Paraphrasing detector — find disguised copying
A paraphrasing detector finds text that has been reworded to disguise its source. Noplag catches the light end of that today: edits that leave five-word runs intact still match, so synonym swaps and minor reordering are found by winnowing fingerprints and cross-checked against live web search. Thorough sentence-by-sentence rewriting is a different problem and we do not detect it yet — near-duplicate matching (v1.1) and semantic matching (v1.2) are what solve it, and they ship when they clear published evaluation.
We detect. We don't help evade.
A humanizer is a tool that rewrites AI- or human-written text specifically to defeat detection. We could build one. We don't. The whole point of this product is to surface unattributed copying, not to help hide it. If detection is going to be useful at all, the people running detectors and the people building humanizers can't be the same vendor. We've picked our side.
When we mention humanizers in the report, it's to tell you they exist and may have been used upstream of the text you're checking — not to sell you one.
Four layers. Two shipped, two on the way.
The cascade is designed to narrow from exact-match through near-duplicate to semantic equivalence, then cross-check against the open web. Two of those layers ship in v1.0; the two that handle real paraphrase are still being built, and we would rather say so than imply otherwise.
Winnowing fingerprints — shipped
Exact substring matches. Light synonym swaps and minor reordering that leave five-word runs intact.
Anything reordered or rephrased.
MinHash LSH — v1.1 roadmap
Will catch sentence reordering and word-level reshuffling. Not detected today.
Real synonyms and semantic rewrites, which is what L3 is for.
Sentence embeddings — v1.2 roadmap
Will catch synonym substitution and semantic paraphrasing. Not detected today.
Rewrites far enough from the source that no statistical signal survives.
Live web cross-check — shipped
Sources surfaced after the corpus snapshot.
Paywalled or unindexed content.
What we catch today — by paraphrasing intensity.
This is v1.0 behaviour, not a projection. Light editing still leaves matchable runs of text; thorough rewriting does not, and no amount of fingerprinting recovers it. We wrote about exactly where that line falls in our post on the honest limits of verbatim detection: /blog/the-honest-limits-of-verbatim-detection.
Synonym swaps and minor reordering. Enough of the original wording survives for fingerprints to match.
The cat sat on the mat.
The feline sat on the mat.
Sentence restructured, verbs changed. Caught only where an untouched run survives. Near-duplicate matching (v1.1) targets this tier.
The cat sat on the mat.
On the mat, the cat had settled.
Complete semantic rewrite, new structure and vocabulary. Scores zero today; semantic matching (v1.2) targets this tier.
The cat sat on the mat.
A small mammal had positioned itself on a fabric surface.
Ten questions about paraphrasing detection.
Mostly about where detection breaks down. We're explicit about the floor; the people who need to know are the people about to act on a score.
- What's the difference between detecting paraphrasing and detecting plagiarism?
- Plagiarism is the broader category — copying text without attribution. Paraphrasing detection is the specific challenge of catching plagiarism after the text has been rewritten to disguise its origin. Noplag does both in one report; the paraphrasing pipeline is what kicks in when the L1 exact-match layer doesn't fire.
- Why is heavy paraphrasing so hard to catch?
- Detection works by finding statistical signals — shared runs of words, and later shared structure and meaning. Light editing leaves those runs intact, so it is caught. Rewrite a passage sentence by sentence and the shared runs disappear entirely: a thorough paraphrase scores zero against a fingerprint engine, which is what v1.0 is. That is a property of the method, not a bug, and it is why near-duplicate (v1.1) and semantic (v1.2) matching are separate layers rather than a tuning knob. We wrote this up in our post on the honest limits of verbatim detection: /blog/the-honest-limits-of-verbatim-detection.
- Does Noplag offer a tool to humanize AI text?
- No. We don't ship a humanizer. A humanizer's job is to defeat detection — that's anti-aligned with our mission. We're on the integrity side; we detect, we don't help hide. See the position callout above.
- How does this compare to Quillbot's paraphrasing tool?
- Quillbot rewrites text TO defeat detection. We detect text rewritten to escape detection. We're on opposite ends of the same problem. If a Quillbot output ends up in your check, the report will flag 'possible paraphrasing-tool use' when we see the signature.
- What's a 'four-layer cascade', and how much of it is live?
- Four detection passes, each less literal than the last. L1 winnowing catches exact and lightly edited matches, and Layer W cross-checks against live web search — both ship in v1.0. L2 (near-duplicate matching) is on the v1.1 roadmap and L3 (semantic matching via sentence embeddings) on v1.2; neither is running today, and neither is in the public repo yet. So the honest summary is: half the cascade is live, and it is the half that handles light rewording.
- What about back-translation as a paraphrasing technique?
- Back-translation (English → Spanish → English) is among the hardest paraphrasing types — semantically equivalent but structurally novel. Where it leaves runs of original wording intact we catch it; where it rewrites cleanly we do not detect it today. It is a target for the semantic layer in v1.2.
- Will you flag heavy paraphrasing as 'plagiarism'?
- Only if confidence is high enough. Below threshold we surface the suspect passage with 'low confidence match' and let the reviewer judge. We never declare paraphrased plagiarism unilaterally — we report a likelihood and the reasoning.
- Can I see how Noplag scored a passage?
- Yes. Every match in the report links to the source URL and shows which detection layer fired. You can replicate the score offline via the open-source engine — see /open-source-plagiarism-checker.
- Does Noplag work for non-English paraphrasing?
- Detection is calibrated for English, Spanish, Portuguese, Polish, and Ukrainian at launch, with more languages rolling out. Other languages return results flagged 'low-resource — interpret with extra caution.'
- Is the paraphrasing detector free?
- Yes on the free tier — 2,500 words per check, no signup. Pro and Premium tiers raise the limit; the underlying detection algorithm is the same on every tier.
Find paraphrased plagiarism — free, up to 2,500 words.
We catch light rewording today and say plainly that thorough rewriting escapes us until the semantic layers ship. Run a check on your own text and read what the engine actually found.
Find paraphrased plagiarism