Perplexity Score

Perplexity measures how surprised a language model is by a sequence of text. When a model predicts the next token with high confidence consistently throughout a passage, perplexity is low. AI-generated text tends to have very low perplexity because the model that wrote it was optimizing for the most statistically probable next word — producing text that is, by definition, easy to predict.

Human writing is messier. People make unexpected word choices, change register mid-paragraph, and introduce structural variety that spikes perplexity in irregular ways. Detectors exploit this: text with uniformly low perplexity reads as AI-generated; text with higher, irregular perplexity reads as human.

AI humanizers work partly by increasing a text's perplexity score — introducing word choices and structural patterns that are less predictable. Coda One's humanizer does this while preserving semantic accuracy, which is the hard part. Blindly raising perplexity by inserting random words just produces bad writing. The target is human-range perplexity with coherent meaning intact.