Methodology

This page describes what each mode measures, how the figures in a report are produced, where the method is known to fail, and what the service deliberately does not do. The same method is applied to every text, whether it is an essay, an article or a manuscript; the audience pages only change the examples and the wording.

1. Plagiarism: overlap with sources

The submitted text is split into overlapping word sequences. These are compared with records from scholarly databases (Crossref and OpenAlex metadata, open-access full texts, arXiv) and with web pages retrieved for the candidate passages. Where a sequence matches a record, the passage is marked and the record is listed as a source with the share of the submitted text it covers.

• Overlap is the share of words that fall inside matched passages, after sources covering less than 0.5 % of the text are set aside. The number of such minor sources is stated in the report.
• Quotations: text inside quotation marks or accompanied by a citation marker is reported separately when the engine recognises it as such. The recognition is heuristic and misses some citation styles.
• Reference list: the bibliography section is excluded from overlap counting when the engine detects it, so matching reference entries do not inflate the overlap figure.

An overlap figure is not a measure of misconduct. Common phrases, assignment wording, methods descriptions, standard definitions and correctly cited quotations all produce overlap. The report exists so that the writer can review each matched passage against its source before the text is handed in, published or submitted.

2. AI indicators: features of machine-generated text

Each unit of running prose (a paragraph in the current version of the method; sentences where the detector supports them) receives a score between 0 and 1 from a classifier trained to distinguish text written by people from text produced by large language models. Units above the threshold stated in the report are highlighted. The share reported for the document is the proportion of evaluated text above the threshold, not a probability that the document was generated.

Headings, tables, formulae, reference entries and very short fragments are excluded from evaluation. Texts with fewer than about 3,000 characters (roughly 500 words) of running prose outside the reference list receive no indicators, because the classifier is not reliable on such input.

Figures are shown only when the detector in use has been validated for the language of the text. While a validated detector for English is being prepared, the report states "AI indicators for English are in preparation" in place of figures; no score from an unvalidated heuristic is ever shown.

Known limitations

• Text by writers whose first language is not English, highly formulaic prose (methods sections, legal boilerplate, abstracts written to a template) and heavily edited text receive more false signals than average. Published evaluations of comparable detectors report false-positive rates in these groups that are several times higher than on native-speaker prose.
• Text that was generated and then paraphrased by a person, or produced with instructions to vary style, receives fewer signals. The absence of indicators does not establish that no tool was used.
• Indicators drift as language models change. The report states the detector version so that results can be interpreted against the version in use at the time.

For these reasons the report never states that a text is machine-generated. It lists passages with indicators so that the writer can reconsider wording, and so that a marker or an editor can ask about a passage rather than draw a conclusion from a number.

3. Reference check

Entries of the reference list are parsed for DOIs, titles and identifiers and looked up in Crossref. Each entry is marked:

Where Crossref carries a retraction notice for a matched record, the report says so with a link. "Not found" is a prompt to verify the entry by hand, not a statement that the work is fabricated.

4. AI-use disclosure draft

Many journals, following the ICMJE recommendations and publisher policies, ask authors to disclose the use of AI-assisted technologies in the manuscript's Methods or Acknowledgements section, and not to list such tools as authors. Universities and publishers of other kinds of writing increasingly ask for a similar statement. The report includes a neutral draft statement the author can adapt: which tool, for what purpose, and that the authors reviewed and take responsibility for the content. The draft does not assert that a tool was used; it is a template.

5. What the service does not do

• It does not issue verdicts ("AI-generated: yes/no"), accuracy percentages or guarantees.
• It does not rewrite text, "humanise" it, or help a text evade detection.
• It does not add the submitted text to any comparison database, and it does not use the text to train models. Texts are deleted with the report after 30 days.
• It does not compare against publishers' closed full-text collections or pages behind logins; coverage is limited to open records and the web.
• It does not use the text for any purpose other than producing the report, and it is not an instrument of surveillance: results are returned to the person who ran the check.

6. Contesting a result

If a report marks a passage as overlapping although it is correctly quoted, or marks a reference as not found although it exists, or lists indicators on a passage you wrote yourself, write to [email protected] with the report link and the passage or entry in question. We review such cases, correct source classification where warranted, and use them to adjust the method. Markers and editors receiving a report are asked to treat it as a starting point for a conversation, consistent with COPE guidance on handling suspected misconduct.

See also: for students · for writers · for researchers · for journals · privacy notice · terms of use.

Last updated 17 October 2026.