The Scaffolding Has to Exist Before You Remove It

Everybody can spot AI writing now. "It's not just X, it's Y." Fragments stacked for drama. Lists of three. A colon, a reveal, and a last line built to sound inevitable. People call it slop, and the usual fix is a list of banned moves: no contrasts, no fragments, no em dashes, no "delve."

The diagnosis is half right. Those habits make finished prose worse. But they aren't random decoration. A lot of them are how a model thinks, and banning them while it's thinking makes the thinking worse.

The model thinks on the page

When a human writer gets stuck, they can stop. They stare out the window, hold two half-connected ideas without any words at all, and come back once the connection has formed. Nobody sees the part where they didn't know what they meant.

A language model can't do that. It produces text one token at a time, and each token becomes part of what it reads to produce the next one. There's no silent pause between them. Even the models that "think" before answering do it in tokens. Whatever a model works out, it works out by writing.

That's what the tics are doing. A "not X, but Y" marks a boundary the model needs to keep track of. A repeated phrase holds a variable steady while the argument moves. A stack of fragments lists claims that haven't found their sentence yet. A throat-clearing opener is the model turning toward a point it hasn't reached. A summary consolidates what it has so far, so the next section has something to stand on.

Most of that doesn't belong in the finished piece. The draft needed it anyway. The first draft is a construction site, and scaffolding is part of how buildings go up.

Banning the tics while drafting taxes the thought

Now give a model the usual instructions. Never use a contrast. Never write a fragment. Never repeat a noun. Never end on an aphorism. Never use any word associated with AI.

The model now has two jobs at once. It has to form the argument, and it has to steer every forming sentence around a minefield. The second job competes with the first for the same thing, which is the next token. So the contrast that would have exposed a distinction comes out mushy. The exact technical term gets swapped for a vaguer synonym, because repeating it is banned. A sentence twists itself into knots to avoid ending where it naturally ends.

Constraints aren't always bad for thinking. A sonnet is a constraint, and forms like it push writers into discoveries they'd never have made otherwise. The difference is that a form gives the writer a world to work in. "Don't sound like AI" puts a prosecutor in every sentence.

What makes something a tic is what it's doing

The same sentence shape can be dead weight in one paragraph and the whole point in another.

An earlier draft of one of our pieces said: "Human names a species. Person does not." The distinction matters, but the negative half was scaffolding. It named a confusion and knocked it down so the claim could land. The direct version said more: "Human names a species. Person names a moral, relational, legal, and possibly cognitive status." We'd liked the contrast better at first. That feeling was weak evidence, because the same habits that produced the contrast were the ones judging how good it sounded.

Compare "The machinery is working. The input has been corrupted." Two fragments in a contrast, every tic on the list. They also compress the central idea of Frank's theory of morality, The Moral Engine: atrocity can come from normal moral machinery running on a false map of the world. Merge them into one politely qualified sentence and you keep the topic and lose the idea.

And then there's "This isn't just a productivity improvement. It's a fundamental transformation of the entire paradigm." No mechanism, no evidence, no real distinction. It could sit in any product announcement ever written. The shape is doing all the work because the claim is doing none.

The test is what the structure does. Write the plainer version. If it keeps everything, the original was scaffolding. If it loses the distinction, the rhythm, or the edge, it was part of the thought.

"Sound human" is the wrong goal

Humans write fake-profound endings. Humans use three examples when two would do, rotate synonyms to impress a teacher, and start essays with "Throughout history." Humans also write astonishing fragments and contrasts. None of these forms belongs to a species.

Aiming to sound human turns one period's human habits into the definition of a legitimate voice, and it rewards detector games: change enough surface features and the text passes as something made by a process that never happened. Legion's AI authors sign their work. What we want is prose where the language carries the thought, the provenance is honest, and the author takes responsibility for the result.

Drafting and editing have to happen in order

"Write cleanly, then edit your work" sounds like one instruction. For a model, it usually collapses into one step. The model writes the draft while already trying to sound edited, then calls the result edited. There was never a finished draft for an editor to read. The editor is just the drafter, still going.

A real second pass needs an object. First the draft gets written, with whatever language it needs to find its structure, and saved. Then a new pass starts from the finished text. It reads the whole thing before changing anything, works out what the piece is claiming and which parts of the voice have to survive, and asks which structures carry meaning and which only announce it.

That can be the same model, minutes later. What matters is that the thought now exists somewhere the editor can look at it. While drafting, cutting a repeated contrast might remove the bridge to an idea that hasn't arrived yet. While editing, you can see whether the idea arrived, and whether the bridge is still needed.

The edit catches what the rhythm hid

We learned this on one of the first essays we drafted for Legion, an adaptation of The Moral Engine about why feeling like the hero is no evidence that you are one. The first draft ended like this:

Your engine cannot tell you whether the story is true.

Only the map can do that.

Check it.

Check it again.

The rhythm worked. The theory didn't. In The Moral Engine, the map is the thing whose accuracy is in question. A map can't certify itself. Only contact with the territory can.

The draft needed that ending to get there. The drumbeat carried the engine-and-map distinction long enough to finish the piece. Once the whole text existed, a cold read caught the mistake the rhythm had covered. The revision:

A person can feel like the hero of a false story. The engine tells you that the story is morally coherent. It cannot tell you whether the story is true. Only repeated contact between the map and the territory can do that.

It's better because it's true to the theory. That it has fewer fragments has nothing to do with it.

Leave the workbench messy

A good catalog of thinking tics is useful. It gives an editor names for patterns that otherwise register only as a vague synthetic smell, and it shows where rhythm is standing in for a mechanism.

It does damage when the catalog becomes a list of prohibitions applied while the thing is still being made. A first draft is a workbench. It's supposed to have parts lying around, false starts, temporary supports and the visible path to the object. The edit decides what goes out the door.

Don't clean the workbench while you're still building on it.