I've Been Writing Em-Dashes Since the 1980s

Get rid of the AI em-dash.

You've seen some version of the note — the little correction that turns up under anything on your feed thoughtful enough to look edited. The em-dash, apparently, is a tell. A confession. Proof that a machine was in the room.

I have been setting em-dashes since the 1980s — long before there was a U+2014 to call them by. I set them in Interleaf, the technical-publishing system I learned the craft on, and back then you built one the honest way: two hyphens, --, typed by hand. (There were proper en-dashes on offer too, but nobody touched those except as bullets.) I entered those double dashes the same way I put two spaces after every period — a reflex drilled into the fingers on an IBM Selectric, and as reliable a tell as any that you learned to type in the era I did. So let me say it plainly: I earned the right to have a machine set them for me.

The Mood in the Room

The complaint didn't arrive in a vacuum. It arrived in the middle of a backlash you can feel in the air like weather. AI executives are getting booed off commencement stages. Billionaires are taunting the graduating class with a message that boils down to we are taking your jobs, there is nothing you can do about it, and your one available move is to speculate on our stock on the way down. The anxiety underneath all of it is real, and mostly earned.

The em-dash sneer is the retail version of that anxiety. A person who cannot boo a billionaire boos a punctuation mark instead.

The Sneer Has Nowhere to Land

Here is the part that gives the whole game away, and it is a point about biology.

When you sneer at a sentence a machine helped write, you are trying to land a blow on something with no limbic system. The prose has no adrenal gland. It cannot flush with shame or lie awake at 3 a.m. Insult it, insult its mother, and the only nervous system in the entire transaction producing any chemistry at all is yours — a small squirt of dopamine for the person who threw the punch, landing on a target built to feel exactly nothing.

That is the quietly absurd thing about scolding a bot for its em-dashes: the whole emotional event happens inside the scold. It is a carbon reflex — the ancient, useful instinct to shame a rival back into line — misfiring against silicon, which has no line to be shamed back into. This is most of what I write about. Carbon runs on chemistry; silicon doesn't. You cannot hurt the thing you are trying to hurt.

And for the record: I use AI to write, and I have never once pretended otherwise.

Yes, It's a Slop Machine. That's the Whole Point.

Let me grant the loudest objection before I take it apart. AI produces slop. It generates coherent gibberish at prodigious volumes — a brain dump so vast that no human, working with our thimbleful of attention, could ever hope to drink it. Point a model at a topic and it will hand you ten thousand competent, forgettable words before you've finished your coffee. Most of what clogs your feed is precisely that: raw model output, shipped straight from the machine to your eyes with nothing in between.

And let me be honest about where I stand. I am not a writer. I did a few months as a technical writer in 1990, documenting telecom report screens, and made nothing anyone would mistake for art — so I have no craft to be precious about. Which is exactly why I can say this next part without a shred of defensiveness: the people who did give years to the craft, who learned to cut and weigh and rewrite until a sentence earned its keep, and who now watch people with no talent and less taste shovel mountains of slop into the same undifferentiated feed — they are right to be furious. The slop is real, and it is burying them. My quarrel has never been with them.

But here is what the people sneering at the surface have never had to learn.

When you write software, the lines you read at the top of the file are a courtesy — a thin, legible skin stretched over an enormous body of machine instructions. A compiler lowers your tidy function stage by stage: first an intermediate representation, then assembly, the terse mnemonics written for one specific chip. An assembler turns that assembly into object code — the actual machine instructions, in binary. A hex editor can render those bytes in base-16 for the rare specialist who needs to read them, but the ones and zeros underneath don't change. Nobody hands you the compiled binary and calls it the source.

A model's raw output is that low-level sprawl: technically valid, machine-shaped, not built for a human to read. The prompt is the real source you wrote. The inexperienced operator runs it once and posts the sprawl as though the value were sitting in it. It isn't. The value is in the editing pass that distills that sprawl to the density a human being can actually use.

And a confession that turns out to be the best evidence I could offer: a machine drafted that compiler paragraph, and it got the pipeline wrong — assembly and object code transposed, the assembler (a tool) mistaken for a layer, hexadecimal treated as a stratum of its own rather than a way of looking at the one below. Fluent, confident, wrong, in the manner of a bright sophomore. I caught it because I've spent my life around this machinery. That catch is the whole argument: the model supplies plausible, and a human who knows the domain supplies correct. Leave the error in, and you've shipped the intermediate language — which is exactly what I'm telling you not to do.

And the part that should make a purist's teeth itch: I didn't fix that paragraph by hand. I never opened an editor to re-order the stages myself. I pasted the wrong version back into the machine, told it what it had botched, and asked it to correct the error and hold the correction up as an example — which is the paragraph you just read. The judgment was mine: knowing the pipeline was wrong, knowing why it mattered, deciding it earned a place in the piece. The keystrokes were the machine's. That division of labor is the whole thesis folded back on itself — an essay about curation-by-prompting, curated by prompting, including this confession, which I also prompted. Tautologies all the way down. And the tautology is the point.

Curation Is the Entire Job

The skill nobody wants to credit is writing the prompt — and then the second prompt, and the tenth — until the machine self-curates its own firehose down to the level worth a person's attention. Entertaining, if you're lucky. Thought-provoking, if you've done the work.

The ideas in these essays are mine. The analogies are often the machine's. And the tells everyone loves to point at — the profligate em-dashes, the reflexive it's not X, it's Y, the way everything in a lazy AI sentence happens "quietly," the breathless "this changes everything" — those I spend real effort prompting back out, because leaving them in is the signature of someone who shipped the intermediate language and went to lunch.

So when someone reaches for the em-dash instead of a single idea in the piece, the reach tells on them. They've judged the work by a surface marker they've decided is disqualifying, without troubling to find out whether the thing underneath is any good. It is the same reflex that once ruled photography wasn't art, that the synthesizer wasn't music, that the word processor was the death of "real" writing — a refusal to weigh the output on its merits because of a prejudice about the tool that made it. The marker becomes the entire argument, so no one has to reckon with the substance.

And here is the joke buried in the sneer: every writer carries tells. Two spaces after a period say I learned on a Selectric. A double hyphen says I set type before Unicode had a name for the character. A profligate em-dash says a machine had a hand in the draft. A tell is a fingerprint of how you were formed, not a confession that the work is worthless — and mistaking the one for the other is the whole of what's gone wrong here.

Written in Sand, Just Above the High Tide Line

So why bother, if it earns you a sneer and a slot in front of the two hundred people an algorithm has decided to show? I'm not viral. I don't expect to be. I do it for three reasons.

I want to read essays about how carbon and silicon intelligence fit together, and almost nobody is writing them — so I have the machine help me write the ones I want to read. I carry far more unarticulated ideas about all of this than I will ever have hours to grind into the kind of prose I'd actually want to read; the mechanics of the sentence are the bottleneck, never the idea. And the alternative to an imperfect, AI-assisted essay is no essay at all — the thought just evaporates, lost like tears in the rain.

These pieces are written in sand, just above the high tide line. The water will take them eventually; I have no illusions about that. But they'll sit there a little longer than the ones scratched down at the waterline — long enough, maybe, for someone to read one before the tide comes in.

I'll keep letting the machine set my em-dashes. I earned them.