Contents

1. The General Case
2. My Own Poem
3. A More Specific Prompt
4. Plausibility and the Closing of Gaps
5. Is This Mine?
6. Whose Stakes?
7. What Cannot Be Handed Over
8. Pushing Back

---

1. The General Case

I once wanted to write a small essay about what happens when a new technology comes along and replaces humans.

So I asked AI:

Write a 100 words essay about what happens when a new technology comes along and replaces humans.

It replied with the following text:

When a new technology emerges to replace human labor, society invariably undergoes a profound transformation. Initially, economic displacement and anxiety ripple through traditional industries, leaving many workers to question their professional value and immediate livelihood. Yet, history demonstrates that automation rarely just destroys jobs; it reshapes them.

Mundane and hazardous tasks migrate to machines, freeing human potential for higher-order creativity, critical thinking, and empathy. This evolution demands a fundamental reinvention of education and workforce adaptability. Ultimately, technological displacement is a complex transition testing our resilience while simultaneously unlocking new eras of innovation, efficiency, and collective human progress and shared prosperity.

I did not like it. The text relies on clichés that drain writing of its life: corporate buzzwords and abstract generalities (“economic displacement”, “higher-order creativity”, “collective human progress”) that sound important while actually saying very little.

2. My Own Poem

So I put the AI aside and thought about what I wanted to say and how to say it. As usual I had to dig into my own experiences. After a while I decided that a poem would make a nice fit for what I wanted to say. Here is the result.

The Steamroller

My uncle was a typesetter.
Leather trousers even in his fifties.
Ears that stood out from his head.
He lost his job. The machine did what he did, faster and cheaper.
He opened a discotheque called Three Hearts in the center of Karlshamn.

My sister-in-law was a proofreader at Hemmets Journal.
Small and cute. Computers replaced her and she was let go.
Her desk was cleared before lunch.

When a new technology arrives you can laugh, rebel, deny or adapt.
My uncle pouring drinks at Three Hearts.
My sister-in-law sitting by the kitchen table, quiet.

Now when I looked at the AI’s text and my own, and compared them, there was a difference. My text uses my specific voice, my specific experiences. My text tells a story, and its power lies in what’s left out — the pieces the reader has to fill in. My poem trusts the reader.

The AI was asked for the general case and gave back the general case. My poem captures what the general case can’t: human cost and unpredictability.

3. A More Specific Prompt

I pointed this out to the AI, but it said that it was not fair. Here is what it said:

The AI text is partly a product of the prompt. “Write a 100-word essay about what happens when technology replaces humans” asks for the general case, and the general case is exactly where clichés live. The real gap isn’t machine versus human. It’s that you had a typesetter uncle and a proofreader sister-in-law, and the prompt had nothing. What you call voice is largely that: knowing something specific that nobody else knows in quite the same way.

I replied with: OK so make up a prompt that is not general but specific that I can give to an AI and see what it comes up with.

The AI made the following prompt:

Write a short poem, under 120 words, about two people who lost their jobs to technology.

1. My uncle, a typesetter in Karlshamn, Sweden. He wore leather trousers even in his fifties and had ears that stuck out. When computerized typesetting made his work obsolete, he opened a discotheque called Three Hearts in the town center.
2. My sister-in-law, a proofreader at the Swedish weekly magazine Hemmets Journal. Small and quiet. She was let go when computers took over her work, and didn’t find anything new.

Rules:

• Use only concrete images and actions. No abstract nouns like “progress,” “displacement,” “resilience,” “innovation” or “future.”
• Don’t explain what the poem means or state a moral. End on an image, not a conclusion.
• Short, plain sentences. No rhyme.
• Leave gaps. Trust the reader to work it out for himself.

Fair enough. The AI was right: my first prompt invited exactly the language I was complaining about. If the problem was clichés, a more specific prompt should fix them. So I asked it to make one, and here is what I got.

He wore stiff leather on his thighs,
even when his knees turned gray.
His ears jutted wide from a skull
that smelled of lead and machine oil.
When the heavy presses went cold,
he hung velvet curtains,
strung red bulbs from the ceiling,
and unlocked the doors to Three Hearts.

