Perplexity measures how predictable a sequence is to a probability model. Lower perplexity indicates that the model finds the sequence more predictable, making the measure useful as one signal when studying generated or human-written text. section 6.9
Text-detection systems use perplexity alongside burstiness, the variation in sentence length and structure, to estimate whether writing resembles machine-generated text. These signals are useful but fragile: human writing can also be predictable, stylistically varied, or unexpectedly worded. section 6.9
Perplexity is limited as a standalone indicator because:
- It can produce false positives, especially for non-native English speakers and neurodivergent writers. section 6.9
- It reflects patterns that are not unique to artificial intelligence, including ordinary variation in sentence length and word choice. section 6.9
- It becomes less reliable when detectors encounter different models, prompts, configuration settings, or changing user behavior. section 6.9
Perplexity therefore works better as one contextual clue than as proof of authorship. A detector may perform well in a controlled setting yet fail when model families, writing styles, or prompts change. section 6.9
