Perplexity
Perplexity is a measure of how surprised a language model is by a piece of text — low perplexity means each word was roughly what the model would have predicted, and high perplexity means it was not.
Last reviewed 15 August 2026 · The HumanFlow team
In plain English
Imagine reading a sentence one word at a time and guessing the next word before you see it. If you guess right almost every time, the text has low perplexity. If it keeps wrong-footing you, perplexity is high.
Detectors use it because machine-generated text tends to be the text a model found least surprising — which makes sense, since a model produces the words it considers most likely.
A worked example
Two sentences carrying the same information. The first is what a model reaches for; the second is what a person who has actually taught the class writes.
A. The results of the study clearly demonstrate that there is a significant relationship between the two variables. B. The correlation held across all three cohorts, which surprised us, because the second cohort was tiny.
In A, every word is the obvious continuation of the one before it. "The results of the study" is followed by "clearly demonstrate"; "a significant" is followed by "relationship". A model asked to predict each next word would score well. That is low perplexity.
In B, "which surprised us" is not a predictable continuation of a sentence about correlation, and "because the second cohort was tiny" is a specific admission a generic model has no reason to generate. Each of those raises perplexity.
Notice what actually did the work: B is not more complicated or better written. It contains information the writer had and the model did not.
Why it matters for AI detection
Nearly every detection tool in this category is built on some version of it, so understanding perplexity explains most of what a detection score is responding to.
It also explains the failure mode that matters most. Perplexity measures predictability, not authorship. Writing that is genuinely predictable — a formulaic lab report, a lease, a competent essay by someone writing in a second language and sticking to safe constructions — is low-perplexity because of what it is, not because of who or what produced it.
That is the mechanism behind the false-positive problem rather than a separate quirk of it. A detector reporting low perplexity has observed that your text was predictable. Whether that is because a machine wrote it or because you write plainly is a question the number cannot answer.
Commonly confused with
- Burstiness
- Perplexity is about how predictable each word is. Burstiness is about how much sentence-to-sentence variation there is. A text can be low in one and high in the other, and most detectors look at both.
- Readability
- Readability scores measure how hard text is to read — sentence length, syllable counts. Perplexity measures how expected the words were. Simple writing can be highly unpredictable, and dense academic prose can be extremely predictable.
Read next
Part of the AI detection glossary.