Note of disclosure: I formed this essay over the course of two weeks from a nagging question in my head: why do I loathe AI writing so much? Several articles showed up in my RSS feeds and email newsletters that nudged the question into thoughts and then eventually into ideas I had to explore. Dan Koe’s article I’m begging you to write more essays and Paul Graham’s The Age of the Essay were the signs from the universe to give long form writing a try.
But the real nudge that kept me moving through the process was Claude’s Opus 4.6. While I am aware of the irony, I haven’t fully concluded if this essay is a satisfying exploration, a useful experiment, or an exercise in hypocrisy. I am certain that it’s the future. The removal of friction, the endless organization and revision suggestions, and even the somewhat sycophantic encouragement, was…needed. I don’t know that I would be able to complete the ritualistic sacrifices needed for Writing, the Process™, given my current work and family demands, were I alive at any other point in human history and technological advancement. It’s a bit of a mindfuck. I’m probably not alone.
The writing starts off well enough. A decent premise, a hook that gets you mentally committed to the five minute reading time promised at the top. Then something starts to poke at the edge of your mind like a smoke detector light blinking in a darkened room. An unnecessary em-dash. The word delve used in a more than acceptably pretentious manner.
Then, finally, a “it’s not X, it’s Y” grates on you like the second runner-up in a JFK speechwriter contest. You stop reading and scroll to the comments hoping someone validates what you’re suspecting.
AI writing.
You don’t really know it’s AI writing. It’s a guess mixed with a nagging feeling you can’t pinpoint. You can’t confirm it unless others confirm it, or you forcibly extract the author’s confession.
It’s weirdly ironic — that suspicion of accumulated weighted averaging of human writing can’t be judged until a requisite weighting of humans achieves a certain average judgement. But once the gavel falls and the scarlet letter is cast, the response is almost universal loathing.
How dare you. Why couldn’t you just take the time and write it yourself?!
Why do we hate AI writing so much, though? Let’s explore that question, because the obvious answers don’t quite hold up.
To preface, most people haven’t written of their own free accord since — well, a teenage diary. Most of us relished the last words of the last paper in our last semester of college or grad school, and promised ourselves: never again.
And so, very few of us spend leisure time formulating thoughts into sentences on a keyboard or paper; especially not when a day’s work has melted our brains, TikTok only requires a single finger, and a new season of Bridgerton needs binging.
We just don’t write.
Among those who do, statistics say half of us write at a quality that can be defined as average. In the US, this is about a 7th or 8th grade level.
So, why hate on a tool trained on the finest writing of humans past and present and perpetually available? Shouldn’t we be greeting AI writing like a wisened guide shepherding us up literary peaks we’d never have reached alone? The quality is better than what most writers would have produced alone. But that’s the polar opposite of the reaction it gets today. The revulsion can’t be about the words being bad.
Maybe it’s the deception. I do put some stock in this idea. From personal experience, I have a much better time with articles that disclose AI use, preferably in the byline (but I’ll accept a clear footnote). I think it frames the intent of what I’m reading, and lets me adjust my mental faculties accordingly — I know there was a human collaborating and experimenting and so I’m more forgiving of that partnership before I em dash the writing on the rocks. And if it’s pure AI writing? I appreciate immediately closing the tab like a work meeting ended early.
Yet, even disclosure doesn’t resolve the mentally prickling feeling. There are pieces where I “get the ick” even after the author has told me upfront. Disclosure helps with the trust violation — it renegotiates the contract between writer and reader — but it doesn’t fix whatever this deeper thing is.
I’m not the only one to notice this gap. In Taylor Pearson’s As We May Work, he describes a stage of his writing workflow he calls the “De-AI-ify pass”:
This is basically a filter for AI voice. It strips out “delve into,” “it’s worth noting,” “landscape,” unnecessary hedging, false enthusiasm, hollow transitions, etc. Some of this residue seemed to creep in during the structural edit.
