Paragraph rewriting
The paragraph is where rewriting starts working, because it is the smallest unit that can move material between sentences.
A sentence rewriter is stuck inside one sentence. A paragraph rewriter can shift things across sentence boundaries, and almost all of the available improvement lives in exactly that move.
What opens up at paragraph scale
Four operations a sentence rewriter simply cannot reach.
Merging and splitting. Two short sentences become one with a subordinate clause. One long one becomes two. This changes sentence length variation directly, which is the strongest single signal in most detectors and the exact thing sentence-level rewriting leaves alone.
Moving a clause across a boundary. A qualifier hanging off the end of one sentence often belongs at the front of the next. You can only see that with both sentences in view.
Pulling out the connective scaffold. "Moreover", "Furthermore", "Additionally" opening three sentences in a row. Individually each looks fine. Together the pattern is obvious, and removing them improves the prose regardless of who wrote it.
Changing how the paragraph opens and closes. Generated paragraphs love a topic sentence followed by a closing sentence that restates the topic sentence. Cutting the restatement is frequently the single best edit in a draft.
Why sentence length variation matters this much
Models produce sentences of remarkably consistent length. People do not.
You can check this on your own draft without any tool. Count the words in each sentence of a paragraph and look at the spread. Measured across our corpus, the standard deviation of sentence length within a document has a median of 8.7 words for pre-2020 arXiv writing and 6.4 for generated text.
In our detector that standard deviation is one of sixty stylometric features, and it is among the most durable signals in the whole system, because it survives rewriting that changes every individual word. A paragraph rewriter can shift it. A sentence rewriter, blind to its neighbours, mostly cannot. The rest of the markers are in what makes text read as machine written.
Keeping the meaning nailed down
Rewriting technical prose comes with a constraint general rewriting does not have. The claims must survive exactly.
Protected, always:
- numbers, units, percentages, sample sizes
- statistical results and p-values
- citations, DOIs, URLs, equation and figure references
- defined terms, which mean something specific and are not swappable for their synonyms
The failure here is quiet rather than dramatic, which is what makes it dangerous. A tool rounds "42 samples" to "over forty samples". It swaps "significant" for "considerable", turning a statistical claim into a vague one. It renumbers a citation. The prose reads beautifully and your paper now says something else.
Ours treats those as protected and rejects a rewrite that would touch one, so it sometimes hands a paragraph straight back. That is deliberate.
Working through an actual document
Rewriting everything at once is rarely right. You lose track of what changed, and paragraphs differ enormously in how much they need it.
A more controlled pass:
- Score by passage first to find which paragraphs actually carry the signal. Plenty will
not, and rewriting those is wasted effort.
- Start with the worst. Usually the introduction and conclusion, the most templated parts
of most drafts.
- Rewrite one paragraph, re-score. If nothing moved, the problem is structural rather
than lexical.
- Leave the technical sections for last. Methods and results carry the most protected
values and the least stylistic room.
Ours scores each paragraph separately rather than only the document, which is what makes step one possible at all. Free and unlimited.
The honest limit
Rewriting changes how text reads. It does not change what the text is, and no rewrite makes a claim about authorship true or false. A lower score means the prose now looks less like generated writing to one classifier, which is a statement about the prose and nothing else.
Detectors also disagree with each other, so a passage that improves on one may not on another. See why AI detectors disagree.