Stylometry

Stylometry is the quantitative study of writing style: measuring features like sentence length distribution, vocabulary richness, function-word frequency, punctuation habits, and syntactic patterns. It long predates AI — it's been used to attribute disputed authorship and identify anonymous writers — and it's the conceptual foundation modern AI text detectors sit on.

An AI text detector is, at heart, a stylometric classifier. Instead of asking "who wrote this?" it asks "was this written by a human or a language model?" and answers using the same kinds of features: perplexity and burstiness are stylometric measures, and so are the finer signals like the ratio of function words or the distribution of sentence openers.

Understanding stylometry explains both why detectors work and why they're unreliable. They work because AI text really does cluster in style — low perplexity, low burstiness, even register. They're unreliable because plenty of human writing clusters there too: non-native writers, people trained in formulaic academic prose, and anyone writing in a clear, plain style can score as machine-written. This is a major source of false positives.

Humanizing text is, in effect, shifting its stylometric profile: raising burstiness, raising perplexity, varying the surface features that a classifier keys on. SynthGuard's layers each target a stylometric axis — restructuring for sentence-length variance, phrase replacement for word predictability, contractions and punctuation for register. Because the analysis is statistical and probabilistic, no shift guarantees a given detector's verdict.