Why good writing gets flagged more
The properties that make prose read as polished are the same properties that make it read as machine-made. This is the central problem with detection and nobody puts it on a marketing page.
Everything you were taught about writing well pushes your text toward the profile detectors were built to catch. Not as an unfortunate coincidence, as a direct consequence of what they measure.
Worth understanding, because it explains a result that otherwise makes no sense: the most careful writer in a cohort is often the one with the highest score.
What detectors measure
Two things.
How predictable your word choice is. Given the words so far, how surprising is the next one. Generated text scores low here, because models are trained to pick likely continuations.
How much that varies across the document. Human writing spikes and dips. Generated writing runs flat.
Both are measures of regularity. Neither has any way to know how the text was made.
The advice that raises your score
"Write clearly and conventionally." Conventional phrasing is predictable phrasing. Choosing the standard construction is choosing the one a model would also choose.
"Use signposting." "Moreover", "Furthermore", "Additionally", "Overall". These are taught as good structure. They open 2.4% of human sentences in our corpus and 5.8% of generated ones, more than double.
"Keep your sentences a manageable length." Read as: make them all similar. Human academic writing in our corpus varies sentence length with a standard deviation of 8.7 words, generated text with 6.4. Even rhythm is the single strongest signal there is.
"Revise it again." Every editing pass smooths. Smoothing removes variation. A fifth draft sits closer to the generated end than a first draft does.
"Cut the digressions." Digressions are irregularity. Trimming a piece to its argument removes the unevenness that reads as a person thinking.
Follow all five and you have produced exactly the profile a classifier flags.
Who this hits hardest
Careful writers, for the reasons above.
Non-native English speakers, because learned English is applied more consistently and reaches for the common construction more often. Published evaluations keep finding elevated rates. See why non-native English writing gets flagged.
Anyone writing in a technical register, where field convention constrains word choice hard enough that predictability is not a choice.
Anyone with a good editor, since the editor's job is to smooth.
What to do with this
The fixes run against instinct, and none of them make the writing worse.
Vary sentence length on purpose. Put a four-word sentence next to a thirty-word one. It reads better as well as scoring better, because rhythm is what carries a reader.
Delete most connective openings. If the logical relationship is real, sentence order already carries it. If it is not, the connective was covering for that.
Keep the specific detail. Name the method, give the number, use the real example. This is the thing most often cut while tightening a draft and it is the strongest human signal there is.
Leave the asymmetry in. Let the point you actually care about take more room than its logical weight deserves. That is what a person does and what a model does not.
Full list in what makes text read as machine written.
Ours is free and unlimited with no account, scores every paragraph separately so you can see which of your best-written passages sit highest, and never stores your text. Wrong-flag rate 0.4%, published, against Turnitin's 4%.