About this tool
AI writing detectors — including the commercial ones — are a genuinely hard, unsolved problem. This page does not run a trained detection model. Instead it computes four simple, transparent text statistics that researchers have loosely associated with machine-generated text: how much sentence lengths vary (bursty vs. uniform), how diverse the vocabulary is over rolling 100-word windows, how often exact 4-word phrases repeat, and how much word length varies within sentences. Each is shown separately, plus a combined 0-100 score, so you can see exactly why the number is what it is rather than trusting an opaque verdict.
None of these four signals is a reliable fingerprint of AI authorship on its own, and neither is their average. Careful human writers can produce very uniform, low-diversity prose (technical documentation, legal boilerplate, non-native-speaker writing) that scores "AI-like" here, and lightly-edited AI output can easily break all four patterns. Use this as a rough, transparent gut-check on a piece of text — a conversation starter, not a verdict — and never as evidence in an academic integrity case, hiring decision, or anything else with real consequences for someone.
Frequently asked questions
Can this prove a text was written by AI?
No, and this is the most important thing to understand about this tool: it computes 4 simple text statistics (sentence-length variance, vocabulary diversity, phrase repetition, word-length variance) and combines them into a score — it contains no trained language model of any kind. Even sophisticated commercial AI-detectors with real ML models have known false-positive/false-negative rates; a hand-rolled statistical heuristic is meaningfully weaker still. Never use this as evidence in an academic-integrity case, an employment decision, or anything else consequential.
Why would human writing score as "AI-like"?
The underlying signals this tool measures (uniform sentence lengths, lower vocabulary diversity, repeated phrasing) can show up in plenty of ordinary human writing too — technical documentation, legal text, non-native English writers who favor simpler, more repetitive sentence structures, or even just a careful, methodical writer's natural style. None of these signals were designed to be definitive proof of anything; they are weak statistical correlations, not causal markers of authorship. This is exactly why the tool shows its work (all 4 sub-scores individually) instead of hiding behind one opaque number.