Sentence rewriter: what one sentence can and cannot fix

Rewriting sentence by sentence changes less than you would expect, and the reason is that the thing detectors measure lives above the sentence.

A sentence rewriter takes one sentence and reworks it. It can reorder clauses, flip voice, change vocabulary, adjust length. What it cannot do is see the sentence before or after, and that turns out to be most of the problem.

What you get inside one sentence

Four things are genuinely available.

Clause order. "Because the sample was small, the effect is uncertain" and "The effect is uncertain because the sample was small" say the same thing with different emphasis. The first leads with the caveat, the second with the finding.

Voice. Active and passive move where the actor sits. Academic writing conventionally goes passive in methods and active elsewhere, and mixing on purpose reads better than committing to either.

Length. Split a long one, join two short ones. This is the change that matters most, for the reason below.

Vocabulary. Word choice, as long as the technical terms stay put. "Significant" is a statistical claim, not a fancier way of saying "considerable".

Why it moves the score less than you would think

The signal detectors lean on hardest is variation between sentences, not the construction of any single one.

A rewriter handed one sentence at a time cannot see that the previous three all ran to 22 words. So it produces a locally decent sentence and leaves the paragraph's rhythm exactly as flat as it was. Rewrite every sentence individually and you can end up with a paragraph where the vocabulary changed completely and the statistical profile barely twitched.

Same underlying reason synonym substitution fails, which we go through in why paraphrasing tools do not reduce detection.

To change the rhythm, the unit has to be the paragraph. See paragraph rewriting.

When it is the right tool anyway

Three cases where sentence level is exactly what you want.

One sentence is bad. A single awkward construction in a paragraph that is otherwise finished. Rewriting the whole paragraph would disturb work that already landed.

The paragraph carries an order. In a results section the sequence of claims often means something. A paragraph rewriter can reshuffle them. A sentence rewriter cannot, which here is the feature rather than the limitation.

You want a decision at every step. Sentence by sentence gives you a yes or no on each change instead of a wholesale replacement to accept or reject.

Ours works at paragraph level by default because that is where the improvement is, but you can select any passage and rewrite only that. The scope gets outlined before anything is sent, so you always know what is about to change before it changes.

The part that must not move

Everything carrying a value: numbers, units, statistical results, p-values, sample sizes, citations, DOIs, URLs, equation references, defined terms.

Sentence level is where this most often goes wrong, and the reason is structural. A rewriter handed a single sentence has no context telling it that "42" is a sample size rather than a stylistic choice. General tools round numbers, swap statistical terms for vaguer near-synonyms, and mangle citation formatting, all while producing prose that reads perfectly well.

We mark those as protected and refuse any rewrite that would alter one, which is why ours will sometimes hand a sentence back untouched. A sentence that still reads a bit generated is a much smaller problem than a sentence whose numbers no longer match your study.

Checking whether it did anything

Rewriting one sentence barely moves a document score, and a document score is what most detectors report. To see whether an edit achieved anything you need per-passage numbers.