Perplexity and Burstiness, Explained
Updated
Perplexity and burstiness are two ideas often used to explain how AI text detectors work. Perplexity is about how predictable each word is; burstiness is about how much that predictability varies from sentence to sentence. Understanding them helps explain why plain, careful human writing sometimes gets flagged as AI.
Limitations: Every result is a probability that can be wrong, not proof of how something was written. How results can be wrong →
3 free scans in total, shared across all five tools (not per day or per month). No signup needed to start.
What Perplexity and Burstiness Mean
- Perplexity: a measure of how surprised a language model is by each next word. Low perplexity means the text is very predictable.
- Burstiness: how much sentence length, structure and predictability vary through a text. Human writing tends to mix short and long, plain and unusual.
- The intuition: language models tend to pick likely words, so their text is often low-perplexity and evenly paced.
- The catch: careful human writers can be predictable and even too.
Why These Signals Mislabel Human Writing
Writing that follows rules closely, such as technical reports, legal text, school essays and writing by people working in a second language, tends to use common words and regular sentence shapes. That makes it look low-perplexity and low-burstiness even when no AI was involved.
Research on seven GPT detectors found they flagged over half of a set of TOEFL essays by non-native English writers as AI-generated, while classifying essays by US eighth-grade students almost perfectly. Predictable wording is a likely part of that gap.
What Detectors Actually Use
Commercial detectors don't publish their full methods, and most now rely on trained classifiers rather than one or two simple statistics. The detection provider behind this site's text tool (named on the How It Works page) doesn't disclose exactly which signals it uses, so we can't say how much perplexity or burstiness matters to its score.
What we can say: the text tool needs at least 300 words, because statistical signals of any kind are unreliable on short texts.
What This Means for Writers
Varying sentence length or adding unusual words to dodge a detector isn't good advice: it makes writing worse and doesn't reliably change scores. If you wrote something yourself and it is flagged, keep your drafts and notes, and point to the known limits of detectors.
Frequently asked questions
A measure of how predictable each word is to a language model. Very predictable text has low perplexity, which some detectors associate with AI.
How much the rhythm and predictability of writing vary through a text. Human writing often varies more than AI output.
We don't know. The provider doesn't publish the exact signals it uses; see How It Works for what we do know.
We don't give advice on evading detectors. If your own writing is flagged, keep your drafts and explain the known limits of AI detection.
Sources
- GPT detectors are biased against non-native English writers (Liang, Yuksekgonul, Mao, Wu, Zou) (arXiv 2304.02819, read on )