Paste any text to instantly check if it was written by ChatGPT, Claude, Gemini, or other AI. We analyze 9 linguistic signals — no guessing, no black boxes.
Burstiness · MATTR · Hedging · Zipf · 5 more signals
How It Works
We don't use a single score. Our engine runs 9 independent linguistic tests and weights them by diagnostic power.
We tokenize your text into sentences and words. We extract syllables, identify entities, and build a word-frequency distribution — all locally, no data stored.
Each signal is scored independently: burstiness (sentence rhythm), MATTR (vocabulary diversity), hedging density, Zipf compliance, transition word frequency, and four more.
Signals are weighted by effect size (burstiness = 25 pts, MATTR = 20 pts, etc.) to produce a 0–100 AI probability score with a confidence estimate.
| Signal | What it measures | Weight | AI tells |
|---|---|---|---|
| Burstiness | Variance in sentence length rhythm | 25 pts | AI writes at uniform "medium" length |
| Sentence Length CV | Coefficient of variation across sentences | 22 pts | AI keeps all sentences similar in length |
| MATTR | Moving Average Type-Token Ratio (50-word window) | 20 pts | AI reuses the same vocabulary |
| Hedging Density | Epistemic markers per 100 words | 12 pts | AI avoids uncertainty language |
| Flesch Variance | Readability score variance across paragraphs | 8 pts | AI keeps all paragraphs equally complex |
| Transition Density | "Furthermore/Moreover/Additionally" per sentence | 7 pts | AI overuses formal connectors |
| AI Vocabulary | Frequency of AI-favoured words (50+ tracked) | 6 pts | "Delve", "leverage", "paradigm" etc. |
| Zipf Compliance | Word frequency distribution vs. Zipf's law | 5 pts | AI overuses mid-frequency words |
| Entity Coherence | Named entity overlap between consecutive sentences | 4 pts | AI references the same topic in every sentence |
FAQ
Yes, completely free. No account, no word limits, no paywalls. Paste text and analyze — that's it.
The score is directional, not certified detection. It's most reliable for texts of 100+ words. Shorter texts get lower confidence. Use the confidence indicator alongside the score. No detector — including GPTZero or Turnitin — is 100% accurate.
Yes. Our signals target statistical properties common to all large language models — uniform sentence rhythm, low lexical diversity, absence of hedging — regardless of which model generated the text.
No. Text is processed in memory and discarded immediately. We do not log, store, or share any submitted text.
A score of 70 means 70% of our weighted signals point toward AI-generated text. Below 30 = likely human-written. 30–55 = uncertain. Above 55 = likely AI-generated. The verdict label reflects the current threshold at each range.
This is a false positive. It can happen with highly formal writing, academic abstracts, or texts that use very structured language. The score is statistical — not a judgment of your writing. Check the individual signals to see which metrics triggered.
You can use it as a screening signal, but do not use any AI detector — including this one — as sole evidence of academic misconduct. AI detectors have false positive rates. Use the score as a starting point for a conversation, not a conclusion.
Our AI Humanizer rewrites flagged text to bypass Turnitin, GPTZero, and Originality.ai — free, no login required.
Humanize Free →