Glossary of AI text detection terms

AI text detector: A tool that estimates whether a text was written by a person or by a language model. It does not verify the author; it measures how closely the text resembles what it knows as machine writing, and none is one hundred per cent accurate.
Detector score: A number, usually from 0 to 100, with which a detector expresses how closely a text matches its idea of machine writing. Each tool calculates it differently, so sixty in one does not mean the same as sixty in another.
Language model: A program trained on a large amount of text that writes by repeatedly choosing a likely next word. ChatGPT, Claude and Gemini are examples, and their texts are what detectors look for.
Perplexity: A measure of how predictable each next word of a text is for a language model. Low perplexity means the model would have written the text almost the same way, which is typical of AI but also of formulaic human writing.
Burstiness: A measure of how much short and long sentences alternate in a text. Human writing tends to be irregular and AI text more even; here this property is measured by the Machine feel detector.
Stylometry: Measuring an author's style through properties people hardly change on purpose — sentence and word length, punctuation or the frequency of the most common words. It is used to establish authorship, and here the Style deviation detector is built on it.
Personal style baseline: A profile of your writing calculated from your earlier texts, against which every new text is compared. Here it is built from up to the last fifty checks and your uploaded samples, and it starts working from six texts.
Phrases typical of AI: Turns of phrase that language models use far more often than people, such as “it is important to note” or “in today's fast-paced world”. A single one proves nothing, because people use them too; what matters is how dense they are.
False positive: A case where a detector flags a text written by a person as AI. In AI text detection it is the most serious error, because it can lead to an unjustified accusation.
False negative: A case where a detector misses an AI text and judges it human. It typically happens with texts someone rewrote by hand or humanised after AI.
Humanisation: Rewriting a text so that it feels less mechanical — with a more irregular sentence rhythm and without phrases typical of AI. Our humanisation keeps the facts, the numbers and the tone, and shows how the machine feel changed.
Watermark: A hidden mark that the generator itself puts into a text so that it can later be recognised as AI-made. In text it is fragile: rewriting or translating usually removes it, so detection cannot rely on it alone.
Hidden characters: Invisible Unicode characters such as zero-width spaces, direction marks or unusual kinds of spaces that travel with copied text. They are rare in human writing, which is why we count them in a separate detector.
C2PA: An open standard for a signed record of where content came from, embedded directly in the file. When such a record with a valid signature states that the content was made by AI, the Hidden characters detector shows 100.

Try it on your own text

Paste a text or upload a file and look at the score and at the specific sentences that came out suspicious.

Run a detection