AI Content Detector

The signals and how each one moves the score

The score starts at a neutral 50 and each signal adds or subtracts points. The result is clamped to 0–100. Nothing here is a trained classifier; it's arithmetic on a handful of measurable patterns, which is what makes it explainable and also what makes it fallible.

Labels are fixed thresholds: 65 and above is “Likely AI-generated” with a red bar, 40 to 64 is “Possibly AI” in amber, below 40 is “Likely human” in green.

Two worked examples

A 55-word passage of five sentences of 10–12 words each, containing “plays a crucial role” and “in conclusion” plus “moreover” and “furthermore”, scores like this: variation ratio 0.10 gives +24; two cliché phrases give +18; 3.6 transition words per 100 is capped at 3 for +18; a lexical diversity of 0.91 is capped at 0.9 and gives −11.25. Total 98.75, shown as 99%, Likely AI-generated. Every sentence in it is plausible human prose; the formula is reacting to rhythm and connectives, not meaning.

A 49-word diary-style passage with sentences of 4, 20, 3 and 22 words has a mean of 12.25 and a standard deviation of 8.8, a ratio of 0.72, for −13; no clichés and no transition words add nothing; diversity of 0.76 gives −7.6. Total 29, Likely human. Notice that rhythm alone moved the score by 37 points between the two.

Why false positives are common

• Academic and business English legitimately uses “therefore”, “moreover” and “in conclusion”. A well-structured essay by a careful human collects those points.
• Writers taught templated transitions, including many non-native speakers, write exactly the even, connective prose the formula penalises.
• Technical documentation and legal text keep sentences uniform on purpose, so their variation ratio is low.
• Lexical diversity falls naturally as text gets longer, because a 2,000-word piece must reuse its key words. Long human texts drift upward for that reason alone.
• Edited AI text passes easily: vary three sentence lengths and delete two stock phrases and the score can drop 40 points. A pass is not evidence of human authorship any more than a fail is evidence of AI.
• The number is an estimate from five crude signals. It must never be the basis for a grade, a disciplinary step or a public accusation.

Text it can't measure at all

Words are detected as runs of letters or digits in any script, with apostrophes kept inside a word, so accented French or Spanish words count once and Arabic, Russian or Hindi text is counted word by word. The counts are therefore right for any alphabet, but the signals that carry most of the score are English-specific: the cliché list and the transition words are English phrases, so for other languages those two signals read zero and the verdict rests on sentence variation and vocabulary alone. Treat a result on non-English text as weak evidence at best.

Sentences are split at full stops, question marks and exclamation marks, except where the stop sits inside a number (“2.5”), after a common abbreviation (“e.g.”, “Dr.”, “vs.”) or inside initials such as “U.S.”. An unusual abbreviation can still split a sentence in two, which makes a text look more varied than it is and lowers the score slightly.

What to do with the result

Treat a high score as a prompt to look closer, not a finding. If authorship matters, ask for drafts, version history or a short conversation about the piece; those establish far more than any percentage. For your own writing, the signals are useful as a style check: a cliché count above zero and a variation ratio under 0.3 usually mean the text reads flat, whoever wrote it. The AI Humanizer on this site rewrites with varied rhythm if you want help, but reading the piece aloud finds the same problems.