How the detector works
The detector scores a text from 0 to 100 using nine statistical signals. Each signal looks for a pattern that is common in AI-generated text and rarer in human writing:
| Signal | What it measures | Why it points to AI |
|---|---|---|
| Character entropy | How predictable the letters are | Language models favor smooth, common wording |
| Word entropy | How predictable the vocabulary is | Low surprise across the whole text |
| Burstiness | How much sentence lengths vary | AI text often keeps a steady rhythm |
| Type-token ratio | Unique words vs. total words | Limited vocabulary diversity |
| Hapax ratio | Words that appear only once | Humans use more one-off words |
| AI-typical phrases | Phrases like “delve into” or “it is important to note” | Typical model filler |
| Sentence starters | How often sentences open the same way | Repetitive structure |
| Sentence complexity | Variance in clause length | Uniform complexity |
| N-gram repetition | Repeated three-word patterns | Recycled phrasing |
It handles English and German text and detects the language on its own. The full explanation is in our post How Our AI Text Detector Works. It groups the nine signals into seven methods: character and word entropy count as one method, and so do type-token ratio and hapax ratio.
Limits
- False positives: clear, formal human writing can score high.
- Non-native writers: simpler vocabulary can look AI-like. Do not judge people by the score alone.
- Edited AI text: light rewriting lowers the score fast.
- Short texts: below a few hundred words the signals get noisy.
Treat the result as a clue, not as proof.
Privacy
The analysis runs entirely in your browser in JavaScript. Your text is not uploaded, logged, or stored, and there is no account.