How to make text sound less like a machine wrote it
Five changes, each of them measurable, each of them something you can do without altering what the text says.
The advice on this topic is usually vague: "add personality", "write naturally". Detectors do not measure personality. They measure a small number of concrete properties, and those properties are the ones worth changing.
Here is what they actually read, with the numbers from our own corpus, and what each one means in practice.
1. Vary sentence length, deliberately
This is the strongest single signal in most detectors, and the easiest to fix.
Human writing in our corpus has a within-document sentence length standard deviation of 8.7 words. Generated text sits at 6.4. Generated prose is not too long or too short, it is too even.
The fix is mechanical. Find a run of three sentences of similar length and split one, then merge another pair. A four-word sentence next to a thirty-word one does more for this number than any amount of word substitution.
2. Cut connective openings
Sentences that begin "Moreover", "Furthermore", "Additionally", "Overall", "Consequently".
These open 2.4% of human sentences and 5.8% of generated ones in our corpus. More than double, and they are visible to a reader as well as a classifier.
Most of them can simply be deleted. If the logical relationship is real, the sentence order already carries it, and if it is not real, the connective was covering for that.
3. Keep the specific detail
Generated prose generalises. It reaches for the category term where a person would name the thing. "Various approaches have been proposed" instead of naming two of them. "Significant improvements" instead of the number.
Specificity is the strongest human signal there is, and it is the one most often edited out in the name of concision. Put the number back. Name the method. Give the actual example.
4. Break the paragraph template
Generated paragraphs tend to run to a shape: topic sentence, two or three supporting sentences, a closing sentence that restates the opening. Repeated down the page, that shape is visible.
Let a paragraph be two sentences. Let another run long. Start one in the middle of the argument without announcing it.
5. Leave the asymmetry in
Real writing is uneven. Some points get four sentences and some get half of one. The thing you care about most gets more room than its logical weight deserves, because you care about it.
Smoothing that out is what editing passes do, which is why heavily edited human writing scores higher than a first draft. Every revision pass pulls toward the machine end of the scale. See does Grammarly get flagged as AI.
What does not work
Synonym substitution. Swapping words for less common ones changes the vocabulary without changing the rhythm, and the rhythm is what most of the signal lives in. It also degrades the prose. Tools built purely on synonym swapping are the ones detectors handle most easily. See do paraphrasing tools beat AI detection.
Adding errors on purpose. Detectors do not score correctness. Typos cost you a reader and buy nothing.
Unusual punctuation. Same reason. It is not what is being measured.
Checking the result
Score it per passage rather than as one document number, because the changes above are local and a document average hides which paragraph is still flat.
Ours is free and unlimited with no account and marks each paragraph separately, so you can see the flat ones and stop when they are gone. The full list of markers is in what makes text read as machine written.