My AI editors keep correcting each other, and I keep sounding like whichever one won

A model doesn’t need to supply the idea to shape the writing. It can impose rhythm, structure and emphasis while leaving the argument apparently intact — and then hand the rhythm to a rival.

I had three Field Notes reviewed that I’d developed with ChatGPT. The diagnosis was that the ideas were good and buried in a recognisable cadence — stacked fragments, bold phrases standing in for argument, lists where sentences should have been.

Once it was named, I could see it everywhere.

A single point split across five short paragraphs to manufacture emphasis. Abstract nouns arranged in a list instead of connected by an argument. The central idea repeated in bold, as though typographical confidence could do the work the prose hadn’t done.

The writing wasn’t bad. The ideas were mine and mostly sound. But the rhythm had stopped being mine, and I hadn’t noticed it happening.

One correction before I go further, because it matters for everything after it. I noticed something was wrong first — I sent the pieces over saying I didn’t like them and couldn’t say why. The contribution wasn’t the detection. It was describing the pattern precisely enough that I could act on it. Those are different things, and the more flattering version of this story is the one where I hadn’t seen it at all.

Why there was a process at all

Using AI hasn’t removed editorial process from my work. It’s given me a way to build one, and this is the first time that process has caught something I genuinely couldn’t see myself.

There are three editors and four passes, and one of the editors is me.

The idea has to come from somewhere, and that somewhere is usually an observation, a conversation, a disagreement, or something I can’t stop turning over. Working out what I actually think, and whether it’s worth saying, is the first editorial decision and the one that can’t be delegated.

ChatGPT gets it next, though not a blank window — it works against the accumulated Editorial Intelligence context of frameworks, research, previous arguments and unfinished thinking. That pass is developmental. What does this connect to? Where’s the argument? Am I contradicting something I’ve published? What’s missing?

Claude gets the same context and a different brief. Read this coldly. Find the weak argument. Tell me if I’ve said it before, or if it isn’t worth publishing at all. Sometimes the second editor should disagree with the first, and that disagreement is most of the value — two models agreeing doesn’t make an argument true, it just means the same blind spot survived twice.

Then it comes back to me, and I change things, cut things, and occasionally decide not to publish.

That’s the structure. What follows is the first evidence that it does anything, and the first evidence that it can go wrong.

Every model has a gravitational pull

People discuss AI writing as though there are two possibilities: the machine wrote it, or a person did.

What I’d underestimated is that a model doesn’t need to originate an idea to shape the writing.

Use one repeatedly and its preferences begin to feel like decisions. Short paragraphs read as directness. Bold phrases read as conviction. Repeated contrasts read as clarity. The habits become invisible precisely because they’re consistent — and consistency is very easily mistaken for voice.

That’s the mechanism worth naming. Not the machine writing for you. The machine’s defaults becoming the thing you think you sound like.

The second model spotted it because its preferences are different: longer argumentative sentences, fewer headings, more concession, less formatting carrying the load. That doesn’t make it right about prose. It made it useful as a detector of a pattern the first model and I had stopped seeing.

Which is narrower than I’d like. Two models don’t automatically produce better work — they could just as easily average two sets of generic defaults into something worse than either. The value came from the contrast being legible, not from one model being better.

The bit that should have been obvious

Here is the detail that makes this difficult to argue with.

I’d already published a piece describing this exact process — three editors, four passes, the whole thing. It was written in precisely the cadence I’ve just been describing. One-sentence paragraphs stacked to create emphasis. A bolded word doing the work of an argument. An arrow diagram standing in for a sentence about how the stages connect.

The essay explaining why I have an editorial process was itself a demonstration of the failure that process is supposed to catch. It has been absorbed into this one, rewritten, which seemed better than leaving it standing as evidence against itself.

I’d like to claim that’s ironic. It’s closer to instructive. A process only catches what somebody points it at, and nobody had pointed it at the prose.

Then the other one overcorrected

This is where it gets funny.

I took the criticism back to ChatGPT and asked it to look at the edits. It agreed, examined them properly, and updated the writing instructions it keeps for me: favour continuous prose, fewer headings, develop arguments rather than listing them, include concessions and counterarguments, avoid conclusions that arrive too neatly.

