction is important. Flotsam may be claimed by the original owner, whereas jetsam may be claimed as property of whoever discovers it. If the jetsam is valuable, the discoverer may collect proceeds received though the sale of the salvaged objects.”

I used jetsam with a type of irascible deliberation. I thought hard about my words: ground my brain for that word: I didn’t just drag my fingers over the keyboard. I used it because AI writing has made all writing weirdly, strangely democratic: words are just everywhere now, sprayed about, throw overboard on the internet. Anyone can salvage them for their own use, AI or human.

And now we arrive at the final pickle: non-random writing before AI was a pretty good sign someone was attempting to tell you something. Sure, there was the corporate jargon or academic post-modern nonsense, but there were still people behind the writing. Now with AI writing, you begin to think that non-random patterns aren’t worth your time because you can’t be sure anyone is there. If no one cared about the writing, who cares about reading it? You become solipsistic to the point where you aren’t sure if anyone else’s mind exists but your own.

That’s why conferring about the rules of jinx is so important. It’s not that jinx, or any game really, can’t be played differently. There really isn’t the correct way to play it. The point is to play the game together.

For that’s the salve on the wound: humans write all the time in non-random ways that are pretty meaningless. I have written many things that didn’t mean much when I initially wrote them out. The only way to figure out your meaning is to ask other humans, converse with other marvelous humans, chatting, even if it’s over frozen waffles at 6 am.

Introduction

Based on the AI hype, you’d think we have full blown robots doing all the work. Forget that nearly every robot-building problem is a materials problem as much as it is a problem of computers. I digress. Do I digress too often? It is annoying when I do? Perhaps. But I care about my audience feeling something, even if it’s annoyance. The last thing I want them to do is skim, forget my writing.

Alrighty, let’s focus: The managers love AI. They very nearly prance about with the pronouncements. Every middle manager seems to be advocating for GenAI writing while the people under them sorta roll their eyes. What gives? Why the difference?

The most cynical, ungenerous reading is that the managers aren’t actually doing any work. If people just did their work without any conflict or stayed on task, the managers would be the ones out of a job. The dirty secret in all of management is that the people most exposed to AI layoffs are the managers, not the makers.

I wanted to know how a GenAI tool would respond to the question, “Why are middle managers and administrators so excited about AI writing?” So I decided to ask Claude why they like GenAI writing. Here is what the machine wrote:

“The most obvious one is that AI writing produces text that looks like competent institutional prose — it’s fluent, organized, uses appropriate register, and checks the surface-level boxes. For people whose primary relationship to writing is managerial rather than compositional (meaning they evaluate and approve text rather than craft it), that surface polish is often the main thing they’re scanning for. AI output passes the “does this look right” test with flying colors precisely because it’s trained on enormous quantities of exactly that kind of professionally adequate text.

But there’s something deeper going on. Middle managers and administrators often occupy roles where writing functions as documentation and compliance rather than as thinking or persuasion. The memo, the report, the strategic plan update — these are frequently performative artifacts. They need to exist more than they need to say something. When the primary function of a text is to demonstrate that a process occurred or a box was checked, the fact that AI can generate it effortlessly isn’t a bug — it reveals something about how hollowed out the communicative purpose of that writing already was.

There’s also a status and labor dynamic. These are roles where being seen as efficient and forward-thinking matters enormously, and where the actual craft of writing has never been a core professional identity marker the way it is for, say, a journalist or an academic. Embracing AI writing tools signals technological savviness without threatening any part of how they understand their own expertise. Compare that to how writers, researchers, or educators respond — people for whom the process of writing is bound up with the process of thinking feel something genuinely at stake.

And then there’s the control dimension. AI-generated text is predictable, on-message, and tonally safe. It doesn’t have the rough edges, idiosyncratic voice, or occasional inconvenient honesty that human writing carries. For someone managing communications up and down an institutional hierarchy, that blandness is a feature. It minimizes risk.

The uncomfortable implication is that enthusiasm for AI writing in institutional contexts often correlates with contexts where writing wasn’t really functioning as communication in the first place — where it was already closer to ritual. AI just makes the ritual cheaper.”

I agree with mostly everything here. If you’re a manager who writes prose that needs to say very little or just be performative, then GenAI writing is perfect for your purposeless purpose. If no one is really going to read your work, why not use GenAI? It’s all just a transaction where the artifact needs to exist regardless of quality or process.

Here’s the problem: the perspective of the managers misses the makers. If you’re a maker—i.e., someone who writes or makes images or produces content, in short someone who makes things, who makes art or crafts prose—then you aren’t excited about a machine doing parts of your job. (But as a reminder, the machines aren’t thinking.)

…Hold up a minute…

Did I really just quote Claude for five paragraphs? Wasn’t I excusing my own thinking about an issue, just letting the machine do it for me? Did you skim it? Maybe I wore you, my thoughtful reader, down to believe those reasons Claude gave me were pretty good. Did it work? If it did, it probably worked through what Steven D. Shaw and Gideon Nave call “cognitive surrender” in a recent publication. This is an excellent paper that I spent five hours reading. I recommend you read it. If you can print it out, do that. Read in the depths.

What is cognitive surrender? The authors write that, in their experiments, the AI system “reflects a reallocation of cognitive control rather than mere effort saving. [The AI system’s] fluent, confident outputs are treated as epistemically authoritative, lowering the threshold for scrutiny and attenuating the metacognitive signals that would ordinarily route a response to deliberation. In the case of cognitive surrender, there is a shift in the locus of control, with an external system [The AI system] occupying the default position.” (p. 45).

Let’s tease two words out: scrutiny and attenuating. First, AI lowers the threshold for scrutiny; in other words, AI systems make the user (i.e., a human being) less discriminating, less picky. A GenAI invariably wears out a human being over time. I get worn down by Claude. Maybe you got worn down by those five paragraphs I copied. I’m sorry I did that, by the way. I promise I won’t do it again.

