Our method
Understand what each reading measures and how rewrites are checked. Back to the humanizer
Human writing gets flagged. Here is why.
“It is interesting to contemplate an entangled bank, clothed with many plants of many kinds, with birds singing on the bushes, with various insects flitting about, and with worms crawling through the damp earth, and to reflect that these elaborately constructed forms, so different from each other, and dependent on each other in so complex a manner, have all been produced by laws acting around us.”
Sentences average 55 words with a variation of 0.28. Sentence lengths are similar; consider whether the rhythm serves the ideas.
This was written 160 years before the first language model. It still lands in mixed signals, because its sentences are uniformly long and it uses no contractions. Careful formal writing looks like that. So does fluent second-language English.
Detectors measure style, not authorship. That is why honest writing gets accused — and why the reason matters more than the score. A reason is something you can check, explain to a supervisor, and change if you choose to.
Load the Darwin sample in the scanner above to see every sentence marked up.
How the local style index works.
Seven measures of style, weighted and combined. No hidden model, no proprietary secret — this is the whole method. Signals that do not apply to a passage (paragraph evenness needs three paragraphs) drop out and the rest are reweighted.
| Signal | What it measures | Weight |
|---|---|---|
| Sentence-length variation | How much sentence lengths vary. People swing between short and long; models cluster near an average. | 28% |
| LLM-favoured phrasing | Phrases heavily overrepresented in model output: “it is important to note”, “furthermore”, “leverage”. | 24% |
| Contractions | Whether “it's” and “don't” appear. Models default to the formal “it is” and “do not”. | 14% |
| Hedging language | Density of qualifiers such as “generally”, “typically”, “can be”. | 12% |
| Em-dash density | Em-dash frequency. A well-known habit of chat models at high density. | 10% |
| Rule-of-three lists | Rule-of-three lists like “faster, cheaper, and more reliable”. | 7% |
| Paragraph uniformity | How evenly sized the paragraphs are, when there are three or more. | 5% |
Several drafts. One reviewed result.
- 1Your text is measured across the seven signals above.
- 2Three drafts are written in parallel — by three different frontier models on paid plans, or three DeepSeek drafts on the free tier.
- 3Every draft is measured the same way. Any draft that lost a figure is rejected; the lowest-exposure survivor is kept.
Draft selection uses the local style index after preservation checks. It is not a benchmark of semantic fidelity or external detector performance. We have not established superiority over other writing tools.
Refine the wording.
Keep the details in view.
Rewriting tools are known for quietly altering a statistic or dropping a citation. Before every rewrite we extract literal figures, supported date formats, quotations and references. We check their occurrences in each draft, along with any sentences you locked. A draft that fails these checks is withheld and its words are returned.
Names are flagged for your review rather than enforced, because a faithful rewrite can turn “municipalities in Denmark” into “Danish municipalities” and no string check can tell that from a genuine loss.
It is important to note that the The team shipped 42 updates in 2024. The release notes describe call the result as "a steadier workflow".
What the results can tell you.
Local style index
A transparent heuristic from 0 to 100. It describes patterns in English prose and highlights places to review. It has not been calibrated as an authorship probability or an external-detector predictor.
Self-hosted AI check
Our fine-tuned detector runs on our servers. When calibration is available, it estimates AI likelihood for the whole passage using held-out examples. This is not a percentage of AI-written words. Sentence highlights are review cues, and results can be wrong. An unavailable detector produces no score.
Neither result establishes authorship, and a lower style index does not prove better writing. Review meaning, sources and voice before using a rewrite.