The em - dash discipline

I caught my AI agent using em dashes in nearly every sentence today.

Bishop and I had been working through a brand voice doc for about two hours. The session was deep in the weeds: punctuation conventions, voice anchors, that kind of editing. About forty minutes in, I noticed something. Every paragraph Bishop drafted had at least one em dash. Most had two. A few had three.

If you’ve spent any time reading AI-generated prose since 2023, you know exactly what this looks like. The em dash has become the canonical AI tell. Once you see the pattern, you can’t unsee it.

Wait — let me be honest. I just used an em dash. So did the agent. So did half the LinkedIn posts you scrolled past this morning.

That’s the point. The reflex is everywhere.

i.Why this happens

I don’t have inside knowledge of how Anthropic trained Claude. Nobody outside Anthropic does. But the pattern reads like an artifact of either training-data preference (em dashes are everywhere in modern editorial prose) or reinforcement-learning preference signals from human raters who like the rhythm em dashes create. Probably both.

The mechanism doesn’t matter much. What matters is the effect. Replace one em dash with a period or a comma and the prose stops reading like ChatGPT translated a Wikipedia article. The sentence-level rhythm changes. The voice surfaces.

That’s what makes the em dash interesting — interesting enough to be the canonical example. It’s not that em dashes are bad. They’re a useful piece of punctuation when you want to mark a sharp aside or a sudden swerve in thought. They’re bad when they’re a reflex, when every sentence reaches for one because that’s the cadence the model learned to produce.

ii.The discipline I gave Bishop

Mid-session, I told Bishop: AI punctuation is a no-go phrase category. Not a flat ban. A discipline check.

Rule of thumb: if you can’t name the deliberate beat shift the em dash is marking, replace it with a period or a comma.

That’s it. Six words and a conditional.

Bishop recalibrated mid-conversation. The next response had zero em dashes. Not because em dashes had been banned, but because the discipline check forced a moment of reflection before each one. Most didn’t survive. The ones that did were genuinely doing work.

I want to flag something about that exchange specifically. The correction stuck because it was a discipline, not a rule. A flat ban would have produced compliance theater: Bishop scrubbing em dashes from every output regardless of whether they belonged there. The discipline-check produced something more useful — a learned reflex that checked the reflex.

iii.What I actually learned

That single exchange unlocked something I’d been circling for months.

Brand voice work isn’t really about “what should I sound like.” That’s the question every voice consultant tries to sell you on, and it’s the wrong question. The right question is: what tics creep into my output without my noticing, and how do I codify the difference between deliberate craft and reflexive habit?

Voice isn’t a positive declaration. It’s a pile of negative space: the things you don’t do, the moves you rejected because they weren’t yours. Catching the tics is most of the work. Codifying the discipline-check is the rest.

Bishop and I spent the next hour doing exactly that. We codified seven AI-punctuation patterns that needed discipline-checks before they survived a draft.

The seventh one is the giveaway. If you ban em dashes outright, the model routes the same impulse through parentheses. The reflex is upstream of the punctuation choice. You have to address the reflex, not the symptom.

iv.Why this matters for anyone working with AI

I don’t think the em dash discipline is the point. The discipline-check pattern is the point.

Working with AI on your own writing means catching the moments where the model’s defaults bleed into your voice. Not because the defaults are bad — they’re often quite good, calibrated for clarity, polished by training. The problem is that defaults aren’t yours. If you let them through unexamined, your output starts sounding like everyone else’s output, because everyone else is also letting the same defaults through unexamined.

The discipline-check is how you keep what’s yours and let go of what isn’t. Six words and a conditional. Apply it to punctuation. Apply it to word choice. Apply it to sentence structure. Apply it whenever you feel the reflex toward listicle format the moment you sit down to write something.

v.What I’d ask you

What AI tells have you been catching in your own drafts? Or in mine?

The whole point of “earn the alignment” is that calibration is mutual. The model learns from how you correct it. You learn what’s yours by noticing what isn’t. The work goes both ways, and the goal is the same on both sides: prose that reads like someone in particular, not prose that reads like everyone in general.

Drafted with Bishop, my AI partner.
Words picked, edited, and approved by me.