What Do Perplexity and Burstiness Mean in AI Detection?
Perplexity describes how surprising a sequence of words is to a particular language model; burstiness informally describes variation or clustering in text patterns. Some detector explanations discuss these ideas, but they are not universal authorship tests and do not prove whether a person or AI wrote a passage.
What is perplexity?
In language modeling, perplexity is a measure related to how well a model predicts a sequence. Lower perplexity means the sequence is less surprising to that model; higher perplexity means it is more surprising. The value depends on the model and tokenization, so it is not an objective score of writing quality or human authorship.
The research paper DetectGPT studies probability curvature under a language model as a zero-shot detection signal. It does not establish that one perplexity value can identify every AI-generated passage.
What is burstiness?
Burstiness is a broad term for unevenness or variation in events or features. In discussions of writing, it is often used informally for changes in sentence length, rhythm, or predictability across a passage. Different tools may define and calculate it differently, so the label alone does not explain a detector’s result.
Some product documentation uses perplexity and burstiness to explain parts of a proprietary approach. That description belongs to that product; it should not be assumed to describe all detector models.
Can either measure identify an author?
No. These measures can be inputs to a classifier, but they cannot reveal a writer’s identity or process. Human and generated writing can overlap in statistical features, and editing, topic, language, and length can change measurements. Treat them as model signals that require context, not as standalone evidence.
For a broader view of detector robustness, see the RAID benchmark, which evaluates multiple detectors across text sources, domains, and perturbations.
Common questions
Is low perplexity proof that text was written by AI?
No. It only says a particular model finds the sequence more predictable. Many kinds of human writing can also be predictable.
Do all AI detectors use burstiness?
No. Detector methods differ, and many vendors do not disclose their complete model internals.
Sources and further reading
Related Ghostiq tools
- AI writing detector — review an estimate, not proof of authorship.
- Text revision tool — review every suggested change for meaning and accuracy.