Pearson isn’t alone. I’ve seen other writers describe the same step, sometimes insisting that anyone using AI should do it. I do it myself when copy-editing my own drafts. The fact that there’s a named workflow stage devoted to scraping a generationally revolutionary assistant’s fingerprints off your own writing tells you something. An anomaly…a non-human one…is being detected. Something specific enough that experienced writers have built a defensive routine around it.
So: a mix of trust violation and uncanny valley, maybe? Fine. Useful even. For future writing obey the following commandment: Don’t lie, don’t delve, do disclose. You’ll be fine.
But even this is a dodge, because none of it explains the recoil. None of it tells me what I’m actually detecting when I get the ick. And until I can name that, the recommendations are just hygiene rules and risk mitigation. They treat a symptom.
What is it I’m detecting?
A little explanation by story: five months ago I was writing with Claude’s Opus model and was floored by the jump in quality. Where I had previously asked for structural outlines because rewriting AI prose into my voice took longer than just composing from scratch, I could now ask for a fleshed draft full of quirky aphorisms and lexical banter and mold it with confidence. The de-AI-ify pass got harder, not because the AI was writing “better”, but because it was writing more “humanly”: there was less to strip out. The “delve” trail was fading. The em-dash was mostly gone. Pieces had tangents and meandered like a well-intentioned, idea-exploring writer. Reading some of the conversations I felt a mix of awe and a rekindling of hope as the world of AI writing bridges idea and execution in more powerful and more accessible ways.
Then, I felt a second set of emotions: fear and defensiveness.
Removed from college by nearly 20 years, I’d finally healed from the scars of forced writing enough to want to try writing again. I even prepped and planned. I’d gone off the zeitgeist grid for years, dodged every social media and streaming distraction and cautiously mustered a modicum of motivation to attempt writing for myself. I even purchased an online writing course for beginners. There’s an email receipt collecting dust in my inbox somewhere.
But if AI writing was this good, why should I try to coax my thoughts out of the shadowy cobwebs to develop my own nascent voice? Why would anyone, even I, want to hear an early-high-school-level struggle with words when I could be spoon fed a fully formed listicle with just the right amount of pithy sarcasm and quip? Quickly. Cheaply. Proficiently.
The odds continue to stack against the human. The bridge curves steeply upward to form a wall of ‘why even bother.’
I expect we’ll see more inflection points as time goes on and AI capital investments keep flowing like someone hosting a Guild Navigator all-you-can-eat buffet.
Soon, will we still be able to tell the difference between human and AI? Will it even matter if there are so few human voices left, drowned by the firehose of AI writing?
I believe….and hope…yes. That prickly feeling won’t magically disappear. But spotting and identifying AI writing has to be the result of something more durable than the occasional “delve”. Those tells are eroding fast. The reason has to be structural. And I think it is the residue of human reasoning.
Reasoning is a fingerprint of unique thought. Our patterns of formulating and chaining thoughts and then transforming them into communication are incredibly complex and unique to each person. It’s a mixture of our life experiences, our biochemistry, our genetics, our cultural upbringing. Even the way we scaffold thinking onto language is shaped by what language we first used to think with — case in point, have you ever heard a new English speaker reach for an English metaphor and miss by an inch? Or watched a multilingual friend try to render a non-English metaphor into English and end up with something more interesting than either original? That gap is the fingerprint becoming visible.
AI, at its current architectural core, is a probabilistic determination across the weighted collective of all those fingerprints, with every culture dumped in like someone emptying the entire spice rack at once into a pot of soup. It tastes off. The output is, by construction, the centroid: the average essay no person actually wrote.
Here’s a more relatable analog: grocery shopping. In aggregate, consumer grocery purchases can be modeled and predicted. You can model a town’s weekly milk consumption. You can trade futures contracts on the underlying commodities all the way up to a national scale, and people make a lot of money doing it. But, scaling down does not work the same. No two grocers shop alike, even to their own selves, on a given trip. The same shopper, on two different grocery store excursions, won’t have the same items, the same bill, or even the same consumption of the purchased goods.
The point? The aggregate is real and useful. It gives us a meaningful and representative picture that can be used to accomplish impactful work such as forecasting prices, inflation, land usage, trade surpluses and deficits, and more. But it cannot predict or serve as a proxy for the individual.
