Blame the machine, as always

Professionals who live inside language (editors, critics, teachers, established writers) know the case against AI by heart. The list is long, it is damning, and each charge, taken on its own, sounds convincing.

The machine averages. It sands the particular grain out of a writer's voice. It writes "not X, but Y" where a person would simply say "Y," and it constantly reaches for an em dash, a mark most living people couldn't find on their own keyboards. It sorts thoughts into threes, like pricing tiers on a subscription page. Where an image should be, it bolts on bridges and fresh chapters. It floods editorial inboxes with manuscripts. It teaches readers to eye every overly smooth paragraph with suspicion. It leaves behind prose that is lifeless and soulless, beyond resuscitation, fit only to be rewritten by hand from scratch. It even drives people to write worse, so that no one will mistake them for it. And it wears down their habits of thinking.

None of this comes from spite. Behind every accusation stand decades of exacting, serious work with words. The complaints are fair, for professionals. They have something to defend.

We will keep talking about the failings of the models. Today, though, let's set them aside for a moment and look at the main problem in your own writing.

That problem is you.

cr.png

Worse than what?

Notice who the professionals always defend.

When they need an example of a good voice, they reflexively turn to Hemingway or Maugham. There is nothing strange in that. These writers are the big leagues, the best of the best, the yardstick against which all writing is measured — AI included.

The professionals say: "The machine ruins your text." The trouble is that "ruin" is a comparative word. To ruin something, you need it to have been better before.

Now, honestly: does your writing resemble Hemingway's? It's flattering, of course, to drape yourself in all the indignation the professionals have been saving for machines that desecrate a beautiful piece of writing, with all its nuances and subtleties. But if you aren't Hemingway, I wouldn't lose sleep.

If you are Hemingway, I wouldn't either — though for a different reason.

cr.png

Two boxcars of exceptions

"A writer is someone for whom writing is more difficult than it is for other people," observes the narrator in Thomas Mann's Tristan, watching a poet agonize over a letter.

Writing is generally harder than speaking. Most of us learn to speak on our own by the time we're four years old. Writing has to be drilled into us, slowly and painfully, with textbooks, dictation and red pens, and what remains at the end is a train of rules trailed by two boxcars of exceptions that we must simply memorize.

Somebody could launch rockets, invent new drugs for obesity, play the stock market, even be a loving husband and a responsible father, and still not know what anaphora is. He may write with complete illiteracy: monstrous errors, no capitals, no commas. He faces no consequences. His professionalism was built around something other than text.

Most of the people around us can read, but writing two coherent paragraphs, in which a thought travels from beginning to end without scattering along the way, is beyond a good many of them. In 2004, the National Commission on Writing surveyed HR directors at major American corporations, who estimated that roughly a third of their employees wrote worse than their positions demanded. (Their roles usually require reports, memos and other business correspondence.) You can run a fresh survey yourself in a minute: open any work chat and watch how people write there.

For most ordinary people, writing is not a native element or an art form. It is a chore they long to shake off. Remember that every time you feel like mourning someone's lost authorial voice.

cr.png

A gift from the gods

The professional looks at machine text and sees a flawed knockoff: a cliché here, a slickness there, not one living thought on the whole page. Perhaps he is right, but a person whose commas and paragraphs have always landed wherever they happened to fall sees something else: clear, readable prose they could never have produced themselves. To them, it is a superpower. A miracle. A gift from the gods that lifts them a whole level at once.

Averaging works simply. It lops off the jutting peak and hoists the sagging bottom. For an editor with thirty years behind him, that can do damage, because he wrote better than average. For someone who wrote worse than average, it is a windfall.

The research tells the same story, though none of it concerns prose, or people who can barely string two words together. In the experiment by Noy and Zhang, 453 college-educated professionals wrote press releases and short reports. Those given ChatGPT finished 40 percent faster, and professional evaluators rated their texts 18 percent higher; the gap between strong and weak writers narrowed. In a call center with more than 5,000 agents, an AI assistant lifted productivity by 15 percent on average. Novices produced a third more, while the most experienced staff gained almost nothing.

For a strong writer, the machine often adds next to nothing. For a weak one, it brings clarity and simplicity, and that's all many ever needed.

cr.png

Writing without a voice

Say a flight is canceled at eleven at night. A traveler spends the hours until morning in the terminal, then three more at the airline counter, and now he is holding a hotel booking, a taxi receipt and a right to compensation that the airline is in no hurry to mention. He neither knows how to write a complaint nor wants to learn.

He turns to a model. It asks for the flight number, has him photograph the receipts, checks the industry rules, lists the violated provisions, names the sum and the deadline. The tone is courteous and firm. Two weeks later, the money arrives.

Did that letter have a voice? No. And it didn't need one. The employee the airline pays to process your complaints doesn't care how you write. What matters to him is finding the flight number, seeing the receipts, and grasping that it is easier to settle with you now than to invite a lawsuit. A letter in which the author "sounds like a living person" loses here to a letter in which they sound like a lawyer.

