About

Large language models produce text with statistically detectable patterns. The average sentence length variance (σ_len) in AI-generated prose is typically 30 - 50% lower than in human writing. AI models optimize for coherence, which flattens the natural "burstiness" of human cognition - the tendency to alternate between short, punchy sentences and longer, complex ones. This tool quantifies that signal alongside 8 other linguistic features: Type-Token Ratio, hapax legomena density, transition word overuse, punctuation diversity, syllable distribution, readability index, and n-gram repetition patterns. Each metric is scored independently, then combined into a weighted composite.

No detector is infallible. Heavily edited AI text or formulaic human writing (legal documents, technical manuals) can produce false results. The composite score assumes standard English prose of at least 50 words. Accuracy degrades below that threshold. Mixed-origin text (human-written with AI-assisted edits) will score in the ambiguous middle range, which is the correct result. Treat scores below 30 or above 70 as directional signals, not verdicts.

Formulas

The composite human-likeness score H is computed as a weighted sum of normalized sub-scores:

H = n∑i=1 w_i ⋅ s_i

Where s_i is the normalized score (range 0 - 1) for each metric and w_i is its weight such that n∑i=1 w_i = 1.

The Burstiness Index B is defined as:

B = σ_len²L

Where σ_len² is the variance of sentence lengths and L is the mean sentence length in words.

The Flesch Reading Ease score:

FRE = 206.835 − 1.015 × WS − 84.6 × YW

Where W = total words, S = total sentences, Y = total syllables.

Type-Token Ratio:

TTR = |V|N

Where |V| is the count of unique word types and N is the total number of word tokens.

Reference Data

Metric	What It Measures	Typical Human Range	Typical AI Range	Weight
Sentence Length Variance (σ_len)	Standard deviation of words per sentence	6 - 14	2 - 6	20%
Burstiness Index	Ratio of variance to mean in sentence lengths	0.5 - 1.5	0.1 - 0.4	15%
Type-Token Ratio (TTR)	Lexical diversity: unique words ÷ total words	0.55 - 0.80	0.40 - 0.60	12%
Hapax Legomena Ratio	Words used exactly once ÷ total words	0.40 - 0.65	0.25 - 0.40	10%
Transition Word Density	Discourse markers per 100 words	0.5 - 2.0	2.5 - 5.0	12%
Punctuation Diversity	Unique punctuation types ÷ total punctuation	0.30 - 0.70	0.15 - 0.30	8%
Flesch Reading Ease	Readability based on syllables and sentence length	40 - 80 (varies)	50 - 65 (narrow band)	8%
Readability Variance	Per-sentence Flesch score standard deviation	15 - 35	5 - 15	8%
Bigram Repetition Rate	Repeated word pairs ÷ total bigrams	0.02 - 0.08	0.08 - 0.18	7%
Average Syllables per Word	Vocabulary complexity indicator	1.3 - 1.8	1.5 - 1.7 (consistent)	-
Paragraph Length Variance	Consistency of paragraph sizing	High (irregular)	Low (uniform)	-
Contraction Usage Rate	Contractions per 100 words	1.0 - 4.0	0.1 - 0.8	-
Unique Sentence Starters	Variety in first words of sentences	60 - 90%	35 - 55%	-

Frequently Asked Questions

Language models optimize for fluency by maintaining consistent sentence structures. The standard deviation of sentence word counts (σ_len) in AI text typically falls between 2 and 6, while human writers naturally produce values of 6 to 14. Humans unconsciously vary rhythm: a 5-word declarative followed by a 35-word compound sentence. AI smooths this variation. The burstiness metric quantifies this pattern as variance divided by mean.

Statistical metrics like Type-Token Ratio and sentence length variance require sufficient sample size to stabilize. Below 50 words, there may only be 2 - 4 sentences, which is insufficient to compute meaningful variance. TTR is also length-dependent: shorter texts naturally produce higher ratios regardless of origin. The tool will still run on shorter text but flags the result as low-confidence.

Yes. If a human rewrites AI output by varying sentence lengths, adding contractions, inserting parenthetical asides, and breaking uniform paragraph structure, the statistical fingerprint shifts toward human-like values. The tool measures the text as-is, not its origin process. Conversely, human-written text that follows strict templates (legal briefs, military reports) may score as AI-like due to enforced structural uniformity.

The tool checks against a list of 80+ discourse markers ("however", "moreover", "furthermore", "in addition", "consequently"). The count is normalized per 100 words. AI models overuse transitions because training rewards text coherence. Human writers often rely on implicit logical connections or paragraph breaks instead. A density above 2.5 per 100 words is a moderate AI signal.

The syllable counting algorithm and transition word list are calibrated for English. Sentence splitting (period/exclamation/question mark) works cross-linguistically, so metrics like sentence length variance and burstiness remain valid. However, TTR, hapax ratio, and readability scores will be inaccurate for non-Latin scripts or agglutinative languages like Finnish or Turkish where word boundaries differ.

Hapax legomena are words that appear exactly once in a text. Human writers tend to produce more hapax because they draw from personal vocabulary idiosyncrasies, rare word choices, and context-specific terminology. AI models favor statistically common words from training distributions. A hapax ratio below 0.30 in a 300+ word sample is a moderate indicator of machine generation.