Drop a document here or browse
.txt · .md · .pdf · .docx — max 20 MB
This document is short — treat the result with extra caution.
Heuristic estimate, not proof. Formal human writing can score high; lightly edited AI text can score low. Use as a hint, not as evidence.
Document stats
Signals
Each signal is scored 0–1 (higher = more AI-like) and combined into the overall score.
Most suspicious sentences
Full text (colored by sentence score — hover a sentence for reasons)
Methodology & limitations
The headline score comes from a neural detector — a RoBERTa language model trained to distinguish human from AI-written text. It scores the document in segments and aggregates them, and the percentage next to each paragraph in the highlighted text shows the model's verdict for that paragraph.
Underneath, a set of style signals (shown in the Signals section) provides supporting, sentence-level evidence:
- Sentence-length variation (burstiness) — human writing alternates long and short sentences; AI text tends to be steadier.
- AI-typical phrases — stock expressions such as “delve into”, “it’s important to note”, “in conclusion”, “a testament to”.
- Transition-word openers — repeated “Furthermore…”, “Moreover…”, “Additionally…”.
- Repeated phrasing — reused word sequences and template-like structure.
- Paragraph uniformity, sentence-start diversity, em-dash usage, vocabulary richness.
Limitations: no detector is proof of authorship. Neural detectors can be fooled by deliberately “humanized” rewrites and can misjudge highly formal or formulaic human writing (legal, academic, non-native English). Short texts are the least reliable. Treat results as a strong hint, not as evidence.
Privacy: documents are processed in memory only on this server. Nothing is written to disk, logged, or shared, and no external services are contacted while analyzing.