How PlagiatX checks texts: methodology and limitations

To trust a result you need to know how it was produced. This page openly describes the check stages, the sources used for comparison, the score formula, quality control and the system's limitations.

Check a text →

1. Preparing the text

Text is extracted from PDF, DOCX or TXT (scanned image PDFs are not supported yet). The text is split into sentences and each one is scored separately, so results are shown sentence by sentence: plagiarism in red, likely AI-written text in yellow.

2. What it is compared with

An academic corpus of more than 200,000 documents (Uzbek, Russian and English) plus open web search: encyclopedias, research-paper databases and the general web. Verbatim copying, paraphrase with replaced words and text translated from another language are each searched with a different method (text fingerprints, lexical matching and multilingual semantic vectors).

3. Detecting AI-written text

Several signals are combined: how "expected" the text is for a language model (perplexity), a multilingual classifier and stylistic features (uniform sentence length, formulaic phrases). The result is a probability, not a verdict: human text sometimes scores high too, so a final conclusion should be made after talking to the author.

4. How the score is calculated

Originality = 100 − (P + A − P×A/100), where P is the plagiarism percentage and A the AI percentage. The formula is the same on the site, in the PDF report and in the QR certificate. In AI-only or plagiarism-only mode only that measure is used.

5. How quality is controlled

Every change is tested on 33 control cases before release (13 Uzbek, 10 Russian, 10 English): verbatim copies, light edits, original and human-written texts, AI-written texts and edge cases. If results get worse, the change is not released. Real errors reported by users with the "Result wrong?" button (with consent) are added to this set.

6. Privacy

A checked text never becomes a source for others and is shown to no one — unless the author explicitly consents via "Add my work to the database". Results are deleted after 30 days (24 hours in private mode); the certificate's QR verification keeps working with the numbers only, without the text.

7. Limitations — stated openly

No plagiarism checker can see every text on the internet and in closed university archives — an unfound source does not mean "the text is original". The AI score is a probability. Scanned PDFs cannot be read. A PlagiatX certificate is a digital confirmation of a check result; it is not an official document of the Higher Attestation Commission or a university, and each university sets its own required percentage.

Questions

What if the result is wrong?

Report it with the "Result wrong?" button on the result page. Every report is reviewed and, with consent, added to the quality test set.

How can a certificate be verified?

Scan the QR code or enter the number at plagiatx.uz/verify — the check date and percentages are shown.

Plagiarism checker · AI detector (ChatGPT) · Originality score · Methodology · Dissertation check · Thesis review · Journal article readiness · Dissertation requirements · Master's · PhD and DSc · GrantGlobus — grants · Motivation letter · CV for grants · Lebenslauf (German CV) · Blog · Terms of offer · Privacy · Telegram bot