You cannot infer the shopper from the futures market.
I think humans, as a species, are accustomed to encountering each other one on one. We have spent millions of years tuning ourselves to that level of granularity and have come to expect a level of deviation from the ‘norm.’ Picking out what makes us unique is what gives us that proof of humanity. We recoil when a single person behaves the way a crowd would. There’s a term for the feeling depending on the context. Cultishness. Possession. Mass hysteria. The feeling that there’s no one home behind the eyes. We have very old detectors for this, and they fire whether or not we can articulate why.
This is what’s actually happening when we read AI writing and feel the ick. AI is the crowd. A piece of AI writing, served up as if from a single author, is a grocer shopping based on commodities futures. The voice of an aggregate posing as the voice of an individual. Writing with no soul behind the eyes. Our category-violation alarm goes off, and we feel it before we can name it. The em-dashes and the delves are evidence we use to confirm the suspicion, but they aren’t the cause. The cause is that we sensed the crowd.
This explains why disclosure helps but doesn’t fully resolve. Disclosure tells us yes, what you’re sensing is correct, this is the crowd speaking, which lets the alarm stop firing as a false positive and start firing as a true one. But it doesn’t make the experience pleasant. A crowd is still a crowd. We just stop pretending we don’t notice.
So what changes as the models get better? Two things, on different timelines.
Short term, lazy option: we get used to it. Some of the alarm dulls just from repeated exposure, the way you stop smelling your own house. This makes me sad and I think we should resist it.
Short term to medium term with a bit more work: we enact our own boundaries. Not by banning AI writing or shaming the people who use it badly — that ship has sailed and the shaming was never going to work anyway. But by writing more, ourselves, in our own voices, with our own meandering structure intact and utilizing AI as the helpful shepherd who has seen the peaks and valleys of writing, and understands the journey. This article is my attempt at that: it is written with the help of Claude Opus 4.6. Working with the model was an astonishing experience as it pushed the evolution of this essay from its birth as chicken scratch shower thoughts dictated to Notion to its final form here at nearly 1500 words where I can hear my own voice. It took me two weeks of effort put in at the end of long work days with nothing left in my mental tank, and I loved every painful moment of it. I feel like I’ve been gently challenged to battle myself and win: I feel bloodied but emergent.
Long term — and this is the part that should keep you up at night — humanity finds a way for technology to recreate the fingerprint. To spool up “unique voices” with synthetic formative experiences, mental anchors, lexical quirks, and the kind of irreproducible coherence that emerges from a specific life. Sci-fi has been exploring this for decades and I don’t think it’s as far off as it feels. Once a synthetic fingerprint is convincing, the question becomes is this human, and is it the human it claims to be? The answer to that question is going to look less like a stylistic check and more like a chain of custody with notarization. The signal we currently get for free from “this prose has the texture of a lived life” is going to get expensive, and the people who can produce it verifiably — through long bodies of work, through visible revision histories, through accumulated public thinking that no one could have generated retroactively — will hold something the rest of the market cannot fake. To me, it’s a storm brewing far off on the horizon, but racing towards us on a hurricane wind.
That’s a strange future. It’s one where the most valuable thing you produce is not what you wrote but the trail of having written it. Where the body of work matters more than any individual piece. Where short form becomes increasingly suspect if it has no chain attached, and short form embedded in a long body of legible thinking becomes more trusted than ever, because that message from that visible chain can only have impact if it connects on the human level. A meme goes viral because it infects at the level of our DNA.
For now, the practical advice is simple. Write in your own voice. Include the parts you’d be tempted to scrape out as too weird or too specific. Write long enough that the chain is visible. Don’t optimize the variance out of yourself trying to appeal to the masses — the mass appeal is the thing that’s about to be free, and what’s free is what’s about to be worthless. In a world where knowledge is a democratized commodity, the costly signal of identifying, showcasing, and proving out your unique fingerprint is all that’s left.
Don’t trade it for delves.