How predictable a text is to a language model.
Perplexity measures how surprising a text is to a language model. AI text has low perplexity (the model predicts its own output easily); original human writing is higher. Alone it is unreliable, so AIX combines it with other signals.