Check text for AI. Human or machine?
A statistical look at rhythm, vocabulary and typography. We show where exactly the text looks templated instead of passing a verdict on authorship.
- kinds of analysis
- 9
- languages
- 4
- texts stored
- 0
Your text is not stored on the server or written to logs.
Question or suggestion
Write to us: a question, a bug, an idea. If you want a reply, leave your e-mail below and we will answer by e-mail. The text you analyse is not sent here; please do not send passwords or payment details.
How the check works
Paste your text
Paste text, upload a file (PDF, DOCX, TXT and more) or give a link to a page. The limit depends on your plan and is shown under the editor.
Several kinds of analysis
Sentence and paragraph rhythm, AI phrases, connectives, vocabulary, typography and hidden characters.
Edit with hints
Every signal is tied to a place in the text and explained. No black box, no verdicts.
What we count
Each kind of analysis is explained in plain words. Some of them open on paid plans.
Overall score
Combines all signals into one score from 0 to 100.
Sentence length
People vary sentence length a lot, AI tends to be even. Measured as spread of lengths (lower = more even).
Paragraph length
People write paragraphs of different sizes, AI nearly equal ones. Needs at least 3 paragraphs.
Hidden characters
Invisible zero-width and similar characters left by AI services or copying. Found - red.
AI phrases
Stock phrases like "it is important to note", "in conclusion". Counted per 100 words; a weak signal.
Connective words
Density of words like "moreover", "therefore". AI uses them more; a weak signal.
Typography and special characters
Curly quotes, long dashes, ellipses and rare marks per 1000 characters. A weak signal: people use them too.
Lexical diversity
Share of distinct words in the text. Depends strongly on text length: higher in short texts. A weak signal.
Word entropy
A measure of word "unpredictability". Also length-dependent; a weak signal.
Who AuthorText is for
Authors and copywriters
Find templated phrasing and even rhythm in a draft before you publish.
How to read the scoreEditors
Quickly see where a text sounds like a template and what to change there.
PlansStudents and teachers
See which places look templated. Remember: the score is a hint, not proof.
Read on the blogDevelopers
Connect the check to your own service through the HTTP API.
APIWhy you can trust it
A hint, not a verdict
The score shows how many AI-like signals were found, not the probability of authorship. People also write evenly and in templates.
Texts are not stored
Your text goes to the server only to be analysed. We do not keep texts or write them to logs; statistics are counters only.
Honest about the limits
The thresholds are provisional and not yet calibrated. On short texts and in untested genres the site holds back a verdict.
Frequently asked questions
How the score works, what we keep and where the method ends.
Help and methodologyWe measure statistical features: spread of sentence and paragraph length, AI phrases, connectives, vocabulary, typography and hidden characters, and combine them into one number from 0 to 100. Every feature is visible on its own.
No. The score is a hint about style, not proof of authorship. A human author can write evenly too, and a text edited by a person can look natural.
No. The text is sent to the server only for analysis and is not kept. Statistics hold counters only: no texts and no addresses.
Russian, Ukrainian, English and Portuguese. In other languages the result may be inaccurate.
Text length and the set of analyses depend on your plan: your limit is shown under the editor. Details are on the plans page.
More kinds of analysis, a detailed breakdown and editing hints
The paid subscriptions open every kind of analysis, longer texts and a fragment-by-fragment review.
The score counts AI-like signals found - it is not a probability of authorship: <20 none; 20-35 weak; 35-55 mixed (review the spots above); 55+ probable AI. "NOTE" remarks mean the verdict is withheld (short text / out-of-calibration genre) - treat them as 'no data', not 'human'.
The score is based on statistical features of the text; thresholds are provisional and not yet calibrated. Not proof of authorship.