Second, the AI system attenuates the users’ metacognitive signals for deliberation. Explained another way: over time, my own ability to think about thinking, to deliberate with my own thoughts, weakens. I’ve had several friends and students tell me they don’t feel so sharp after using an AI tool. I liken it to riding in a car. After a long time in a car, your legs are stiff and sore. But in the case of an AI tool, instead of your legs, it’s your brain. Your brain gets stiff, sore, crampy, in need of free space, in need of the joy of thinking hard.

I can’t help but draw a connection between the cognitive surrender of using GenAI to write and the manager who just reads reports in a bureaucratic fashion. You see, the manager doesn’t make the report, for the manager is only skimming it to make sure it looks right (I mean looks in the literally sense of sight), sounds like it means something because they aren’t crafting it. I think the same thing happens with Shaw and Nave’s study as well as my own use of Claude.

Managers just read to make sure everything seems like it’s in order because they assume the stuff they’re reading is correct. They surrender to the meaning of the text, instead making sure it looks like from a literal visual perspective. I feEl like I’Ve HaD ThIs ExpeRiEnce bEfore? I’ve had a parallel experience where I gloss over things I’m reading.

Academic literature reviews

I’ve had this managerial experience when reading literature reviews in published papers. Every academic I know has skimmed their fair share of literature reviews. We’ve read very bad ones that proclaim they’re filling some “gap.” (I have a colleague who once joked to me that “maybe some gaps don’t need to be filled or that we can just step over some gaps.” I always chuckle when I think of someone just stepping over a research gap.)

Certainly, there are lots of mediocre literature reviews, where the author kinda meanders through the background section, writing standard assertions, occasionally identifying a good paper, likely citing the canonical ones, either because the author is supposed to cite it or simply because a canonical paper (by definition) is relevant for citing. I’ve done this plenty of times.

We’ve read plenty of good literature reviews, too. A good literature review synthesizes beautifully. It crafts a conversation together that exists while creating the conversation. It can pull in seemingly unrelated threads or find obscure papers (especially those pesky books and book chapters that aren’t indexed by the internet). For you see, a scholarly conversation may exists external to a publication but no literature review exists outside of the paper. A literature review thus is as much constructed as it objectively exists.

Why am I harping on this?

Because as a consumer of writing, it’s easy for me to just read a literature review from a transactional point of view: what does the literature review do for me, the reader? As readers, we often take a managerial understanding of the writing of a published paper. We readers often do not even realize we’re taking such a transactional view. We don’t stop to consider the maker view of writing. I saw this managerial view on display in February of 2026 when a professor at Dartmouth posted this on social media:

I’ve been thinking about this post for long time. There were a lot of people who agreed with the sentiment, along with a lot of people who weren’t engaging with Dr. Westwood’s sentiment in a genuine way. The post made me ponder as much about the literature review (and peer review) as it did about AI.

The post divorces the literature review away from the results. It takes a solidly managerial, a commodity view of the academic paper. It ignores the maker view.

It forgets that there are many reasons for a literature review to exist for the writer. In order for studies to exist, the researcher isn’t starting from scratch, nay they cannot start from scratch. I replied to Dr. Westwood’s post with the following: “I suppose I see my writing of the literature review as analytically integral to my studies’ designs, including as a guide for what I can contribute. The review and analytical schema are deeply related.” He never replied. This deflated me because I think he saw me as being sarcastic. He probably thought I wasn’t engaging with his ideas. This saddens me.

It’s difficult to identify when someone engages your ideas seriously, particularly with social media. But it’s just as tricky in literature reviews. At a transactional level, literature reviews are performative: we cite things because we’re supposed to cite them or because we’re told to cite them (by editors, certainly by peer reviewers, often by our colleagues).

It’s easy to imagine readers skimming the literature review because readers only care about results; they see the results as the contribution, not the entire scholarly paper. But a literature review represents many things I read along the way (and only a small part of the stuff I read!) to form my study, or coalesce my ideas into something a reader can, well, read. Without the literature review, my results might still exist but they lack context. My papers, without the literature review, lack the “why does this stuff matter” claim.

I can’t help but think that Westwood’s view is the logical conclusion of the managerial view of writing. If you really think the literature review is vestigial, then you think of much of the writing humans do as managerial. Custodial even. And this ends up with a misguided belief in transactionalism. You view the literature review as a perfunctory means to present the paper. It’s just checking a box.

Same goes with genres of workplace documents. The annual report becomes not a representative of a worker’s diligent labor but just a document that some higher up will skim and scan without much consideration. That memo? No one is going to read with the exception of maybe lawyers in case there is a lawsuit.

A maker view of writing

There is another view here: the maker view.

Because workplace documents actually matter sometimes. When I write my annual report (me, John now, not the royal “I”), I like doing it because I get to think about all the stuff I did. When I write up the literature, I get to cite all of the papers that helped me work through my data or my argument. Not everything can be distilled down into a product bought and sold.

For a life well lived cannot be entirely about transactions. Politics, writing, art, family, insert another feel good abstract concept here. We don’t do these things just to get them done. I don’t want to speed through my craft, my makings. I don’t want to speed run being a parent. I actually greatly enjoy my teaching; I’m a little forlorn when a class period ends.

Making isn’t about transaction. To be tautological, making is about making. And yes, we do want rewards for making our stuff. But we often enjoy producing our products more than we enjoy consuming. That’s why the managers love AI writing but the writers don’t. The managers want the writing to be done. They want the document to be in their inbox. The writer wants to keep writing, hacking, polishing their words. A writer can’t skim their work. If the manager cares solely about the product, then the maker gets lost in the process.

