AI Content Detector & Plagiarism Checker
Two independent checks. AI detection estimates how closely text matches statistical patterns typical of machine-generated writing. Reference comparison finds passages copied from a source text you paste in. Neither check is proof of misconduct, and this tool does not crawl the live web.
AI content detector and plagiarism checker
Minimum 40 words. Best results with 150 words or more of continuous prose. Lists, code, poetry and text that is mostly non-Latin script are not scored.
Ready. Paste at least 40 words to begin.
Analysis
Confidence: very lowSignal breakdown
| Signal | Measured value | AI signal | Score |
|---|
Each signal is scored from 0 (human-typical) to 100 (AI-typical). The overall estimate is a weighted average, pulled toward 50 when the signals disagree.
Ready. Paste both texts to compare.
Similarity
Matched passages in Text A
Common phrases
How to use this tool
- Choose a check type. AI detection analyses a single text for machine-like statistical patterns. Reference comparison checks one text against another.
- AI detection: paste at least 40 words and click Analyse text. The tool scores seven independent signals and combines them into one estimate. It refuses to score very short text because the result would be meaningless.
- Reference comparison: paste the text you are checking in Text A and the source in Text B. The tool finds five-word sequences that appear in both, highlights them, and reports the share of Text A that overlaps.
- Read the confidence rating. A score is only as useful as the confidence attached to it. Short text, mixed scripts, lists and code reduce confidence.
Both checks run entirely in the browser. Nothing is uploaded, and nothing is stored after you close the tab.
AI detection and plagiarism checking are different tools
The phrase "AI plagiarism checker" combines two unrelated checks. Plagiarism detection asks whether text already exists elsewhere: it compares a document against databases, academic journals and the open web, and reports matching passages[reference:5]. AI detection asks whether text was drafted by a model: it looks at internal writing patterns and gives a probability score based on how closely the text matches AI-like behaviour[reference:6].
They fail in opposite directions. A plagiarism checker will not flag AI text that is original wording. An AI detector will not flag human-written text copied from another source. A tool that claims to do both in one click is usually strong at one and weak at the other[reference:7].
This page keeps them separate. AI detection is a client-side statistical estimate. Reference comparison is a client-side text-matching check. Neither one crawls the web, and the page says so instead of implying otherwise.
What the AI detection signals measure
No client-side tool can reproduce a large detector's training data. What it can do is measure linguistic properties that correlate with machine-generated text in independent studies. This tool uses seven signals, each scored independently so you can see which one drove the result.
| Signal | What it measures | Why it correlates with AI text |
|---|---|---|
| Vocabulary diversity | Unique words divided by total words | Language models tend to reuse a narrower vocabulary and repeat phrases more than human writers. |
| Sentence-length variation | Coefficient of variation of sentence lengths | Human prose mixes short and long sentences; machine output is often more uniform. |
| Contraction rate | Contractions per 100 words | Models default to a formal register and use fewer contractions such as don't or it's. |
| Punctuation variety | Distinct punctuation types used | Human writing uses dashes, parentheses and semicolons more erratically; model output clusters around commas and periods. |
| Three-word repetition | Repeated three-word sequences | Models loop on stock phrases more than people do. |
| Sentence-opener diversity | Unique first words divided by sentence count | Models reuse openers such as Additionally, Moreover and Furthermore. |
| Rare-word rate | Words appearing exactly once | Human writing introduces more one-off vocabulary than model output. |
Each signal is converted to a 0–100 score using published ranges for human and machine text. The overall estimate is a weighted average. When the signals disagree strongly, the result is pulled toward 50 rather than presented as a confident verdict — a technique used by open-source detectors to avoid the overconfidence that causes false accusations[reference:8].
How reference comparison works
Reference comparison splits both texts into overlapping five-word sequences, called shingles, and finds sequences that appear in both. This is the same fingerprinting approach used in plagiarism-detection research, simplified for client-side execution.
The output is a word-overlap percentage and a highlighted view of Text A, with every matched passage marked. The percentage is not a plagiarism verdict: quoting a source with attribution produces overlap, and so does a common technical phrase. The highlighted passages are the useful output, because they show exactly what matched so you can judge whether the overlap is legitimate.
Reference comparison is also the honest way to check self-plagiarism: paste your earlier draft as Text B and the current draft as Text A. The tool has no database and no persistence, so it can only compare what you give it in this session.
