Burstiness.

Burstiness

Burstiness is a classical NLP feature that measures variance. In the AI-detection context, it almost always means variance in sentence length and syntactic complexity. Human writers produce a mix of short punchy sentences and long winding ones; base-rate AI output clusters around a median length.

Burstiness alone is not enough, highly edited academic prose is also low-burstiness. Modern detectors combine burstiness with perplexity and a learned classifier.

What bursty writing looks like

A bursty paragraph might open with a 24-word complex sentence, drop into a five-word fragment for emphasis, then build a 38-word sentence with two embedded clauses. Compare that to a low-burstiness paragraph: every sentence in the 14-22 word range, all subject-verb-object structure, no fragments. Most professional human writing varies sentence length deliberately; most untouched LLM output does not.

Where burstiness fails

Burstiness is unreliable on short passages (under 200 words there's not enough variance to measure), on heavily-edited academic prose (revision tends to flatten variation), and on certain genres (legal briefs and lab reports are intentionally low-burstiness even when human-written). Modern detectors weight burstiness less heavily than they did in 2023, in favor of embedding-based features that generalize better.

How Burstiness Interacts with Perplexity

Burstiness and perplexity function as complementary measurements in AI detection systems. Perplexity quantifies how predictable a text is overall, while burstiness examines the consistency of that predictability across sentences. A document may exhibit low perplexity (indicating predictable word choices) yet also display low burstiness (uniform sentence complexity throughout). Conversely, human academic writing often shows moderate perplexity with high burstiness, reflecting the natural variation between explanatory sentences, topic sentences, and complex analytical passages.

Detection algorithms weight these metrics differently depending on context. Research papers analyzed in 2024 studies demonstrate that burstiness alone achieves approximately 68% accuracy in distinguishing human from AI text, but combining it with perplexity measurements increases accuracy to 89%. The relationship becomes particularly diagnostic when burstiness remains low while perplexity varies, a pattern characteristic of AI models that alter vocabulary without changing structural rhythm. Institutions implementing detection tools should therefore examine both metrics rather than relying on burstiness scores in isolation.

Edge Cases and Known Limitations

Burstiness measurements produce unreliable results in specific document types and writing contexts. Technical documentation, legal contracts, and standardized report formats naturally exhibit low burstiness due to deliberate uniformity in sentence construction. Similarly, texts under 200 words lack sufficient sample size for statistically meaningful burstiness calculation, as variance metrics require adequate data points to distinguish intentional patterns from random fluctuation. Writing by non-native speakers or individuals with certain learning differences may also show atypical burstiness patterns that do not correlate with AI generation.

Domain-specific writing conventions further complicate burstiness interpretation. Scientific abstracts typically maintain consistent sentence length and complexity throughout, producing low burstiness scores despite human authorship. Journalism following AP style guidelines similarly prioritizes uniform sentence structure. Detection systems trained primarily on general academic prose may therefore flag legitimate human writing in specialized genres. Practitioners should establish baseline burstiness ranges for specific document types before applying thresholds, and should never use burstiness as a sole determinant of AI involvement without examining content, citation patterns, and contextual factors.