A good argument gets overtaken by a style question

If you write with AI, and most people now do, the thing worth understanding is that a good argument will lose to a question about style. This is early, I’m working it out, and I’d rather record what I got wrong than pretend I planned it.

A few days ago I published a piece I was pleased with. It argued that spotting AI writing has become the least interesting question available, and it had decent evidence behind it: The Economist had compared its own journalism against four models across 55,940 sentences and found that the tell everybody relies on has inverted. Only one model now uses em dashes more than human writers. The one that taught everybody to be suspicious of them uses fewer than anyone in the study.

The response was almost entirely about whether the post itself was AI-written.

I had noticed the tells before publishing — short paragraphs, quick two-beat contrasts, a fair amount of white space. I left them because that format works on a feed and the post read well to me, and I didn’t think the choice would need defending.

It did, and the lesson underneath it is more useful than the argument I was making.

The tells were real, and I said so first

I use AI heavily, I’ve written about the mechanics of it, I disclosed it in the post itself, and when somebody said they could tell, I agreed and asked what they were seeing.

They weren’t wrong. I’ve since measured my own corpus, and the pattern is real and quantifiable: in the weakest pieces on this site more than 60% of paragraphs are a single short sentence, and the mean sentence length collapses to around ten words. That produces a particular rhythm, and people who read a lot have learned to hear it. It isn’t punctuation and it isn’t vocabulary. It’s pace.

One thing fell out of that measurement worth passing on, because it contradicts the folklore. Em dashes tell you nothing. One piece on this site has fifteen per thousand words and none of the cadence; another has none at all and the cadence throughout. Anybody hunting em dashes is using a signal with no information in it.

Style is checkable from the doorway. Argument isn’t.

The mechanism is worth naming, and it isn’t unfairness.

Noticing a cadence takes seconds and needs only the first paragraph. Deciding whether an argument holds means reading the whole thing, knowing something about the subject, and forming a view you would defend if pressed. One of those is free. The other is work.

So the style question wins by default, in the same way that a comment about a typo always outperforms a comment about a thesis. Not because people prefer the shallow question, but because the shallow one can be answered from the doorway while the deep one requires coming inside.

That is the part I would want anyone using these tools to absorb. However good your argument, a question about how the text reads will reach it first, and will set what the conversation is about. You can be right and still lose the room on pace.

What I actually got wrong

I have been treating prose as the layer that matters least.

The reasoning seemed sound. Production is what models are best at and getting cheaper at, so the durable value must sit upstream in evidence, argument and the decision about what deserves to exist. I still think that is broadly right, and it is most of what this project is about.

But it misses something I have now had demonstrated. The surface decides what the work gets discussed as — not because the value lives there, but because it is the first thing anybody forms a view about, and that view frames everything after it. Judgement nobody discusses is hard to distinguish from judgement that isn’t there.

That is a correction to my own position rather than a defence of it. The post had the judgement in it: a corpus study rather than an anecdote, numbers I had checked, a question I meant. On my own account of what matters, those are the things it should have been discussed for. They weren’t, and the reason was pace.

The construction error was mine

There is a second thing, more specific and entirely self-inflicted.

Write about how to detect AI and you have instructed your readers to treat your own post as a specimen. That instruction is more available than the argument and it arrives first. Nothing else I publish carries it.

So the response was not people missing the point. It was people doing exactly what the piece had just told them to do, on the nearest available text. If I write about detection again, it needs to make the interesting question easier to reach than the boring one, and I don’t yet know how you would manage that.

What I’m taking from it

I am not going to start writing to evade detection. That is the trap the original post identified: make this sound less like AI is a problem with a shelf life, and optimising against it means chasing a target that moves with every model release while accepting a frame I don’t accept.

The target that doesn’t move is whether something reads as though a person thought about it. The two overlap today. Only one of them still works in three years.

