Why are you paraphrasing?
A paraphrase should have a reason to exist. Maybe the original is too dense for its audience, repeats itself, or is more formal or casual than where it's headed. Name that problem first. "Make this score lower on a detector" says nothing about whether readers will follow the revision or whether the facts survived.
A natural style sounds more conversational, a clear style trims needless complexity, and a formal style suits professional settings. None of them should alter the underlying claim. If all you want is cosmetic variation, ask whether the original already works.
What has to survive the rewrite?
Before rewriting, list the facts and relationships that matter: people, organizations, dates, measurements, causes, qualifications and citations. Note which phrases are direct quotations and which are your own explanation. Mark technical terms that shouldn't be swapped for approximate everyday words.
For a short paragraph that's a few notes. For a long document, keep a structured review next to the draft. The point is to judge the output independently of how smooth it sounds, because a revision that drops an inconvenient caveat can read better and be less accurate.
How do negation and hedging get lost?
Compare "the test did not establish an improvement" with "the test established no improvement." The first can mean the evidence was insufficient. The second sounds like evidence of absence. "May reduce costs" and "reduces costs" split the same way, and changes like these hide easily inside an otherwise natural sentence.
Antonyms are especially unsafe as an automatic technique, since reversing one word can reverse the conclusion. Synonyms need context too: a scientific term may have a precise definition, and near-synonyms differ in intensity or implication. Read the whole claim, not just the individual replacements.
How do you check numbers after paraphrasing?
Compare every number with its unit and what it refers to. Keeping "25" while changing "percent" to "people" is still an error. A date can keep its digits and land on the wrong event. A range can lose an endpoint, and a decimal can get rounded with no justification.
The paraphraser flags numeric strings that were added or removed. Treat that as a prompt to review rather than a full consistency test. It can't tell whether a number still refers to the same thing, and numbers written as words, changed units and alternate date formats need a human eye.
How do you protect quotations and citations?
A direct quotation stays exact unless you deliberately change how it's presented and follow the relevant quotation convention. Paraphrasing a quoted passage while leaving the quotation marks on it misrepresents the source. A citation should keep supporting the claim right around it.
If you revise a passage that summarizes someone else's work, compare the new version with the original source. Keeping the citation doesn't make the rewritten statement accurate. Models can strengthen a claim, merge two findings or add a detail the cited work never contained, and you should never accept a newly generated reference without checking it.
What's the best way to review a paraphrase?
Start with meaning. Put the original and the revision side by side and look for omissions, additions, reversals and shifted relationships. Ask whether a reader would decide differently after the new version, and restore any qualification that affects interpretation.
Then look at expression. Cut redundancy, pick the right tone and check that the paragraph flows into what surrounds it. A model only sees the passage you gave it, so its revision may repeat something said earlier or refer awkwardly to material it never saw.
If the output is truncated or incomplete, don't stitch it into the document without a fresh review. The original stays your reference point, and a failed rewrite should never overwrite it.
Does paraphrasing change who wrote it?
Editing generated text doesn't erase its writing history, and a model-assisted paraphrase stays part of that history. Follow the attribution, disclosure and originality rules of your setting, and only claim things about your process that you can support.
A detector may react differently after an edit, but that's an observation about the detector. It doesn't prove the revision is human-written or that every possible watermark is gone. If you're studying robustness, keep the original and revised text, describe the transformation and record the exact detector configuration. If you're just improving a draft, ask whether the revision is accurate and useful.
What should the final checklist include?
Before you use a revised passage:
- Verify names and dates.
- Compare every number with its unit.
- Check negation and uncertainty.
- Keep direct quotations exact.
- Confirm each citation still supports its claim.
- Read the passage in its surrounding context.
Then ask whether the rewrite solved the problem you named at the start. If it only sounds different, it isn't an improvement.