There are far more genres like this than people used to writing for print tend to realize. A reply to a client, an application to a school, an invoice, a summary to go over on Monday: each does a single job, and the author is less offstage than simply superfluous. Those whose voice is their profession take this hard. Everyone else sees no problem at all. All of the above documents were always written from templates, and everyone was fine with that.

A voice matters where the reader is choosing you from among others. Knowing where it is wanted, and where it is better kept to yourself, is the skill. The machine hasn't abolished that skill. It has only shown how narrow it has always been.

cr.png

It's hard to talk a thinker out of thinking

The gravest charge runs like this: the machine trains writers out of thinking. Every new marvel of science and technology, we are told, is out to make humanity unlearn something useful and good. Television taught us not to reflect. The calculator taught us not to compute. The search engine taught us not to remember. Each time, the implication was that the tool strips a person of the right to stir their own brain.

In 1946, in his essay "Politics and the English Language," Orwell wrote about ready-made phrases that assemble themselves into sentences and will "even think your thoughts for you, to a certain extent." He was not speaking of machines. He was speaking of verbal clichés. What dulls the mind is not the tool but the habit of reaching for the prefabricated, and people had that habit long before the models.

Think of the encyclopedia. You opened the paper encyclopedia a couple of times a year: you had to gather your strength, blow off the dust, and count along the spines to reach the right volume. The CD made it easier, but you still had to fish the disc out of its case and feed it to the drive. The internet made it easier and faster still, and whoever was curious began opening the encyclopedia more often, while whoever wasn't, didn't. AI models made research faster and more convenient again. The curious received a fantastic conversation partner who explains, tracks down sources and builds a chronology. The incurious received a way to dodge tedious work.

I remember school. When I didn't feel like doing an assignment, I copied it. When I was interested, I did it myself. Pen, paper and friends technically helped me either way — but the trouble was never with them.

Large language models have been unlucky in one particular way. They trespassed on our language, and language is sacred, so for doing the dirtiest and most thankless work they have received no joy and no gratitude, only an endless torrent of abuse and curses. Other inventions get applause. An airplane carries you across the ocean without your so much as flapping your arms: a miracle, plain and simple. A washing machine launders your socks: everyone is perfectly happy, and nobody weeps that it does so soullessly and identically, rather than in your inimitable style. But let a machine rewrite a clumsy text so that it can finally be understood, and out come the words "humanitarian catastrophe" and "loss of a unique authorial voice."

Everything, of course, has a price. The calculator wears away your long division, the car your marathons, and a menu translated from a photograph in a little Thai restaurant, your chance to learn Thai. In exchange, life is a little easier. And the professional Thai translator needn't weep: he didn't lose your menu, because you would never have hired a translator for it anyway.

cr.png

The baseline

"Better or worse?"

When machine prose is compared with human prose, the verdict depends on five different things.

  1. Who is judging: experts and ordinary readers often vote differently.
  2. What is being measured: clarity and rhythm give one answer, originality another.
  3. Who you choose as the "human" benchmark: the average writer or the acknowledged master.
  4. Whether readers know the text was written by a machine (once they learn, they usually score it lower).
  5. How well trained the model itself is.

The last is plain to see in the work of Tuhin Chakrabarty and his colleagues. Readers were given excerpts in the manner of well-known authors: some written by graduates of creative writing programs, some composed by a model. When the model was merely asked to write "in the style of" a particular author, experts rated the humans' work higher, but when the model was fine-tuned on all of that author's books, experts and ordinary readers alike preferred its excerpts.

Our own small journalistic test pointed the same way. We asked five models to write a paragraph "in the style of" Hemingway and three other authors. All of them reproduced the set dressing and the vocabulary, and none could capture the style itself, as described in "Hemingway's Writing Style, as Faked by Five AI Models."

It all comes down to one thing: the baseline matters.

There are three reference points, and they differ from person to person. For the professional, zero is the finest models of the craft, and everything beneath them counts as failure. For most people, zero sits far lower: it is their own text before the model, the business letter they would have labored over for an evening. For a few, there is nothing to push off from at all. Without outside help, they could not have produced a single paragraph.

Now look at these three versions of "worse."

Which one is yours?

cr.png

Text that simply works

An author is writing to an editor he has worked with for years. The piece will arrive a week late, and he needs to say that in a warm, conciliatory way. The letter opens with "Hope this finds you well" and closes with "Best regards." The formulas are polished to a shine: millions wrote them before this author, and millions will write them after.

Between those lines sits a living person. Stuck on the third paragraph. Three days of staring out the window. It'll all be done by Thursday.

The clichés are no enemies here. They hold up the frame and signal that the author is courteous and observes the forms. The author's singular voice lives inside, and the frame doesn't get in its way.

Here is another case. A friend of mine was selling a trinket box on eBay, and a buyer wrote to him in Kazakh. My friend doesn't know Kazakh. The model does. He got a translation — the buyer wants the box and asks about shipping — asked the model to answer, and the reply came quickly. It was probably polite and smooth.

Was it good? He couldn't tell. He couldn't see either style or errors. The only way to check the results would have been from the outside: ask the model for a back-translation, show the text to an acquaintance who knows the language, or wait to hear what the buyer says.

