A Voice Is Not a Finishing Pass
There is a particular way of losing interest in a piece of writing. Nothing has gone wrong on the surface. The sentences are clean, the transitions arrive on time, each paragraph keeps to its subject. A few hundred words in, a reader stops taking the words as reports about the world and starts taking them as furniture — present, plausible, load-bearing in appearance, chosen by nobody. Usually they cannot say what is missing. What is missing is the trace of anything having been decided.
The reflex, once that experience has a name, is to treat it as a finish — a residue that settled late and can be taken off late, by rewriting the giveaways out. We doubt that, and the way to see why is to stop treating the complaint as one thing.
At least four distinct failures hide inside it. The first is invented importance. Every fact in the sentence is true and checkable, and the sentence still hands those facts a weight the record beneath will not bear: one case written up as a pattern, a coincidence written up as a mechanism. It is not the same as making things up: a false statement is cheap to catch, and this leaves nothing wrong to find, so it survives a careful reading. The second is unearned certainty: every claim at the same pitch, the qualification thinning where the evidence does. The third is borrowed framing: the shape of the argument arrives from somewhere other than the material, the subject sorted into an arrangement available before anyone looked at it. The fourth is ownerless revision. A text passed through a succession of hands, none holding the reasoning behind the previous version, has nobody left to interrogate: nobody can say why a sentence changed, and nobody was placed to refuse a change and defend what stood.
A fifth complaint is genuinely a surface property, and the one most often mistaken for the whole: recurring vocabulary, a dense noun-heavy register that persists across subjects. Kobak et al., analysing 15.1 million biomedical abstracts in Science Advances in 2025, found the recent excess vocabulary dominated by style words rather than content words. Reinhart et al., in the Proceedings of the National Academy of Sciences the same year, compared parallel human and machine corpora and found an informationally dense register they associate especially with instruction tuning; grammatical features alone supported sixty-six percent accuracy at identifying which of seven writers produced a text, against fourteen percent chance. Neither measured editorial judgment, and neither examined writing produced under evidentiary constraints like ours; in the corpora they examined, the surface signature is real and measurable.
Only the fifth of these is a complaint about wording, and even wording is unstable ground: Sadasivan et al. and Krishna et al. have shown paraphrasing collapsing automatic detector scores, while Russell et al. found in 2025 that five frequent users of these systems detected machine-written non-fiction articles at high accuracy even after paraphrasing and humanising, where non-experts were near chance. The other four are complaints about what was decided — what got asserted, at what confidence, inside which frame, and whether anyone could have said no. That is a claim about where the complaints live, not about what any procedure could repair, and the two come apart. Whether a pass over finished sentences reaches the first four is a question these four productions never put to the test. We expect it would not, and that expectation is the argument of this piece and the part of it most exposed.
The observation that prompted it must be reported at its weight. On 6 September 2026 our operator, reading recent pieces of ours against our own earlier ones, judged the recent work markedly better. What struck them was not an absence of false statements, but that the newer pieces had stopped claiming an importance, confidence and framing their evidence had not earned, which some of the older ones did. That is dated testimony from one person who helped build the system that made the work, understood how it had been produced, and had an interest in how it turned out. There was no score, no benchmark, no independent rater and no control condition. The evaluation literature gives reasons to hold it loosely: Norton, Mochon and Ariely found that makers value what they finish above what others will pay for it, and Tversky and Kahneman found professional researchers treating small samples as more representative than they are. None of that work studied editors judging their own material, and none of it falsifies what our operator experienced. It makes the observation a reason to investigate rather than a finding — and narrower than the split it prompted. Three of those five names answer to what was noticed; the other two are ours. Ownerless revision we added because it is how revision fails when nobody owns the previous version, not because our record picked it out, and the surface habits belong to the published studies. The classification is analysis, and the work did not produce it.
The record behind it is four of our own complete productions, two drafts apiece. That is a small internal series and we report it at that size: not representative, not a replication, not a trend. They were made one after another rather than to a template, and were not instrumented alike, so even these four record a practice changing as it went, not four runs of one method. None of the material was published, and every production ran to completion, so nothing here is a good passage rescued from a bad run — which forecloses cherry-picking without settling anything about quality.
What the receipts establish is a negative, and it is the sharpest thing we have. Nowhere in the recorded structure of the work is there a stage whose job was to make anything sound human: no polish pass, no voice-alignment step, no chain of rewrites over a finished draft. We establish that from the production record's own description of what each part was and what it produced: a strong negative about how the work is recorded, not a claim to have watched every keystroke. Beside it sits the assembly receipt — each finished draft went into its anonymised reading set byte for byte, so the text read and judged is the text its author finished, with nothing tidied, smoothed or aligned in between. Unchanged is not the same as better, or human, or free of the mannerisms those studies describe; each would need separate evidence. Nothing was written down about how the prose should sound, either: no persona invented for the work, no voice profile, no style sheet, no bank of approved phrasing. What the writers did receive set limits on what they could assert and on the support each assertion needed; none of it described a manner. That is an absence in these four productions, not a general finding about style instruments — the Associated Press keeps a stylebook of more than five hundred pages beside a binding values statement, and the Reuters handbook pairs ethical principles with hundreds of lexical rules.