Why this tool does not scan the live web
A web-wide plagiarism scan requires a crawler that has indexed billions of pages, a comparison engine, and a paid API or server to run it. Services that offer this — Turnitin, Copyscape, Quetext — charge for it because the database and infrastructure are the product. A static page served from Cloudflare cannot crawl the web, and a free browser tool that claims to do so is either proxying to a paid service without saying so or fabricating a result.
Rewind.ai, a competitor with a similar client-side model, discloses the same limitation directly: "we do not crawl the live web" and recommends Turnitin for graded academic work and Copyscape or Quetext for paid web-wide scans[reference:9]. This page takes the same position because it is the accurate one.
Accuracy, false positives and responsible use
Independent comparisons of commercial detectors report a wide accuracy range: Turnitin 95–98%, Originality.ai 97%, GPTZero 87%, Copyleaks 80–100%, QuillBot 78–84%[reference:10]. Every one of those tools produces false positives, and the risk is not evenly distributed. Multilingual and neurodivergent writers face higher false-positive rates, which is why several universities warn against using detector output as disciplinary evidence on its own[reference:11].
This tool inherits those limits and adds its own, because a seven-signal heuristic is less accurate than a trained classifier. Treat a high score as a reason to review the text more carefully, not as a finding. Never use the output to accuse someone. Never submit it as evidence. If a result matters, verify it with a system that has a documented accuracy record and a human review process.
Limitations to know
- Minimum 40 words. Below that, the tool refuses to score. A confident result on a single sentence would be misleading.
- English prose only. Signals are calibrated for English. Text that is mostly non-Latin script, code, lists or poetry is not scored reliably.
- Paraphrased AI text defeats the signals. Heavily edited or humanised model output often looks human on every signal. This is a known failure mode of statistical detectors, not a bug.
- Formal human writing scores higher. Academic prose, legal writing and technical documentation use fewer contractions and more uniform sentences, which raises the AI estimate even when a person wrote every word.
- No web database. Reference comparison only sees Text A and Text B. It cannot tell you whether a passage came from a third source neither text contains.
- Five-word matching is literal. Synonym substitution, reordering and translation evade it. That is a property of shingle matching, not a defect that can be fixed client-side.
- No persistence. Nothing is saved. Closing the tab clears everything.
Frequently asked questions
Is this an AI plagiarism checker?
It is two separate checks in one page: an AI content detector and a reference-text comparator. It does not combine them into a single verdict, because they answer different questions and fail in different ways.
Does it scan the internet for plagiarism?
No. It has no web database and cannot crawl the live web. It compares the text you provide against a reference text you also provide. For a web-wide scan, use a service that indexes the web, such as Copyscape, Quetext or Turnitin.
How accurate is the AI detection?
It is a statistical estimate, not a measurement. Commercial detectors reach roughly 80–90% on raw text and degrade on paraphrased or mixed writing. A seven-signal client-side heuristic is less accurate than a trained classifier. Use it as a first-pass signal and read the confidence rating before drawing any conclusion.
Why does the tool refuse to score short text?
Linguistic signals need enough text to be meaningful. Below 40 words, the estimate would be noise presented as a result. Open-source detectors impose the same floor for the same reason.
Can it detect ChatGPT, Claude or Gemini output?
It can detect patterns that are common in model output, but it cannot attribute text to a specific model. No client-side tool can. Claims of model-specific attribution require a trained classifier and a large labelled dataset.
What if my human-written text scores as AI?
Formal, technical or academic writing often does. Fewer contractions, uniform sentence lengths and a narrow vocabulary all raise the estimate. If you wrote the text yourself, the score is wrong — that is a known limitation of statistical detection, not evidence against you.
Can I use this to check my own earlier draft for self-plagiarism?
Yes. Paste the earlier draft as Text B and the current draft as Text A. The comparison will show any passages you reused. The tool does not store either text, so you need to paste them again each session.
Is my text uploaded or stored?
No. Both checks run in the browser. No request carries your text anywhere. Closing or reloading the tab clears everything.
Can I use the result in an academic or workplace dispute?
No. The output is a statistical estimate with a known false-positive rate. It is not evidence. Academic and workplace processes require a documented system with human review; a browser heuristic does not meet that standard.
Does it support languages other than English?
The AI signals are calibrated for English. Reference comparison works on any language that uses spaces between words, but the AI score should be ignored for non-English text.