Here is a quote from the abstract:

Recent analyses revealed that scientific publications already exhibit evidence of AI-specific language. But this evidence is inconclusive, since scientists may simply be using AI to copy-edit their writing. To explore whether AI has influenced human spoken communication, we transcribed and analyzed about 280,000 English-language videos of presentations, talks, and speeches from more than 20,000 YouTube channels of academic institutions. We find a significant shift in the trend of word usage specific to words distinctively associated with ChatGPT following its release. These findings provide the first empirical evidence that humans increasingly imitate LLMs in their spoken language. Our results raise societal and policy-relevant concerns about the potential of AI to unintentionally reduce linguistic diversity, or to be deliberately misused for mass manipulation. They also highlight the need for further investigation into the feedback loops between machine behavior and human culture.

The article is short, so I recommend reading it in full (something like 9 pages). Pages 3-4 have the best figures (Figs. 1-3). I’d like to see how these finding compare to non-recorded audio for social media (meaning the speakers know it’s recorded) or academic language for the following reasons:

1) This kind of recorded orality could in fact be written, which in turn could be that those speakers are using ChatGPT to write their script. How then does spoken impromptu language function compare to this study?

2) My sense is of GenAI is that it uses (perhaps democratizes) the use of fancy-sounding academic words, such as (as the paper argues) “underscore” or “intricate.” These two words (along with “meticulous”) are fairly common academic words (”I underscore [followed by hedges]).

All of those caveats aside:

I’ve been talking about the influence of GenAI on our spoken language with my students since 2024. I’ve noticed that the word “insights” is leaking into my meetings and class discussions. I never really encountered “insights” orally or in writing unless it was in a methods book. Before GenAI, people typically used the word “findings” or “claims.”

Personally, I find the rhetorical difference fascinating. Insights assumes something of note or worthwhile. Findings/claims don’t have the automatic value, meaning it’s ready to be critiqued.

These writing desires crystallized when I read Michael Crichton’s Jurassic Park. My mom and I struck a deal: I could see the movie (summer of 1993) but not until I’d read the book. After reading hundreds of pages, I knew then that writers needed the ability to produce a lot of words. After reading Crichton’s dino novel, I read a novelization of Tim Burton’s Batman Returns. The book was alright but I thought I could do better.

So that’s what I did. I set about writing all sorts of tripe. I ended up writing about Batman fighting a Toxic Thing. I wrote a story about a summer camp invaded by aliens. I wrote by hand. None of my stories made much sense.

I never stopped writing. I firmly believe that writing requires repetitions, or reps, just like with weightlifting or math problems.

Would you believe that as an adult I have written five full (but unpublished, terrible, no-good) novels. Here are the five:

1. The World Below (84,000 words), a fantasy novel about a Roman soldier who sold his soul to God but the Devil granted him immorality. It’s set in modern America. It’s a retelling of mythology but with all of the classic gods reimagined. Death had blue skin. I wrote it as an undergraduate, somewhere around 2003-4.

2. Oak County (74,000), a novel of a man who returns to his hometown after being gone for 20 years. He goes to work for an opulent Harold Teague, who is styled after Jay Gatsby. I wrote it during the first year of graduate school, 2006-07ish.

3. Washing Away (58,000 words), a novel about a drug addict who meanders around the world, finding enlightenment with other drug addicts. Each drug addict tries to get the other drug addicts to get off the sauce. I wrote it between 2009-2011.

4. Falling Words (70,000), a novel of a preacher who deals drugs. I wrote it around 2016.

5. A young adult fantasy novel for my kids during COVID (60,000 words). It’s about…well I’m still revising it.

I started, too, a romance novel about a divorce lawyer who falls in love with a competing attorney (30,000 words). I wrote it in the throes of a whirlwind undergraduate romance. A number of friends think it’s decent enough to publish but I don’t have the desire to wrestle with the lawyer. (She is too in charge of the storyline.) I wrote a novella about a Gothic murder mystery set in the racist south (25,000 words). I have two collections of (a) horrendous poetry (16,000 words) and (b) misshapen short stories (80,000 words). Like, they are downright atrocious garbage. I even have two collections of Buddhist mantras that I put up on Amazon Kindle years ago. Lest you think I’m joking, you can still purchase them for $1 on Amazon (if you subscribe to Kindle, they’re free). The first is called Becoming and the second is called Incineration. I wrote them in 2009 after I visited an ashram in Puttaparthi, India. I still, sometimes, have mantras rattle around in my mind.

One thing I’ve always known is that I am not a gifted writer. I don’t have a knack for eloquence. Instead, I am a workhorse of a writer. A yeomen writer, prolific. My writing skill consists of a willingness to sit down at a computer, day-in, day-out, banging away at a keyboard. Sometimes, I will put off going to the bathroom to the point where it hurts. As an undergrad, I wrote a page a day. As an academic, I write 500 words a day, Monday-Friday.

Do you know how I know I’m not a good writer? Because I keep word counts like they’re a basketball score. That’s how I think about my writing. I don’t write with the goal of making good prose or thoughtful stories. I just chuck words counts like a baller throwing up 3-pointers. Shooters gonna shoot. Writers gonna write. Look at how I added those word counts. I knew how to write them innocuously, sliding in those parenthetical tags. Until recently, my talent for rote production of words was a type of talent.

AI has taken that talent away.

The devaluation of my talent is what makes me so jealous (or is it anger? worry? I remain unsure) of AI writing. I could always rely on pounding out word counts. But now AI writing tools make my talent irrelevant.

The snatching of my talent instigated (or is the word prompted? What is the word I’m searching for?) some introspection. I have started to reflect on the craft of writing more than ever. And one element of writing that seems to be left out of public discussion on social media and in the educational system is this: relentlessness.

Why relentlessness? Because writing is hard.

I could have chosen “obsession” rather than relentless but it doesn’t capture what I want. Obsession isn’t a careful enough word. I don’t treat writing as a job. I don’t treat it as a vocation or even a “calling.” A calling implies some sort of transcendental vision. A “calling” implies that a person was always meant to write. It’s teleological. Or telos for short. That’s a fancy five-dollar word for saying something has some innate purpose. Writing has never been an innate purpose for me. There is nothing teleological to me about being a writer.

