A rewrite pass, a mechanical checker and a composition read. Editors judging quality alone, with authorship never mentioned, preferred the processed draft 22 times out of 22.
Four studies · ~773 blind judgments · The negative results are published too
A rewrite can make a draft sound better. It cannot demonstrate that it did, because a rewriting pass cannot be its own judge. So the work splits into three, and each one is honest about what it is known to do.
Siqi Chen's MIT-licensed humanizer, vendored. It strips inflated significance, AI vocabulary, negative parallelisms and filler, rewriting rather than deleting.
Carries the measured benefit — 16 of 16A standard-library sweep over the trope inventory. It reports FAIL and WARN counts, readability, anaphora runs and duplicate sentences, and it never edits.
Not proven to help — 12 of 21, chanceFive tells stay manual forever, because no regular expression catches them without lying about its precision: dead metaphor, fractal summaries, analogy stacking, one-point dilution, near-verbatim repetition.
Where the author has to show upEditors were shown two passages and asked one question: which is the better piece of writing? Every register, every length.
| Comparison | Preferred |
|---|---|
| raw draft vs fully processed | 22 / 22 |
| raw draft vs rewrite pass alone | 16 / 16 |
| rewrite pass vs full pipeline | 12 of 21 |
| fully processed vs real human prose | 18 / 18 |
Row three is the uncomfortable one. The checker I spent the most time building emitted its input byte-identical in 11 of 18 cells, so twelve of those twenty-one trials were judges comparing a file to itself. It is shipped because a reporting pass that never edits is cheap and feeds stage three, not because a measurement says it improves prose.
Row four is the strangest. Judged on quality alone the processed machine prose also beat genuine human writing, by 22 rubric points — the same passages a detector called human every single time. Reads as human and is good writing behaved as separable properties.
This started as an attempt to beat detection. Three studies say it cannot be done this way, and they are published here rather than buried.
Against a design that could not resolve anything under about four points. Genuine human prose was never once misidentified, 143 times out of 143.
An explicit ruleset built from what judges kept citing fooled nobody. Neither did telling the model to simply try harder: 0 of 15.
A 2,940-word essay was caught whole and caught in every one of six standalone slices.
The reason is worth carrying into your own work: the pipeline crushes the tells it encodes and grows the ones it does not. AI vocabulary fell from 70 judge citations to zero while uniform beat rate rose from 60 to 101. A subtractive rewrite makes prose more uniformly well-shaped, and uniform good shape is itself the tell.
Read the full write-up ▸Swept over 107,191 words of public-domain literature, US government works and IETF RFCs, the checker raises 8 failures — 0.07 per thousand words — and blocks 4 of the 24. Raw AI drafts score 1.02 failures per thousand against 0.00 for human prose across two independent corpora.
That gate runs in CI. Rules were demoted against evidence rather than taste: typography defaults to a warning because two rules produced 94.6% of all failures on human prose, and negative parallelism became a counted warning after 10 of its 17 hits turned out to be the correlative form as written by Thoreau, Du Bois, Russell and Strunk.
Its rules encode one person's taste in prose. They do not identify machines.
The results that went against it are published beside the one that did not.
Four studies · Every corpus and judgment file committed
Every number on this page is reproducible from what is committed: the corpora, the judgment files, the tally scripts, and the two figures I had to correct mid-study.
github.com/itsryanlenk/humanist ▸MIT for this repository's own work · the vendored pattern catalog additionally carries CC BY-SA 4.0