what’s perplexity?
Perplexity measures how ’surprised’ a language model is by text. Low perplexity = predictable (AI-like). High perplexity = unexpected (human-like).
perplexity is the workhorse signal of AI-text detection. intuitively: imagine reading a sentence one word at a time and trying to guess each next word. “the cat sat on the ___” is easy... probably “mat” or “floor.” “the cat sat on the fjord” is surprising.
a language model assigns probabilities to each possible next word given everything that came before. perplexity is a function of those probabilities: a measure of how predictable the text was, on average. low perplexity means “the model wasn't surprised by what came next.” high perplexity means “the model didn't see this coming.”
human writing veers toward the surprising. we make idiosyncratic word choices, switch registers, throw in slang, get distracted, chase a metaphor we didn't plan. AI-generated text, by contrast, is the output of a model optimized to pick high-probability next tokens, so its average perplexity tends to be lower and more uniform than human prose on the same topic.
detectors like the ones in amige.'s text panel score a passage's perplexity against a reference language model and flag low scores as suspect. at scale across a paragraph it's surprisingly predictive. but it's a statistical lean, not a smoking gun.
low perplexity = predictable = AI-like; high perplexity = surprising = human-like. that's the headline.
perplexity is measured against a specific reference model. the same text can score differently against different detectors. that's one of the reasons amige. runs a panel rather than a single classifier. different reference models notice different things.
perplexity is a probability signal, not a proof. a clean, formulaic human writer can score low (think: technical documentation, legal copy, an English learner's carefully-constructed essay). a chaotic prompt can make AI score high (creative writing in an unusual voice). the verdict is always probabilistic.
the most common misconception is that a low perplexity score means “this is AI.” it doesn't. it means “this text is unsurprising to a language model,” which correlates with AI output but is not synonymous with it. non-native English writers, technical writers, and people writing in constrained formats all score lower-than-average for reasons unrelated to AI use.
questions
what is perplexity in AI detection?
perplexity measures how surprised a language model is by a piece of text. predictable wording scores low and correlates with AI writing. unexpected wording scores high and leans human. most AI-text detectors read this signal across a whole passage, since a single sentence carries too little to tell you much.
does low perplexity mean text is AI-generated?
no. a low score means a language model found the wording unsurprising, which overlaps with AI output without matching it. non-native English writers, technical and legal writers, and people writing in tight formats all tend to score low for reasons that have nothing to do with AI. detectors give you odds, not proof.
why does the same text get different perplexity scores?
every detector scores against its own reference model, so one passage lands differently depending on which model does the reading. amige. runs a panel of classifiers for this reason. different reference models catch different things, and the spread between them tells you more than any single number.
can AI-generated text have high perplexity?
yes. an odd prompt pushes AI output toward a higher score, say creative writing in a strange voice. the signal points in a direction rather than settling a case. a high score leaves room for AI authorship the same way a low score leaves room for a human who writes cleanly.
sources.
- 01Jelinek et al. — Perplexity, a measure of the difficulty of speech recognition tasks (1977)The original information-theoretic definition.
- 02Mitchell et al. — DetectGPT: Zero-Shot Machine-Generated Text Detection using Probability Curvature (arXiv:2301.11305)Modern perplexity-based detection.
- 03