The more accurate version of why I chose to write is that I have the ambition to write. I pour myself over my poor prose until writing oozes from my pores. Terrible word play. I go back and forth over my writing relentlessly. Every word is a choice. Pass after pass. Every phrase is another set of choices. Sentences are sets of massive choices. Paragraphs, ideas, punctuation. Repeatedly. All the time. Over and over. Sets. Repetitions. Like going to the gym. Repetitions. Sets. Over and over.

Go back a few paragraphs. Did you see what I was trying to do with these sentences:

• “The devaluation of my talent is what makes me so jealous (or is it anger? worry? I remain unsure) of AI writing”

• “The snatching of my talent instigated (or is the word prompted? What is the word I’m searching for?) some introspection.”

Look at the possible word choice there. Those are just two words that I have to choose from, as a writer. No one owns words. We don’t own writing. But we do make choices.

That is precisely what makes writing so difficult: you make hundreds, thousands of choices when writing just a few paragraphs. That’s why writing is exhausting. It depletes your cognitive power. I like to say, after an intense writing session, I’m all out of juice.

Let me go back to the opening sentence I wrote.

• “I have wanted to be a writer since I read the Boxcar children between 2^nd and 3^rd grade (summer of 1992).”

This was not the first sentence I originally wrote. Below is the first version.

Version 1: “I’m not a good writer. I’m serious. That’s not a humblebrag.”

That’s not even entirely accurate. There were about a DOZEN different sentences here. (There is another choice! I capitalized DOZEN. I could have put it in bold but bold doesn’t have quite the effect I want.) I came back the next day to rewrite Version 1.

Then came the next version.

Version 2: “I have always wanted to be a writer, a novelist specifically.”

Then came:

Version 3: “I’m not a good writer but I have always wanted to be a writer, a novelist specifically.”

I asked my spouse and mother what they thought of the sentence. They liked Version 3 the most. Then I completely rewrote the damn thing. Now, if Sam Altman (OpenAI CEO) or Dario Amodei (Anthropic CEO) or an AI evangelist saw me do that, they’d think I was a bad worker or bad writer. I’d suffer accusations of inefficiency.

That criticism implies a teleological purpose.

But I had to get through those other versions to really understand what I want to say. I wanted to think. I had to think in order to learn about what I was trying to write. I pushed through all the brutal thinking, the painful uncomfortable frustrating hand-wringing mind-wandering thinking. I was using my writing to think in order learn about what I wanted to write. That’s what no one talks about: Good writing means you’re getting better at thinking. All of that requires putting in the reps.

Relentlessness is about putting in reps

People suddenly care about writing because AI, as the zeitgeist, can write. There are so many AI gurus who talk about either the potential it has for tutoring writers or just rote document production.

But that’s not the point of learning to write. To be a writer, you must put in the reps. Imagine a basketball machine (BamAI if you will) comes along, able to sink threes from around the court. Would we suddenly ask our 11^th graders or college seniors who play basketball to become BamAI engineers? Of course not. We’d ask them to continue training, shooting, dribbling, lifting, so on, and so on.

I don’t write my lousy novels because I expect them to be published. (I hope they will someday.) I write because it’s practice. It’s a way to flex a writing muscle that my academic writing doesn’t test. The same thing goes for this blog.

The same thing should be for learning to write. When my students sit down to write, they are training to be writers.

For good writers must write, typing scribbling scratching clawing smashing relentlessly.

Being a cashier is remarkably difficult: you end up standing in one place, not moving, straining your back. Somehow you stand in one spot but end up exhausted. Your feet hurt when you come home. Wrists hurt from handling items of varying size and weight. It’s never been clear to me why it’s tough.

Reading AI writing, too, is bizarrely, inexplicably exhausting. I say bizarrely because, on the surface, reading AI prose should not be so tiring. The syntax of AI writing is correct; the grammar is impeccable; the transitions are smooth; mechanically, AI writing is a better writer than me. I have a tendency to leave off “-ed” (or add “-s”) from verbs when I’m typing. On the surface, AI writing should be good. It would never use three semi-colons in row. That’s uncalled for.

Yet, each AI written (or, more accurately nowadays, AI derived) sentence takes far longer than it should to read. I find myself re-reading complex paragraphs because nothing actually happens in them. As if: AI writing is too well written without enough ideas to justify being that well written. All of that complex syntax, without ideas, places too much onus on the reader to suss out meaning.

(It’s a bit like eating aesthetically pleasing food around the holidays that is devoid of either nutritional sustenance or exquisite quality. Trader Joe’s sells a lot of appetizers that are like this. They look good but all kinda have the same filo dough-type generic flavor. Or store-bought Christmas cookies that have a lot of decorations but just taste off for some reason.)

When I read AI writing, it makes me feel a bit like I’m losing my mind. Sometimes, when I read posts on LinkedIn, I literally don’t understand what the sentences are saying because they only have like 10% of an idea in each sentence. Thankfully, it’s not just me.

The problem of average writing

I’ve become fascinated with finding comment threads around the internet on which people complain about the poor mental effects of reading all this synthetic text.

In this Hacker news post titled, “Why does AI slop feel so bad to read,” I found several people discussing why there was a bad feeling of reading AI prose. In response to the post’s question, one commenter wrote the following (I edited to make it easier to read) about why it feels so bad:

Because it falls into the 10,000 bowls of oatmeal problem, i.e., the perceptual similarity of generated content is homogeneous. The procedural generation community has been familiar with this failure mode since 2016. The problem is that LLM output suffers from this failure mode too. Differences in authorial voice, ideas, and personality all get collapsed down into the average across all perceptual dimensions. It’s not unlike the fighter cockpit problem—the average of everything ends up not being representative of anyone. It’s bland and meaningless and no one writes that way even though it’s the mean of all written text. We can get away from this a little, but prompting LLMs to respond as if they were a certain author or a blend of different authors, but this doesn’t create anything new, just a mimicry. Personality is an unsolved problem that’s synonymous with avoiding oatmeal in my mind.

