How ZeroClaude detection works
Two independent signals, blended into one score
1. Sentence segmentation
Text is split into sentences with language-aware rules, so English, Chinese and other languages are handled correctly. Words and characters are counted at the same time.
2. Language-model judgement
A large language model (Llama 3.3 70B running on Cloudflare Workers AI) reads the numbered sentences and estimates, for each one and for the document as a whole, how likely it is to be machine-written. It looks for uniform rhythm, generic hedged statements, stock transitions, balanced lists of three and the absence of personal detail.
3. Statistical signals
- Burstiness — people vary sentence length a lot; language models tend to be even.
- Stock phrases — the density of phrases models overuse (“furthermore”, “it is important to note”, “综上所述”).
- Informal markers — questions, exclamations, slang and first-person asides that people naturally leave behind.
4. Blending
Each sentence score is 70% model and 30% statistics. The overall score averages the document-level view with a length-weighted mean of sentence scores. Sentences scoring 60 or more are highlighted.
Limitations
Short texts, heavy editing, and very formal human writing all reduce reliability. Treat the score as evidence to weigh, not a verdict.