Skip to content
Technology

Perplexity

A measure of how surprising a piece of text is to a language model — roughly, how unexpected each word is given the ones before it. Low perplexity means the text follows predictable paths.

Why it matters

Perplexity is one of the two signals AI detectors rely on, so it explains most of how they work and most of why they fail. It is also a standard metric for evaluating language models themselves, which is where it originated and where it is genuinely reliable.

A concrete example

The sentence "the cat sat on the mat" has very low perplexity: every word is the obvious continuation. "The cat sat on the escalator" has higher perplexity at the last word. A model generating text tends to pick the low-perplexity option, which is why generated prose reads smoothly — and why so does careful, plain human writing.

How to use it

You will meet this term mainly as one of the two things AI detectors measure, and knowing what it is takes most of the mystery out of how they work. Low perplexity means the text takes the predictable next word repeatedly, which is what a model optimising for likely output produces. It is also what careful, plain, formulaic writing produces — which is the whole reason detection built on it is unreliable. As a model-evaluation metric it is genuinely useful and largely a concern for people training models rather than using them.

The common mistake

Assuming low perplexity means machine-written. It means predictable, and plenty of human writing is predictable by design: technical documentation, legal boilerplate, anything written to a house style, and prose by someone writing carefully in a second language.

Related terms

Put Perplexity into practice

Access 750+ AI models and 60+ tools through Vincony — start free with 100 credits.