Let me parse that fighter cockpit problem because it’s one of my favorite examples of problems with averages. I’ll explain briefly by quoting from this great mini-essay about it, which itself is an excerpt from Todd Rose’s The End of Average (2016), a book recommended to me by my colleague Paul Prior years ago. The military, in their attempt to reduce crashes from pilot error, attempted to redesign their cockpits.

In 1950, researchers at Wright Air Force Base in Ohio measured more than 4,000 pilots on 140 dimensions of size, including thumb length, crotch height, and the distance from a pilot’s eye to his ear, and then calculated the average for each of these dimensions. Everyone believed this improved calculation of the average pilot would lead to a better-fitting cockpit and reduce the number of crashes — or almost everyone. One newly hired 23-year-old scientist had doubts.

The scientist Lieutenant Gilbert S. Daniels went on measure all the pilots. And here is the astonishing thing:

Out of 4,063 pilots, not a single airman fit within the average range on all 10 dimensions. One pilot might have a longer-than-average arm length, but a shorter-than-average leg length. Another pilot might have a big chest but small hips. Even more astonishing, Daniels discovered that if you picked out just three of the ten dimensions of size — say, neck circumference, thigh circumference and wrist circumference — less than 3.5 per cent of pilots would be average sized on all three dimensions. Daniels’s findings were clear and incontrovertible. There was no such thing as an average pilot. If you’ve designed a cockpit to fit the average pilot, you’ve actually designed it to fit no one.

No one fits the average. It’s mythical, not in the sense of being fake; it’s mythical in the sense that it’s an amalgamation of many stories, many data points. The problem with reading AI writing is that AI writing really doesn’t sound like anyone you know, or any author you’ve read, or any actual human being. I’m fairly certain that even if we trained large language models (LLMs) to adopt a specific style, these models would run into the average pilot problem due to their architecture. The mechanics of style are there but there is nothing deeper.

Here is the main point I want to make in this essay. We, as human readers until 2022, generally understand mechanics to be a strong indicator of conceptual sharpness. If periods, commas, semi-colons (hopefully), and other punctuation are all in the “right” spots, there was a general sense of trust that the ideas had a level of “hard work” put into them. To use a machine learning term, mechanics largely functioned as ground truth for a reader’s ability to understand the writer. Mechanics largely told me, as a reader, that there was something there in the writing. Maybe I didn’t agree with the writing. Mechanics didn’t inoculate the writing against criticism. But it provided a baseline for understanding something was there.

Now, though, we have something that was simply not possible a few years ago: mechanically perfect prose with complex layers of imprecise ideas. Beautifully written cues of mechanical correctness can now hide malformed ideas. If we expect fully formed ideas when encountering mechanically correct prose, there is a subtle expectation that there are, indeed, fully formed ideas already present in the text. It’s like eating a beautiful mass produced store-bought cookie: pleasing aesthetics hide cheap ingredients. Reading all of this synthetic, AI-driven text could actually be bad for us, cognitively and physically.

Notice that I used the word “malformed” in the previous paragraph when describing AI writing. I did not use the words incorrect or wrong. I’m not in denial that these GenAI tools are doing something interesting. I’ve tried making that clear in my blog posts. I don’t hate GenAI tools; I am skeptical of their ability to be precise as well as their role in education. I don’t even hate AI more generally; in fact, I’m a strong advocate of machine learning techniques for research. A mass produced cookie is still a real cookie; it’s just not a very good cookie.

My use of “malformed” implies that these tools are making mechanically correct prose awash in weirdly imprecise ways that, as a writing teacher and writer, I think needs a hatchet. I mean this almost literally: I end up as a reader hacking my way through the words and sentences. As readers, we have to work too hard with AI writing. You end up hatcheting your way through the various buzzwords or how they fit together with logical connectors. And in doing so, as a reader, you end up overloading your mind. Let me take you through what I mean.

AI writing makes readers work too hard

In another thread, “Reading students’AI writing is triggering my derealization,” (written all the way back in December of 2024), the poster writes (again text cleaned up a little bit to make it easier to read, and edited to shorten it):

I’m a writing instructor. I’m on the point of giving up. I’ve been teaching for almost 20 years, and I’ve been prone to derealization for about a decade. It used to be a rare thing. It was manageable….But in the last year I’ve had to read and grade so many essays written by AI, and they just...short-circuit my brain. I get that creeping “this isn’t real” sensation and brain fog starts to set in. It feels like I’m in a nightmare. I think it’s something about the uncanny valley quality of a lot of generative AI writing. Derealization episodes (at least for me) can be triggered by something seeming both familiar and “wrong,” or something that seems unread/nonsensical but other people are treating it as normal. It sucks, and it’s impacting my mental health. Wading through these essays while fighting my brain is grinding me down, and making it harder to stay focused and grade the non-AI essays. A tiny part of me imagines venting all this to my students and asking for some compassion, but I don’t have any actual hope that would make a difference to them. Does anyone have a similar experience? Thoughts on the remote prospect of ever getting accommodations for a legit mental health issue like this, when it’s all-but-impossible to prove that a student is even using AI?

I share a lot of these experiences with the user, whose username is SwampYankeeMatriarch.

May I pause for a tangent? Humans really do have outstanding screennames. There is something truly witty about humans on the internet. Yes, it’s true there is a bot problem with plenty of usernames with a name followed by a string of numbers (“Bob42893”). But I love it when the screen name is crafted in ways that my own mind doesn’t think about. Screennames give you a connection to complete strangers in unfounded ways.