All sensible. All roughly what I’d asked for.

Then I gave it a new Field Note to draft.

It came back enormous. Dense, heavily qualified, every claim wrapped in its own counterargument, the central point buried in the middle under several layers of reasonable hedging. In trying to stop sounding like ChatGPT, it had produced an exaggerated impression of Claude — all the concessions with none of the reason for making them.

I told it that it had gone too far. It agreed immediately, and said it had treated the criticism as a complete style prescription rather than as an editorial correction. Which is exactly what had happened, and a rather good description of it.

So one model found the other’s fingerprints on my writing, and the other tried so hard to remove them that it became a parody of its rival. My prose has spent the week being fought over by two machines, and at no point did the ideas change.

A correction is not a constitution

The useful part isn’t that one model writes better than the other. It’s what a model does with criticism.

When a human editor tells you your writing is too fragmented, you understand the scope of that instantly. It means these paragraphs, this piece, probably this habit — not that every sentence you write from now on must be long, subordinate and qualified. You know which parts were working. You keep those.

A model has no equivalent sense of proportion. Give it a critique and it will apply that critique with total consistency, because consistency is the thing it’s best at. It cannot tell the difference between “this bit is weak” and “your entire register is wrong”, so it treats every correction as though it were the second.

Which means the correction gets applied to the parts that didn’t need it, and the qualities that were actually working — pace, directness, the ability to land a point without three sentences of throat-clearing — get sacrificed along with the fault.

That’s not a failure of writing ability. It’s a failure of scope, and scope is a judgement about how much of a thing a criticism covers.

I should be fair about who caused it. I passed on a critique without qualifying it. I didn’t say this applies to prose rhythm in these three pieces, or keep the pace, lose the fragments. I handed over a diagnosis and let a system that applies everything consistently apply it consistently.

Getting a caricature back was a reasonable response to the instruction I actually gave.

The part I’m least comfortable with

There’s a version of this where human judgement remains reassuringly central, and it would be true and almost meaningless.

The specific version is worse. Judgement here meant accepting criticism of work I’d already published, and of a process I’d spent months building. It meant separating a stylistic preference from an actual weakness. And it meant noticing something I’d rather not have: that I’ve been defending AI-assisted writing against people who dismiss it as slop while letting fairly obvious model habits into my own work.

That doesn’t invalidate the argument. AI-assisted writing can involve research, judgement and care, and mine does. But running a sophisticated process doesn’t exempt the output from sounding generated — and claiming to care about the craft creates a stronger obligation to notice when the system is flattening it, not a defence against the accusation.

I should also disclose that this piece was produced in the same session as the rewrites it describes, using one of the two models it keeps discussing. If you wanted a demonstration of how hard it is to get outside this, it’s probably right here.

What the human job turns out to be

Good editing here wasn’t choosing between two models’ defaults. It was deciding which parts of the criticism were correct, which parts of the original were already working, and changing only the things that genuinely weakened the writing.

That’s three separate judgements, and the middle one is what both models are worst at. Neither has any attachment to what was already good. Ask either to improve something and it will improve it, in the direction you pointed, whether or not that direction should have been applied to the whole.

The multi-model process is genuinely useful — one model exposing another’s habits is real, and it worked. But it produces raw material for a decision rather than the decision itself. The models can show you the pattern. They can’t tell you how much of your writing the pattern is allowed to eat.

The other outcome was that the criticism became specific enough to write down. Not “sound more natural”, which is far too vague to act on, but: avoid stacked fragments, don’t use bold as a substitute for reasoning, turn lists into prose where the relationship between the ideas is the point, make room for a real counterargument.

That instruction is stronger than a general preference because it came from an observed failure in real work. Editing one article fixes one article. Changing the system is what lets the lesson reach the next one — which is the argument I keep making about workflows outliving models, arriving this time as something that happened to me rather than something I recommended.

The joke is that I asked one machine to sound less like a machine and got a very good impression of a different machine. The part worth keeping is duller: a criticism without a scope is just a new set of instructions, and somebody still has to decide how far it reaches.

The ideas can be yours and the cadence can still belong to whichever model corrected it last. Caring about the craft means being able to hear the difference — and accepting that you’ll usually hear it late.