How the AI Detector Works

This tool is a heuristic, client-side analyser that computes over 30 stylometric, structural, and lexical signals in your browser — no text is ever sent to a server. It does not use any external AI APIs.

Core signals analysed

• Burstiness — variance in sentence length. Human writing tends to have high variance; AI-generated text often clusters around a mean.
• Perplexity proxy — vocabulary predictability using unigram and bigram frequency against a public English frequency list.
• Type-Token Ratio (TTR) — ratio of unique words to total words. AI output often shows unusually moderate TTR.
• Hapax ratio — words appearing only once. Human creative writing tends to use more hapax legomena.
• AI transition phrases — patterns such as "it is important to note", "furthermore", "in conclusion", "delve into".
• Punctuation profile — em-dash, semicolon, and colon density, which often differ between human and AI writing.
• Sentence-start variety — how many unique opening words appear across sentences.
• Latinate word ratio — proportion of words with Latin-derived suffixes (-tion, -ment, -ance).
• Contraction density — AI text tends to avoid contractions (don't, it's, we're) compared to human writing.
• Paragraph uniformity — variance in paragraph length and sentence counts per paragraph.

Model signature families

The detector compares each signal against patterns publicly documented for three AI model families. These are heuristic signatures, not vendor APIs, and should be treated as indicative not absolute:

• GPT-family signature — formal tone, high phrase-template density, moderate sentence-length variance, frequent "delve / tapestry / navigate" style tokens.
• Claude-family signature — longer median sentence length, richer qualifiers ("however", "moreover", "in particular"), low contraction density.
• Gemini-family signature — more list-friendly structure, "additionally/firstly/secondly" sequencing, direct address tone.

Confidence scoring

Each signal contributes a weighted score. Signals that agree (all pointing AI or all pointing human) yield higher confidence; signals that conflict reduce confidence. Final confidence is displayed alongside the AI probability.

What this tool is not

• It is not a trained neural network. It does not access GPT-4, Claude, or Gemini. It does not claim 100% accuracy.
• It cannot definitively prove authorship. Short texts, poetry, technical documentation, and second-language writing can all mislead any detector.
• It should be used as one input among many when assessing a document — never as the sole basis for accusation or disciplinary action.

Ethical use guidelines

AI detection is inherently probabilistic. False positives are common with legitimate human writing. We recommend using this tool to support human judgement, not replace it. See Wired's coverage for context on the limits of AI detectors.

Frequently Asked Questions