Burstiness
Burstiness measures variation in sentence length and complexity. Human writing tends to be more bursty than AI-generated text.
Definition¶
Burstiness is a statistical measure of variation in a sequence. In the context of AI text detection, it refers to how much a piece of text varies in sentence length, complexity, and structure.
High burstiness means a text alternates between simple and complex sentences, short and long constructions, formal and informal phrasing. Low burstiness means every sentence is roughly the same complexity and length.
Why Human Writing Is More Bursty¶
Human writers naturally vary their sentences because they're thinking about the idea, not just completing a prediction task. A writer might make a short, punchy statement, then follow it with a longer sentence explaining the nuance, then drop in a quick one for emphasis.
AI models, by contrast, optimize for coherent token prediction — which tends to produce sentences of similar statistical length and structure throughout a passage.
Burstiness as a Detection Signal¶
AI detection tools combine perplexity and burstiness scores. A text that is both low-perplexity and low-burstiness is a strong candidate for machine generation. A text that is low-perplexity but high-burstiness might simply be domain-specific writing (like legal or technical prose) that a human expert wrote.
This is one reason why AI detectors have false positive rates — they can flag dense, technically precise human writing as AI-generated, because that writing naturally has low perplexity and burstiness.
Improving Burstiness Through Humanization¶
The most effective humanization tools specifically address burstiness by: - Varying sentence length in the rewrite - Mixing simple and compound-complex sentence structures - Adding shorter "punctuation" sentences among longer explanatory ones
After humanization, the statistical profile of the text should show measurably higher burstiness.