Perplexity, burstiness and stylometry: three different routes to detection
Perplexity measures how predictable the next word is. Burstiness measures the variation in sentence length and complexity. Stylometry compares a text against one specific known hand. The first two judge a text on its own; the third needs something to compare with — and therefore answers a different question.
Perplexity: how much the text surprises
Language models pick likely continuations, so generated text tends to be low in surprise. Low perplexity is therefore suspicious.
The weakness is obvious: a person who writes plainly and factually also has low perplexity.
Burstiness: the variation between sentences
People alternate long complex sentences with short ones. Models keep a more even rhythm. Burstiness measures exactly that variation.
It loses its force wherever an even rhythm is desirable — in a methodology, in a legal text, in a manual.
Stylometry: comparison with a specific writer
Stylometry does not ask "does this look like AI" but "does this look like you". It computes measurable properties of the text — sentence and word length, vocabulary richness, the frequency of commas, dashes and function words — and compares them with a profile built from your older texts.
This is the layer DetekceGPT has in addition. It does not work from the very first check — it needs at least six of your earlier texts in the history to have something to compare against.
FAQ
Which approach is the best?
None on its own. Each answers a different question, and their combination says the most.
Why do two detectors give me different numbers?
Because they measure different things. The difference between them is information, not an error.
Does stylometry need a lot of texts?
In DetekceGPT six earlier checks longer than two hundred and fifty characters are enough. The profile then keeps sharpening up to fifty texts.
Try it on your own text
Paste a text or upload a file and look at the score and at the specific sentences that came out suspicious.
Run a detection