What changes in practice is narrower. I will stop treating pace as cosmetic, because it isn’t — it is the first thing a reader uses to decide whether to spend four minutes on you. That means fewer stacked fragments, sentences long enough to hold a qualification, and noticing when white space is doing work the argument should be doing. None of that is writing worse. The strongest pieces on this site are the ones furthest from that cadence, and they are strongest because a longer sentence can carry a because while a fragment can only carry a claim.

I would also separate two jobs I had been running together. A post on a feed can keep some rhythm, because scanning is real and line breaks genuinely help. An essay should not borrow that rhythm, because an argument needs room to qualify itself, and it is the essay where I had been letting the feed habit leak in.

The awkward part: it was the week’s best-performing post

I nearly published this piece without checking the numbers, which would have been careless given what it argues.

That post was the strongest thing I published on LinkedIn that week — roughly two thousand impressions against 588 for the next one in the same period, and more engagement than anything else I posted. Not a record, and I should be precise about that, because the panel I read it from covers a period rather than the whole account. But whatever the cadence did, it plainly did not stop people reading it.

So the claim I was reaching for — that style stops an argument being received — is too strong, and the evidence I have contradicts it. The accurate version is narrower and more interesting. The post travelled. The argument didn’t. Nearly a thousand people saw it, and the conversation underneath it was about prose rhythm from beginning to end. The cadence bought the attention and then spent it on itself.

There is a second thing I should concede, because it undercuts the tidy version of this essay. The cadence people call “AI style” on LinkedIn is substantially LinkedIn style. Short paragraphs, line breaks, two-beat contrasts and a hook on its own line were the house format of that platform for years before ChatGPT existed, which is presumably how models learned it — that material was in the training data. Calling it a machine’s voice is close to backwards. It is a platform’s voice that a machine absorbed and handed back.

Which makes the detection game odder than it first appears. Somebody objecting to a LinkedIn post for reading like a LinkedIn post is describing a genre rather than a provenance, and a lot of the writing being confidently identified as AI is people using a format that has worked on that platform for a decade.

I would also allow the obvious proportionality point, which I did not want to hear at first. A social post is not War and Peace. It is a short form with its own conventions, read at speed, and holding it to the standards of an essay is a category error. I can see tells in most posts I read now and mostly do not mind, because the form is doing a different job. It would be inconsistent to demand a licence for my own and grant it to everybody else’s.

And there is a simpler explanation I should not skip past, because it does not require AI to be involved at all. Publishing an argument in public attracts objections. That has always been true, on every platform, for everybody who takes a position rather than describing a situation. Some proportion of responses to anything with a claim in it will be dismissive, and the dismissal reaches for whatever handle is nearest. At the moment the nearest handle is AI. Ten years ago it would have been something else, and if I had written every word by hand a version of this thread would still have happened — about the argument, the tone, or my standing to make it. That is not a reason to ignore what the responses said. It is a reason not to treat the AI framing as the whole explanation when a good deal of it is just the cost of saying something.

Where I’m least reliable on this

“They didn’t engage with my real argument” is what everybody says. It is the standard complaint of people whose argument was engaged with and found thin, and I can’t rule that out from the inside. I think it held. I would think that.

The piece may simply not have been good enough to earn the deeper conversation. A strong enough argument drags people past the surface, and blaming the medium is a comfortable way of avoiding that possibility. I am aware I have just spent a thousand words doing something adjacent to it.

And I am treating a small number of responses as a finding. Two people, one thread, one platform, one week. That is an anecdote with measurements attached rather than evidence. The corpus numbers are solid because they run across 89 pieces; the claim about how readers behave rests on considerably less, and I would want to see it hold somewhere else before leaning on it.

All of this is very new, which is the only thing I would ask anybody to grant. The norms around AI-assisted writing are being invented in public by people who disagree with each other, myself included, and anyone claiming to have settled the question is describing a position rather than a finding.

What I would offer is small. If you are using these tools, and most people now are, assume your argument will be overtaken by a question about your style. Attend to the style — not to hide the tools, but because it is the toll you pay to get the argument read at all. I didn’t think I would have to defend leaving the tells in. I did, and I would have been better off spending that hour on the prose.