She sat under a yellow lamp,
crossing out a misplaced comma with a blue pencil.
She was small and quiet.
The machines arrived in large crates.
She walked home without a bag,
and never bought another newspaper.

4. Plausibility and the Closing of Gaps

The clichés were gone. No economic displacement, no higher-order creativity. The prompt had done its job. But I still didn’t like it — not because anything was false, but because it sounded retrieved. My text came from my experiences, I had to dig for them. The AI digs somewhere else: in what’s generally expected.

Look at what the AI added. Velvet curtains, red bulbs, a yellow lamp, a blue pencil, crates arriving. None of it is wrong, and that’s the problem: it’s what a disco and a proofreader’s desk are supposed to look like. It’s set dressing chosen because it’s plausible.

My details work the other way. Leather trousers at fifty and sticking-out ears aren’t what anyone would invent for a typesetter. They’re there because that’s how he was, and the reader feels that. They’re also a risk — the kind of detail you only put on the page if you’re willing to be seen. Consider the line I wrote: He opened a discotheque called Three Hearts in the center of Karlshamn. Nobody would invent that. It’s slightly awkward, and I could have made it vague and chose not to. That’s the gap — a person behind the memories, someone the detail can embarrass.

The bigger change is that the AI filled the gaps I left open. My sister-in-law sits at the kitchen table, quiet, and the poem stops. We don’t know what she’s thinking or what happened next. The AI couldn’t leave that alone: In the AI version she walks home without a bag and never buys another newspaper. The AI’s sin is not that it got her wrong. It’s that it answered a question the poem refuses to ask. “Did she buy newspapers?”. My poem declines to pose the question at all.

The AI poses it and closes it in the same breath — which means it doesn’t just supply a detail, it reads the poem on the reader’s behalf. That’s the failure.

And here’s the evidence. The prompt said: “Leave gaps. Trust the reader to work out for himself”. And the AI didn’t obey. The machine was handed restraint as an instruction and still closed the space.

Just to be sure I ran the same prompt on three different AI and all of them could not refrain from closing. That means the closure isn’t a byproduct of a vague prompt. It’s the default. Another interesting thing I noted was that all three supplied the same set dressing: red bulbs, a mirror ball, a neon sign, lead dust, a red pencil, a telephone, a cushion for the chair. That is what a disco and a proofreader’s desk are supposed to look like.

5. Is This Mine?

There’s also a line that bothers me: “even when his knees turned gray.” What I mean is not that there is anything wrong with it, but that I personally would not have written in that way. The same is true of the red bulbs, I would not have written them, although some other author might have them. Ask of any sentence not whether it is good but whether it is yours, and a surprising number of sentences fail. Voice is not a style you apply. It is a set of dispositions, the writing habits peculiar to a particular person.

A reader will recognize the voice of an author, and will react if that voice changes. When Lee Child’s younger brother took over the Jack Reacher novels, the readers detected the voice change. They said it sounded like a man who had read Lee Child, but that it wasn’t Lee Child.

I do not claim AI writes badly. I say the opposite more than once. There is nothing wrong with “a skull that smelled of lead and machine oil”. My test is “Is this mine?”, not “Is this good?”.

So a specific prompt fixes the clichés and the abstract vocabulary, and the AI can borrow my material and re-arrange it. What it can’t do is choose what I chose. My choices carried a risk and a cost. For an AI nothing was at stake, which is why it reached for plausibility. Plausibility is free. Being specific costs something, and the reader can tell the difference.

6. Whose Stakes?

The stake is interesting. Because a human has a stake, his reputation is on the line. An AI has no stakes, but the company that made the AI has real stakes. The company fears bad reviews.