Unpause. I totally get brainfog when reading AI writing. I get that brainfog from human writing, too. Prior to GenAI, I had the fog when reading Lacanian psychoanalysis and a lot of Marxist writers. When you read something, like really read it to understand it, with sincere effort, it can you make feel downright awful when the prose doesn’t make sense. There is one commenter, named Own_Narwhal_3297, who crystalized all this for me when answering SwampYankeeMatriarch’s post. They write (tightened, trimmed again):

…It’s really taking a toll on my sharpness. I definitely think there are some unstudied, harmful psychological effects of reading AI generated writing. Especially when readily detectable. As instructors, we’re encountering AI more than the average person. We enjoy reading. We identify as readers. So in a way it could also be tearing apart our sense of self identity. It’s an awful feeling of dread that 1) this is the present and future and 2) humans are capable of deceit in social situations that were once sacred. The relationship between the teacher and student is not what it once was.

I get all of this. I do. But the interesting moment for me is the suggestion that “there are some unstudied, harmful psychological effects of reading AI generated writing.” I think we need to study this. (BTW, if you want to design a academic study around this, please email me. I’d be excited to do that.)

The commenter, Own_Narwhal_3297, pokes at something here: the suggestion that some kinds of reading can actually be bad for you.

Wait, hold on.

Did you actually think I wouldn’t comment about that username? Come on now. The user owns a narwhal? Perhaps it’s an aspirational username. A funny thing about me: I didn’t know a narwhal was a real creature until I was 37(!). I thought narwhals were like unicorns: mythical.

Back to reading. Reading holds a weird place in our culture. We tend to think of it as a respectful version of media consumption. Watching movies and tv? That’ll rot your brain. Listening to too much music? Hurt your ears. But read? That’s good for you, like vegetables. Even now, you may think to yourself, as you read this long blog post, reading is somehow better than watching videos. I certainly do.

Reading holds such a valuable place in our culture because it makes you work to get the information. Reading is hard, in other words. Screen consumption is easier. Reading AI writing, though, makes me work through the prose not to imbibe meaning but rather to question if any exists.

I am not close reading, in the English literature sense, or even carefully reading, but rather what I call firm reading.

Firm reading

Firm reading is a response to the possibility that reading AI-driven synthetic text might actually be bad for you. If reading has always been like eating vegetables, then firm reading opens up the possibility that reading synthetic AI-inflected text may, in fact, be like eating trans fats or ultra processed foods.

To give you a better sense of what I mean, let’s contrast firm reading with a type of reading perhaps more familiar: close reading. Close reading, in the literary sense, represents a deep engagement with the source material. When you close read, you are actively engaged with a passage, often down to the level of sentences or even words. We close reading Hamlet’s soliloquy or Shelley’s “Ozymandias.” Yet, the reason we can close read in the first place is that we assume there is worthwhile meaning already there. Close reading privileges something thoughtful lurking beneath the surface. A reader needs to slow down, to get “close,” to find it.

With firm reading, we instead ask, “is there any meaning at all in this text?” Firm reading wonders whether anything is there in the source material, conceptually or mechanically. Firm reading, cynically perhaps, is thus a disposition toward disbelief. If close reading asks, “What are the possible interpretations here?” then firm reading asks, “Is interpretation possible?”

I find myself firm reading by pushing back against the very structure of sentences, often rewriting them. AI writing layers on adjectives without connecting them in coherent ways. As I’ve written twice before, in The Mental Tyranny of AI Writing and The Curious Question of AI-written Lists: Or, LLMs are Genre Machines, AI writing loves itself those lists (X, Y, and Z) and pairs (X and Y). But now I find myself asking numerous questions about the ideas themselves. Let me use a passage from an article I recently re-read.

…an algorithm is a set of instructions for performing a task or solving a problem. A cooking recipe, for instance, is an example of an everyday algorithm. In the context of this article, algorithms as input/output functions process a text and yield a result. On the World Wide Web, algorithms sort, distribute, and organize websites, writing, and content. In doing so, algorithms evaluate, structure, and influence writing and other discursive information.

On the surface, nothing seems wrong. Yet, those final two sentences are a garbled mess. Algorithms are sorting/distributing/organizing websites/writing/content. Algorithms also evaluate/structure/influence. Everything is everything. Algorithms do anything. I know, however, the author didn’t use AI to write these sentences because I was the author. I wrote these sentences in 2016 (the article, Writing for Algorithmic Audiences, was published in 2017). With firm reading, it doesn’t matter if the writing is human, AI, or a fusion of the two.

What I am sure about is that there is no specificity to these claims. Re-read that paragraph again. Did I define what “process” means? The use of “structure” and “influence” are NOT the same thing; I need to tease that apart. What I am really doing here is taking a mental hatchet to my claims, perhaps unfairly. I am constantly re-reading my old writing, hands quivering to revise.

Bizarrely, then, AI writing inadvertently made me a more careful, precise writer by molding me into a firm reader. I read much more carefully now, much more slowly. I read deliberately in ways I didn’t just three years ago. Weirdly, perhaps to the chagrin of the trees, I read much more now on paper, away from screens. I have found that the weird, malformed sentences of AI writing are glaring to read on paper.

I parse usernames much more carefully now. I wonder about SwampYankeeMatriarch and Own_Narwhal_3297 as I crawl over their prose. I’m more curious about them.

I enjoy reading for mechanics now, paying attention to sentence structure. I haven’t enjoyed the craft of writing as much as I do now. When reading on paper, I find myself highlighting good prose itself, something I haven’t done since I was an undergraduate.

I have found that firm reading makes me engage in the daily grind of writing and of life. Did you noticed how deeply I quoted my source material? Did you catch my seemingly offhand reference toward Paul Prior? Those are intentional to make a rhetorical point about how AI writing seems to be changing my preferences as a writer. A colleague (not Paul) once mentioned that she never read long quoted passages in writing. For a long time, I felt the same way. I find intense quoting now, however, an indicator the writer is engaging with material deeply. I prefer that to a passing reference in a post-ChatGPT era.

With firm reading, I find myself taking far more handwritten notes. I have been filling up all sorts of little booklets. I previously highlighted texts digitally and typed out notes. I realize that I was rushing past the writing, rushing past the ideas. Handwritten notes are my firmament to reading, to thinking.

