Perplexity and Burstiness in AI Detection, Explained
Perplexity concerns how predictable text is to a particular language model. Burstiness describes variation, although its exact definition depends on the system. Neither number, by itself, proves who wrote a passage.
What perplexity actually measures
A language model assigns probabilities to the next token in a sequence. Perplexity summarizes how well those probabilities explain the supplied text: a lower value means the model found it more predictable. It is a model evaluation measure, not an authorship percentage.
The chosen model, tokenizer, and available context affect the result. A score from one setup is not directly interchangeable with a score from another. The Hugging Face explanation of perplexity describes these dependencies and how context windows affect the calculation.
A prediction example
Compare “Please bring your notebook to the meeting” with “Please bring your notebook to the lighthouse.” Depending on the preceding context, a model may find one ending less expected. Both sentences could have been written by a person or generated by software.
This example illustrates predictability; no perplexity values were calculated. An unusual word may add surprise without adding anything useful for the reader.
What burstiness means—and what it does not
In writing advice, people often use burstiness to mean changes in sentence length or rhythm. Historical detector discussions also use it for variation in predictability across a passage. These are related descriptions, not a single universal measurement.
Three similarly sized sentences can be clear and well written. A short sentence followed by a long one can be awkward. Counting words alone does not establish authorship or quality, and there is no universal target you should reach to make a document human.
Does GPTZero still use these two scores?
GPTZero's own support documentation says it moved away from perplexity and burstiness to a deep learning architecture in autumn 2023. Advice that treats the current product as a simple two-score formula is out of date. That statement concerns GPTZero; it does not describe every other detector.
For a flagged draft, identify the actual service and read what its result means. Our GPTZero score guide and false-positive guide explain the next review steps.
Use the concepts to ask better editing questions
Look for repetitive openings, vague claims, and transitions that add no information. Change them when the result is easier to understand. Keep a familiar phrase when it is the clearest way to express a technical point. Do not introduce errors or unusual synonyms just to make the text less predictable.
Try the worked example in our editing guide. Evaluate what meaning survives and what the reader gains, rather than treating either metric as a score to optimize.