The buyer wrote: "I'll take it." The letter worked.

Perhaps the text was bad: averaged, even a little soulless. It still did exactly what was asked of it.

cr.png

Fighting problems that no longer exist

Most of our everyday complaints about how the machine ruins text belong to the past. Algorithms learn and improve faster than we notice.

In 2023, the magazine Clarkesworld temporarily shut its doors to submissions, swamped by AI-generated stories. The editor said they were easy to recognize because they were bad.

Then the models began to write better, and "bad" stopped being the chief giveaway. People went looking for other tells: the em dash, the "not X, but Y" construction, the whole list you already know. How this hunt works, and why the quarry keeps slipping away, is laid out in detail in "A Brief History of Slop." What matters here is that these tells are not eternal. "Not X, but Y" still holds for now — nobody quite knows why — while the machines have already dealt with the em dash.

By 2026, the nature of the complaints had changed. What is now at fault is no longer the text but its smoothness: flawlessness looks suspicious. Too good to be human.

"The only true counterstyle is good writing."

— Jessye McGarry, quoted in Emma Madden, "More Typos, Fewer Em Dashes: Writers Are Creating an Anti-AI 'Literary Counterculture'," WIRED, July 2026

The models themselves are growing fast and unevenly. On the whole, though, they seem to write better than they did a year ago. Across different texts and tasks, results can drift: one thing improves a little from version to version, another for some reason slips.

Check AI writing more than once. If it came out badly today, that is not a verdict. If it came out well, that is not forever either.

cr.png

How to check yourself

You and the machine have written a text, and it lies before you. Is it better?

Ask yourself a few different questions.

Was it easier for you to begin? If you used to sit for ages over a blank page and now you have finally gotten moving, that is already something.

Can you explain in your own words what this text is about?

Now look at the text itself. Is it clearer and simpler? Are the thoughts laid out better than before? Did you find words you had been missing? Are there fewer mistakes? More commas where they belong?

Is more of it actually checked? The machine helps with more than stringing words into sentences. With it, it is easier to find material and pin down a quotation — what used to swallow an evening now takes minutes.

Is it easier for the reader? Does the reader understand you?

One last thing: If you are writing a quarterly report, make sure no trace of your authorial style and voice has survived.

cr.png

Don't worry about it

So what is the verdict? Did your text get better or worse after AI?

Don't feel bad if it seems the model makes your writing better.

If your own voice matters to you (and not everyone needs one, nor does anyone need it everywhere), you know perfectly well where to find it and how to hone it. Read a lot and write a lot. Hoard words, so you don't have to hunt down the right one. Notice how different writers say the same thing differently. Hear where a text pauses and where it is time to stop. Read aloud, and catch where you stumble. Learn to write a sentence so that the reader feels dryness, a body, irritation. Know the rules so you can break them on purpose. Keep striking out whatever is superfluous.

And certainly don't feel bad if it seems the text got worse because the model has once again hopelessly wrecked your individual style and voice with its soulless averaging. Just try again in a month, or in a couple of weeks. Models learn fast — and far better than you do.

Sources

References cited in this piece. Last verified on the published or revision date.

  1. 01
  2. 02
  3. 03

    The Most Famous AI Writing Tic Is Also the Most Mysterious

    www.theatlantic.com/technology/2026/07/ai-chatbot-writing-tic-negative-parallelism/687892

  4. 04

    Inside the Unofficial Movement to Save the Em Dash From A.I.

    www.npr.org/2025/11/10/nx-s1-5596088/inside-the-unofficial-movement-to-save-the-em-dash-from-a-i

  5. 05
  6. 06

    A Concerning Trend

    neil-clarke.com/a-concerning-trend

  7. 07

    The Biggest Tell That Something Was Written by AI

    www.theatlantic.com/technology/2026/05/how-to-tell-ai-writing/687345

  8. 08

    More Typos, Fewer Em Dashes: Writers Are Creating an Anti-AI 'Literary Counterculture'

    www.wired.com/story/more-typos-fewer-em-dashes-writers-are-creating-an-anti-ai-literary-counterculture

  9. 09

    Too Many Workers Are Bad Writers

    www.cbsnews.com/news/too-many-workers-are-bad-writers

  10. 10
  11. 11

    Generative AI at Work

    doi.org/10.1093/qje/qjae044

  12. 12
  13. 13
  14. 14
  15. 15

    Hemingway's Writing Style, as Faked by Five AI Models

    silentroom.media/the-method/hemingway-writing-style-ai-test

  16. 16

    What Happens When Robots Write Sci-Fi?

    www.marketplace.org/episode/2023/03/08/what-happens-when-robots-write-sci-fi

  17. 17

    A Brief History of Slop: Why Flawless Text Triggers Both Delight and Disgust

    silentroom.media/the-craft/brief-history-of-slop-why-flawless-text-triggers-disgust

  18. 18

    How to Spot AI Writing

    www.economist.com/culture/2026/07/30/how-to-spot-ai-writing