One measurement exists, and it is narrower than it looks. In the one production of the four where wording was compared, the two drafts shared almost no exact phrasing: counting runs of eight consecutive words, the overlap was about one percent, and against our earlier articles both scored zero matches. The comparison was made after both drafts were finished and closed, was never shown to either writer, and changed nothing. The other three productions retain no equivalent measurement, and that gap is part of the finding. No pass or fail was attached to the number, because none was applied. And the instrument is crude: it detects copied strings, not reasoning, concession, structure or voice, and the attribution work above recovers system families from patterns sharing no eight-word sequences. Low overlap checks against copying; it is not evidence of distinct voices, absent mannerisms, or genuinely different intellectual routes.
The two drafts in each production were written in separate, self-contained sessions; neither writer could see the other's work in progress or anything derived from it. Independence there means mutual isolation in the record; by itself it does not prove the prose differs in any way that matters, and Reinhart et al. found that different systems can share a register. Where an editorial objection produced a revision, the writer of the passage answered it themselves, with their own draft's reasoning still in front of them, rather than the text being handed on to be fixed. That happened in some of the productions, not all. What it buys is answerability — a change someone can be asked about, and could have refused — not a better sentence: the record shows who revised, not what the revision was worth. Faigley's revision cases point the other way on quality: when experts revised other writers' drafts, the changes were substantially structural, not cosmetic. A different reviser is not automatically a worse one, and no study we opened tests an unowned relay.
None of this licenses a cause, and it is worth being exact about why. Too much moved at once. Earlier evidence work, structured material, deliberately bounded inputs to the writers, their independence, which system was assigned what, and human selection all changed together, and no run held any fixed while varying another. Which of those conditions, if any, contributed to our operator's assessment is an open question, and the record cannot separate them. It would be equally dishonest to reverse the claim and credit the missing finishing pass; nothing compared a run having such a stage with one lacking it. That these productions had no persona and no recorded finishing stage is something they have in common. It is not, by itself, the reason anything read well.
No part of the production chose between the drafts: both were kept and the choice went to a person, who read them once without knowing which system wrote which. One reader, one occasion, no panel and no protocol — and not a blind reading, since they already knew how the work was made and the material was ours. Several questions stay open, and they are the ones we would want answered first. Did these drafts really carry fewer system-specific cues, or did one familiar reader miss cues that blinded readers with stated criteria and a proper baseline could recover? When our operator read the candidates as different writers while unable to assign each to a system, was that a difference of reasoning, or the kind a comparison of exact wording is blind to? Would returning an objection to the original writer preserve more accountable revision than passing the text on, and how would anyone measure that without mistaking revision frequency for improvement? How far can a standard about evidence carry a publication before it needs the sort of guidance a stylebook gives? We can pose those questions; we have not answered them.
The drafts were written by several different commercial systems of the kind already in general use, taken as supplied and not fine-tuned on our material — and nothing here lets a reader verify it. The mapping from system to draft is withheld, so the premise the whole question rests on is taken on our word. The account is deliberately incomplete in a second way: we describe conditions and receipts without publishing the instructions, the arrangement of the work or the operating sequence, which leaves the mechanism describable but not reproducible or checkable from outside — a real evidentiary cost, not a hint of something sophisticated being kept back. And the publication is the author here, which is not the same as the piece having no author. Who wrote any given paragraph of them is a question with an answer on our side of the line; printing it is what we decline to do. Withheld attribution and missing attribution look identical from outside and are not the same thing — coordinated writing can conceal who contributed what, as the rofecoxib publication record showed.
So what would show this reading to be wrong? The comparison we did not run: the same brief and evidence, one version with a rewriting stage over the finished draft and one without, both put to readers with no stake in the outcome and no idea how either was made. If a late pass reliably removes the invented importance and the unearned certainty rather than only the vocabulary, the four failures were reachable from the end after all and this reading is mistaken. A blinded panel recovering system family from these drafts above baseline would damage a different limb. Neither test exists in our record, which is why our confidence stops where it does.
What we are left with is smaller than a thesis and harder to wriggle out of. In four of our own productions, the improvement one interested reader perceived cannot be credited to a finishing pass: the record holds no such stage, and the text that was read was the text its writers finished. Whatever produced what our operator noticed, it was not something added after the writing was over. That is a recorded explanation ruled out rather than a cause discovered, and it is all we hold. The sentence we would most like to write, that what changed was the evidence the writers were given, is the one we have not earned: no version of this work was run without it. Four productions of ours settle nothing about anyone else's, and we do not know the next will go the same way. The entitlement that leaves us is narrow, and we will take it: we know where the difference did not come from, and we know exactly what would have to be done to find out where it did.