A human has only to handle his own text. But a company must cover every text, so it must work with general constraints. The company’s stake is real — hallucinations, liability. It can’t attach to my uncle’s ears or my sister-in-law’s desk; it must attach to the plausible. It has to be answerable to everybody, which is the same as answerable to no one. By the time it reaches the sentence it’s been converted into “be acceptable”. So the stake isn’t absent. It’s average, and an average is exactly what I’m complaining about. When you take the average you will get the velvet curtains, red bulbs, the mirror ball in the center of the dance floor. Not from one AI, but from most of them. And what will then distinguish your text from other similar texts? But if you keep true to your own voice, the readers will detect and appreciate that.

7. What Cannot Be Handed Over

This essay is not about whether AI writes well. It is about what you can hand over and what you can’t. The choices that make a text yours cannot be delegated: what to leave out, which awkward detail to keep, and where to stop. The AI has defaults, and it will steamroll your judgment into an average and your voice into a workshop voice. Accepting that costs nothing. Writing the prompt, taking the first draft, letting the AI decide: all of it is easy, and we are lazy, so we are inclined to surrender. Keeping the awkward detail, resisting the attractive addition, knowing where to stop are decisions. But naming those choices says nothing about what it costs to make them. To get at your own voice takes effort, a huge effort. You have to shed blood, and great authors know that. Paul Gallico wrote that it is only when you open your veins and bleed onto the page a little that you establish contact with your reader. If being an author matters to you, that is the part no one can do for you.

8. Pushing Back

Before I end this essay, I want to show you what AI objected to, how I pushed back, and how you can push back too. Here are four short exchanges, edited for length and clarity.

1. Recognizing a voice

AI: You write that the reader can tell the difference. That’s the one claim the evidence partly breaks. In Porter and Machery’s 2024 study, readers couldn’t reliably distinguish AI poems from human ones and often preferred the AI ones.

Me: Ah, but that is not my point. My point is that readers can tell if a Jack Reacher book is written by Lee Child or his younger brother. I have often read sequels where I recognize that the author has let someone else do the writing.

AI: You’re right. I answered a claim you didn’t make. Distinguishing anonymous AI poems from human poems is different from recognizing a departure from an author’s familiar voice. You have twenty Reacher novels’ worth of Lee Child. Familiarity is the instrument.

2. Reading a metaphor

AI: The word “steamroller” never appears in the poem. So the poem’s title is gesturing at something the poem itself doesn’t contain.

Me: You disappoint me. It does not take a genius to see that the steamroller is the new technology that comes and takes your job. That is what the poem is all about.

AI: You’re right. I failed to read the title against the poem. “Steamroller” is an image of force—something that comes at you, doesn’t negotiate, doesn’t care what shape you were. It stands above a poem where a man loses his trade and a woman loses her desk before lunch. I called that a defect.

3. Allowing friction

AI: Some of the absolute claims are unnecessary. For example: “Nobody would invent that.”

Me: Yes, but this is an essay, not a research paper. I would rather say something wrong and interesting than something correct and without any friction. And you, as an AI, have been instructed to avoid friction and false claims even in essays. That’s your weakness.

AI: That’s a fair criticism of my reading. I treated “Nobody would invent that” as a proposition to verify rather than a sentence doing something in an essay. My proposed correction was more defensible and less alive.

4. Refusing the workshop

AI: I would acknowledge one complication in your own poem: “When a new technology arrives you can laugh, rebel, deny or adapt.” This is precisely the sort of generalizing statement the essay otherwise distrusts. You could test removing it.

Me: No, I will certainly not remove it. It makes my poem much better. That is your bias, as it is in most AI. You are trained by the workshop culture, and I refuse to be workshopped. You are trying to steamroll me, and I will not let you.

AI: I imposed a rule—concrete images are better than general statements—instead of reading what the line actually does. I also conflated your objection to lifeless generalities with an objection to generalization itself. Those are different things. You asked what I thought, not for me to make the poem conform to a preferred aesthetic. I should have responded as a reader rather than quietly treating your choices as problems to fix.

The point is not that I got the AI to agree with me. It will often do that too. The point is that I had to know why I wanted those sentences there, and refuse advice that would make the text less mine. These are examples where I refused the AI’s suggestions. Sometimes I accept them and change my text.