Readers who see a lot of generated text begin recognising it quickly, and usually cannot say what gave it away. The tell is structural rather than lexical, and it follows directly from how the text is produced.
Human prose varies more than it appears
Writing by a person contains sentences of markedly different lengths, including very short ones placed for effect.
Paragraphs vary too, some carrying a single thought and others accumulating several before turning.
That unevenness reflects a writer's attention shifting, and it is the texture readers are accustomed to without noticing.
Likely continuations are moderate ones
A model selecting probable next words tends toward the middle of every distribution, including sentence length and clause structure.
The result is prose where most sentences fall within a narrow band and few take a risk in either direction.
Nothing about any individual sentence is wrong, and the aggregate reads as level in a way that sustained human writing rarely is.
The structural habits are strong
Generated text favours certain shapes: the three-part list, the sentence that qualifies itself with a subordinate clause, the paragraph that opens with a claim and closes by softening it.
These appear frequently in the training material and are reinforced by the preference training that follows, which rewards balanced and comprehensive-seeming answers.
Repetition of a shape is more noticeable over a page than repetition of a word, which is why editing individual phrases does not remove the impression.
Openings and closings are the most formulaic
The first and last paragraphs carry the strongest signature, because they are the most conventional parts of any piece.
Restating the question before answering it, and summarising what was just said, are habits that human editors remove and models produce by default.
Cutting the first and last paragraph outright improves generated prose more reliably than any other single edit.
Editing has to attack the shape
Substituting synonyms leaves the rhythm intact, which is why heavily reworded generated text still reads as generated.
What changes it is structural: splitting a long sentence, cutting a hedge, deleting the transitional clause between paragraphs, letting a section end without rounding off.
Those edits are also what make the writing better, so the work of removing the signature and the work of improving the prose turn out to be the same work.