The tell you can't ban

Someone sent me three paragraphs last month and asked whether a person had written them. I knew before the end of the first line. Not from a mistake in them, the words were right and the grammar was clean. It was the distance, the way you place a stranger by their walk before you can see the face. If nothing in the writing is wrong, what’s left to give it away?

The tells you can remove

Most of what marks text as machine-made comes off, and I’ve spent a while taking it off, one rule at a time.

The em-dash scattered through every paragraph like seasoning: gone, by rule. The small icons that sit beside a heading to look friendly: gone. The flat committee voice that says everything and risks nothing, traded for a written character with its own vocabulary and its own things it would never say. Each one is a surface feature. You can name it, write a rule against it, and watch it go.

After enough rules the text comes out clean. And clean still reads as machine-made. That was the part I didn’t expect.

The tell that has no switch

The last one is rhythm, and rhythm has no switch.

You cannot forbid a shape the way you forbid a character. The move where it was never about the plan but about the distance, the list of three with the surprise saved for the third, the short fragment dropped in for punch, the closing line that bows back to the title: none of these are words. They are the meter underneath the words, and a list of banned characters never sees them.

These shapes fill generated prose because they filled the prose the model learned from.

Why the rhythm is the average

A model writes by guessing the most likely next piece, over and over. Do that across a paragraph and the sentences drift toward the most likely shape of a sentence. Do it across a whole essay and the essay drifts toward the most likely shape of an essay. The cadence is not something the model chose. It is the middle of everything it has read, and the middle belongs to no one.

What you sense from across the room is that average: the texture of the exact middle, with everything that would have stuck out smoothed away.

A voice you recognize is built from the parts that stick out: an odd sentence length, a plain point where the form wanted a flourish, the place where a real writer got bored with the pattern and broke it. Prediction strips those first, because each one is the unlikely move. The same pull that makes the model read so smoothly drags it toward the middle of the road.

You cannot subtract your way to a voice

So the rules only carry you part of the way. Strip the em-dashes, the icons, the committee tone, and the obvious marks are gone. You arrive at clean. Clean isn’t the same as alive. I could ban one more thing tomorrow, and the text would come out cleaner and not one degree more alive.

What is missing cannot be added by another ban, because the human signal is the variance, and you do not get variance by taking things away. You get it by putting back what the average removed: a rhythm that follows one particular train of thought instead of the remembered shape of thought in general. Sentences a person could have said out loud while thinking, rather than ones built to sound like writing.

It is fair to say this argument reached you through a machine. Read it back for the rhythm anyway. If it ever settles into a meter you have met a thousand times before, you have caught the exact thing it warns about, and you should trust that feeling over anything I have told you.