Perplexity
Updated June 10, 2026
A measure of how surprised a language model is by a text — the lower the surprise, the more machine-like the text.
Definition
Perplexity quantifies predictability: run text through a language model and measure how well the model predicts each next word. Text built from likely words scores low; text with unexpected turns scores high. Because generated text is literally assembled from each-step-likely words, low perplexity is the residue of generation — the founding insight behind detectors like GPTZero.
An example
"The meeting was productive and we covered several important topics" — every word is the expected next word; minimal perplexity. "The meeting limped through its agenda like a parade in the rain" — "limped" and the simile spike the surprise. Both are grammatical; only one could have been autocompleted.
The limits
Low perplexity isn't proof of AI: disciplined formal writing — academic, legal, corporate — is deliberately predictable, which is why those genres false-flag. And maximal perplexity is just gibberish. Human writing lives in the middle: mostly fluent, genuinely surprising where it counts. Paired with burstiness, it forms the interpretable core of AI detection.