Notes used to be like those customers when I cashiered. All the customers ran together in my mind. They were always a feverish mess of memory. The same thing happened with my notes. Every quote, author, page number blurred together.

Now, my notes look quite different. I ask more questions, one of which: “Is this real?” And I’m not talking about AI writing exclusively anymore. I’m talking about all the books I’ve been reading, many published long before 2022. Books from the 1970s, from the 1800s. Journal articles from the 1980s. In my notes, I have phrases such as, “How though?!” or “But why!?” peppered throughout. I was never such a malcontent before.

For firm reading makes me dwell, ready to craft, hatchet ready.

A little academic story

I was only able to defend my dissertation once I internalized the way my advisors provided me with feedback. That is, I was only able to write a dissertation once I had enough interactions with Donna and Anne to understand the intense expectations of a PhD. I naively thought that Donna simply wanted intensely organized chapters, with scaffolding paragraphs at the beginning of each. I realized that this was an expectation of a dissertation as a genre and academic writing more generally. I can now think about Donna, imagine her reading a document, and make choices around organizing my writing. She is with me, in my own head, an imagined construct. With my other advisor, Anne wanted consistency when writing about specialized terms or phrases. She was intensely critical of any terms that seemed to “slide” around. These were signals of quality and expectations of the deep care that goes into a dissertation. I channel Anne to become critical of the terms I’m using.

It’s become a weird power. I carry around ways of writing for Donna and Anne. I have them in my head. This ability, channeling or whatever you want to call it, only develops after I’ve received feedback several times. I need feedback on the order of something like 7-12 times.

I have four imagined readers that take the form of my undergraduate professors: Harriet Pollack, Saundra Morris, Jeff Turner, Richard Fleming. Harriet tells me to slow my writing down. Saundra reminds me to be precise. Sometimes, she and Anne have arguments in my head. Professor Snape (aka Turner) tells me to stay true to the text. Fleming stares at me silently, sitting in a corner. I want to impress him.

I wanted to impress Donna and Anne. I wanted Harriet not only to read my essays about Willa Cather but to think about them. I wanted her to sit with my writing. I wanted Saundra to read my criticism of Emerson’s poems and smile at my prose. Snape…I wanted Snape to see how much virtue ethics had changed my life.

I’ve never wanted to impress ChatGPT.

AI writing isn’t an audience

AI chatbots can never be a true reader. Never. They’re trained on the entire internet but they can never replace a human reader. Human readers are idiosyncratic. We have limbic systems. We don’t just experience pain or hurt. We feel it. AI chatbots have math. Humans get hangry. We love pizza with mercurial toppings. We cry. We take joy in random things. I love woolly caterpillars. I save them if they’re stuck on a sidewalk. My children have drawn me pictures of woollies. They’ve drawn them for me. They know their audience.

Unimpressive rubrics

I recently read a few articles about students being upset at their teachers for using AI to grade or provide feedback. They should be. Writers need audiences. Writers yearn to be read. Student writing needs to have a reader but no one is going to read novice manuscripts unless it’s a teacher. Students need readers, not grades. Real readers, with limbic systems, not rubrics.

I’ve never wanted to impress a rubric.

“A little didactic”

My colleague, Bill Hart-Davidson or Bill HD as I like to think of him, told me my writing was didactic. Didactic means to instruct, usually in a bad way. My writing tried to control readers too much. Bill told me this over lunch in Pittsburgh in April of 2019. He had a veggie burger. I had a beet salad.

When Bill passed away suddenly in spring of 2024, I made pizza with his dough recipe. His dough recipe is very instructive but it called for being flexible about the amount of oil and milk you added to it. “Add to taste,” the recipe says. I try to write less instructively nowadays. Since 2019, in the limit sections of my academic papers, I try to add genuine suggestions for future research.

I wanted to impress Bill.

AI writing is too dense, too even

Recent research shows that AI writing is quite different from human writing. AI writing is too informationally dense. That’s why it’s so difficult to read. It’s like a sundae with everything on it. Just too much. Multiple concepts in a sentence, concepts connected in only vague ways. It loves listing things, cataloguing loads of nouns. The sentence structure of AI writing is decidedly unvaried, awful in fact. Human writing has short sentences. Long ones. They burst. They drone on. Good human writing excites readers. It makes us feel.

AI writing adds “-ing” to everything, or what the technical linguists call present participial clauses. These machines rarely mention people. They rarely use epistemic stance markers, which is a fancy term for the writer revealing their perspective on the subject matter. Humans add something of themselves to a text. As far as I know, I’ve never seen ChatGPT use epistemic stance markers.

It’s been three years since I lost my father

When my stepmother called to tell me my father died, an intense wave of pain rippled through me. I cried out from the visceral hurt. It felt like a heart attack. My older child still talks about hearing me call out. Remembers it. Recalls it. Feels it in a limbic system.

A month after my father died, I was granted indefinite tenure. If you know anything about academia, this is a career milestone. I felt nothing. Indifference. All I wanted to do was call my dad. I listened to his last voicemail dozens of times. (It was about the NBA and whether Joel Embiid could stay healthy.)

My dad was terrible at typing. He hunted and pecked. His emails obeyed no standard rules of discretion. No formatting to the emails. Maybe half sentences. He didn’t craft his email. Just made his point and moved on.

When I was in seventh grade, I typed out my first school paper. I had written out the entire draft by hand. The Sunday before it was due, I started typing but was so slow that I asked to have lunch in front of the computer. My father said no.

My dad was impressed with my typing.

We ate lunch together, then he sat down to help me write. He was slower at typing than I was!

I imagine my dad hunting and pecking as I write. Other occasions, I imagine him typing wildly fast, nothing at all how he actually typed. A hallucination of my own.

