The tells that survive into published work

The residue is mechanical before it is stylistic

The most damaging leftovers are not matters of voice. They are placeholders that were never filled, brackets containing an instruction to insert a name, a heading numbered from a template that does not match the sections beneath it, and formatting characters that meant something in one context and appear as stray symbols in another.

These get through because the surrounding text is fluent, and fluency is what a proofreading eye latches onto. A page that reads smoothly is scanned rather than read, and a bracketed placeholder in the fourth paragraph is exactly the kind of thing a scan glides over.

They are also the cheapest problems to eliminate. A search for brackets, for the words that appear in placeholder text, and for formatting characters that should not be present takes under a minute and removes the category of error that most obviously signals nobody looked at this.

Structural sameness across a set

Read one piece and nothing stands out. Read six from the same source and the pattern is unmistakable: the same number of sections, an opening that restates the question, three points of equal length, a closing paragraph that balances two considerations without choosing between them.

Individually each of those is a defensible choice. Collectively they are a signature, and a reader who has encountered the pattern a few times recognises it before they have finished the first screen. The impression is not that the writing is bad but that it was not directed by anyone.

The fix is to vary the shape deliberately. Different lengths, a piece that is mostly one argument, a section that runs long because the material warranted it, an ending that commits to a position. This is editorial work rather than prompting, and it is what a real editor does to a set of submissions anyway.

Phrasing that has become a marker

Certain constructions have become associated with generated text — the balanced pairing that qualifies every claim, the concluding sentence that begins by summarising what the paragraph just said, the tendency to describe something as playing a crucial role rather than saying what it does.

Which specific phrases carry that association changes over time and varies by field, so a fixed banned list ages badly. The durable version of the advice is to look for sentences that could appear in a piece on any subject, because interchangeability is the underlying property that all the individual tells share.

The test is to read a sentence and ask what it would look like with the topic swapped. If it survives the swap unchanged, it was not saying anything about your subject, and cutting it usually improves the paragraph as well as removing the tell.

Facts, names and references need checking specifically

The most serious residue is factual. A reference that does not exist, a person’s title that is out of date, a statistic with no source, a quotation attributed to somebody who did not say it, a link that goes somewhere unrelated. These are ordinary-looking and they do concrete damage when a reader checks one.

Give them their own pass rather than folding it into the proofread, because the mindset is different: proofreading is about the sentence and fact-checking is about the world. Every name, number, date, quotation and link deserves to be confirmed against something outside the document.

This applies with particular force to anything that will be quoted onwards. A wrong figure in an internal note is embarrassing; the same figure in a published piece gets repeated by other people, and correcting it afterwards never reaches everyone who saw it.

A short pass, run every time

The whole thing is a checklist rather than a skill: search for placeholders and stray formatting, check every name and figure against a source, read one paragraph aloud, compare the shape against the last few pieces, and confirm that at least one sentence in the piece could not have appeared anywhere else.

Reading aloud deserves its place because it catches what the eye forgives. Sentences that are grammatically fine and rhythmically identical become obvious the moment they are spoken, and so do the transitions that sound like a form being completed.

None of this makes writing indistinguishable from unaided work, and that is not really the goal. The goal is that the piece shows evidence of a person having made decisions about it, which is what a reader is actually detecting when they say something reads as machine-produced.