Why paraphrasing tools do not reduce detection

Synonym substitution changes the words a detector reads but not the properties it measures, and in several cases it makes a text score higher than the original did.

The intuition is reasonable: a detector reads words, so change the words. It does not work, and understanding why is useful whichever side of this you are on.

Detectors do not primarily read words

The signals that carry most of the weight are structural:

Sentence length variation. How much the length of consecutive sentences differs.
Punctuation and clause rhythm. Comma density, subordinate clause frequency,

paragraph shape.

Function word rates. How often the, of, and, to and similar appear, which is

a property of syntax rather than subject matter.

Predictability. How expected each word is given what came before.

A thesaurus pass changes none of the first three. Swapping "demonstrate" for "illustrate" leaves the sentence exactly as long, with the same clause order and the same function words holding it together.

In our own detector, sixty stylometric features and a lexical model both feed the final score. Substitution moves the lexical component a little. It barely touches the stylometric one, and the stylometric signal is the one that generalises best across models and topics.

Why substitution can raise the score

Sometimes a paraphrased text scores higher than the text it came from.

Register flattens. Thesaurus output reaches for the more formal option almost every time. Formal, uniform vocabulary is closer to the generated end of the distribution, not further from it.

Collocations break. Language runs on conventional pairings. "Conduct research" is idiomatic; "administer research" is not. Unusual pairings are unpredictable, and enough of them makes text read as machine-processed rather than as written by a person.

Variance drops. Automated substitution applies the same transformation everywhere, which makes prose more uniform. Uniformity is the signal.

What does change a score

Editing that changes structure rather than vocabulary:

merging two sentences, or splitting one
cutting a leading connective
reducing stacked hedges to a single qualifier
varying how paragraphs open, rather than opening each the same way
removing a closing paragraph that only restates

These change sentence length variance, connective density and rhythm, which are the properties actually being measured. We set out the specific markers in what makes text read as machine written, and what a tool has to do to move them in AI rewriter: what actually changes.

The honest limit

No rewriting reliably clears every detector, and anyone claiming otherwise is describing a moving target as if it were fixed. Detectors are retrained, thresholds shift, and a text that scores low on one tool can score high on another for structural reasons we cover in why AI detectors disagree.

What editing does reliably do is make prose read better: shorter, less hedged, less signposted, more varied. That improvement is real whether or not any detector is involved.

What breaks in technical text

Paraphrasing tools applied to academic or technical writing have a failure mode general prose does not: they alter values.

Numbers get rounded or reworded. "42 samples" becomes "over forty samples". Statistical terms get swapped for near-synonyms that mean something else, so "significant" becomes "considerable" and a statistical claim silently becomes a vague one. Citations get renumbered, and DOIs and URLs get mangled.

Our rewriter treats those as protected and refuses a rewrite that would change one, which is why it sometimes hands a paragraph back untouched. A paragraph that still reads as generated is a smaller problem than a paragraph whose numbers no longer match the study.