On the limits of audience

There are limits to audience awareness. There are limits to AI writing. There are limits to academic writing. We don’t always write for transactional purposes. Sometimes, you just gotta get something out. Expurgate it. Ignore your audience. Get it out of your limbic system. You write with tears streaming down your face because you carry around audiences, readers who never leave you, ghosts who help you write.

Subscribe now

AI writing is deskilling students. Have conversations about deskilling

We need to have long, difficult conversations that AI might be deskilling an entire generation of students. These conversations need to happen with students and parents. The parents need to be involved. Teacher-led initiatives always need a large coalition.

Let me explain the deskilling problem as briefly as I can. My entire life, older generations always had one major weakness in terms of job competition: a lack of comfort with new technologies. That’s simply not true anymore, mostly because the technologies are so easy-to-use. When I was a kid, I often had to fiddle in a terminal to get my videogames working. Or learn HTML to make a website. None of that is true anymore, hasn’t been for years. AI chatbots, such as ChatGPT, are radically easy to use. There isn’t some secret to using them, even though corporations (and unscrupulous online gurus) act like there is.

GenZ now faces intense job competition from older generations who are no longer technophobes. In fact, GenX remains a generation of top programmers (they know the programming language C and *cough cough* maybe Fortran). Python is easy to learn if you know C. Not the other way around.

Writing with AI is the same way. If you already know how to write, using ChatGPT is easy to use. If you don’t know how to write on your own, then people who do will always outperform you with respect to AI writing tools. Always. As a result, when a person doesn’t learn to write on their own, they’re deskilling themselves from the job market. They won’t be able to outperform those who have the habit of perseverance learned from writing. Consider, for example, summarizing. The point of summarizing, ideally, is not to quiz students but to make them glean and subsequently distill important ideas from a reading. Written summaries aren’t meant to be valuable themselves. Rather, summarizing is meant to indicate having thought through some ideas, putting them through your mind’s meat grinder. It’s an exercise designed to make you think. In fact, summarizing should never be graded. That defeats the entire activity of summarizing.

Subscribe now

Move writing assignments in-person, sometimes with computers, sometime with blue books.

One major response to AI writing is to move the bulk of writing back into the classroom. Have students write in class where we can watch them write, including how they write, including the apps they use (with AI or without), offering them real-time feedback. Real-time feedback requires an intensity from the teacher because you need to be “on” the entire time. The bright side here: that real-time feedback shows students you’re engaged.

I think there is a general dislike of moving writing to the classroom that stems from our valorization of the author composing in isolation. It hearkens back to that weird “writer in the garret” image. Many people mistakenly believe good writing only happens through some sort of mysterious muse or when a writer is inspired.

I disagree. Treat writing like any other skill: practice it in challenging circumstances. Treat it like free throws in basketball. Try different techniques. Think about your form, i.e., posture. Think about your writing’s form, i.e., your genre, sentence technique, deliberate word choice, and many other dimensions. Have students set a time slot where all they do is write (say 15-20 minutes). There is an element here that teachers need to practice writing as well, if only to practice what we preach. I write 2 hours Mondays-Fridays. I treat it like work because it is my work. And I tell my students about my own frustrations with writing. I often recount days when I wrote a negative number of words (revision days).

I think the “blue book mindset” receives the bulk of the criticism here. This mindset usually isn’t writing as much as it’s taking a test via writing. There are other ways to do blue book writing that is more helpful. Blue book writing, as a journal, allows students to keep track of brainstorming. It’s useful for record-keeping around freewriting or quickwrites (never grade these, just check them to make sure students are doing the writing). Part of my own knee jerk reaction is that these kinds of in-class blue book activities are juvenile, reserved for grade school. Perhaps. But perhaps, grade schoolers are onto something, with their love of story telling built around low-stakes, ungraded writing. The entire point for them is to practice writing.

Subscribe now

Teachers could start using the oral exam more extensively.

Oral exams have become a standard suggestion for circumventing AI use. This strategy has serious potential because there are strong relational elements between writing and speaking. I wrote an entire book about this, one I’ve made publicly available. I see this response as taking a forking direction. First, the oral exam could literally just be an exam where a teacher asks the student questions. We do this for PhD exams. The other direction involves having students write essays, hand them in, and then teachers ask them questions about their writing. If students can’t defend (or even recall) their own claims from the essay, this type of exam will reveal those problems.

There are a couple of drawbacks here. The obvious one: oral exams are time-consuming. That’s historically why we turned to writing papers. I don’t foresee this working for classes with more than 15-20 students. But if teachers are both supposed to read a paper and then meet with students, that’s an extraordinary amount of work.

The second drawback is one I haven’t seen discussed much: Teachers aren’t trained to ask students exam-type questions about writing in real-time. I thus advocate for training teachers in asking challenging, probing interview questions about writing. A new skill of 21^st century teachers is likely being a good interviewer. These interviews should take the high-stakes form of talking, like the job interview. Not like podcasting because these have audiences and are too meandering. A corollary here might end up reuniting writing and oral education components.

Show students we actually read the papers. And design assignments with that in mind

We need to show students we are reading their papers. True, this indicates we care. But it also means that we take our jobs seriously by offering legitimate suggestions for improvement. This is the one I’ve adopted the most. I make it clear to students that I read their papers by offering a substantive mid-term letter that I wrote without ChatGPT (copy editing mistakes included, artisanal writing, if you will). It takes a ton of time. I take these letters, print them out on letterhead, and sign my name by hand.

Use AI tools as modeling audiences, situations, and genres

AI tools have been trained on the entire internet and, possibly, millions of books. They simulate cultures, people, situations, and genres. It’s all incredibly biased, but so is reality. I think “Big Tech” misses the true power of their technological terrors (Darth Vader voice) by framing the chatbots with branding around time saving and “write for me” slogan. If they really believed in their products, they’d be trying much harder to design a thinking machine that responds to writing, not write for people. There is untapped potential here. For me, it’s the most promising, technically and socially.