Guide

Where "burstiness" and "perplexity" actually come from

These two terms get thrown around constantly in AI-detection marketing copy, usually without much explanation of what they mean or where they came from. Both have a real history in computational linguistics and language modeling that predates the current wave of AI detectors by decades, and understanding that history makes the numbers on this page far more useful than treating them as a black-box score.

This tool exists specifically to show you the raw numbers — sentence-by-sentence length data, vocabulary diversity, repeated-word frequency — without wrapping them in a single "AI or human" verdict. If you want a percentage-style verdict instead, our AI Text Detector uses this same underlying math to produce one. This page is for people who want to see the mechanism directly.

Burstiness in detail

Burstiness, in linguistics and information theory, describes how unevenly a feature is distributed across a sequence — the term is borrowed from physics and network theory, where it originally described clusters of events (like bursts of network traffic) separated by quieter periods, rather than events spread evenly over time. Applied to sentence length, a "bursty" passage has a genuine mix of short, punchy sentences and long, complex ones; a "non-bursty" passage has sentences clustering tightly around a similar length throughout.

This tool measures burstiness as the coefficient of variation (CV): standard deviation of sentence length divided by mean sentence length. This normalizes for the fact that a passage of short sentences and a passage of long sentences can both show the same relative variation. A CV near 0 means every sentence is nearly the same length; a CV above 0.6 or so indicates substantial variation.

Perplexity in detail

Perplexity has a precise mathematical definition in language modeling: it's the exponentiated average negative log-likelihood the model assigns to each word in a sequence, given the words before it. In plain language, it measures how "surprised" a specific language model is by a piece of text — lower perplexity means the text closely matches patterns the model has seen many times before (highly predictable to that model); higher perplexity means the text contains more statistically unusual word choices and structures.

Perplexity is inherently model-specific — the same passage will produce a different perplexity score depending on which language model is doing the measuring, because different models are trained on different data and learn different notions of "predictable." This tool's perplexity proxy sidesteps the need for a specific model entirely by combining two measurable stand-ins: lexical diversity (the ratio of unique to total words) and the density of phrases that are disproportionately common in AI-generated text specifically, based on documented patterns from independent analyses of AI-generated writing samples.

How to read the sentence-length chart

Each bar in the chart represents one sentence, with height proportional to its word count. Bars highlighted in amber are within roughly half a standard deviation of the passage's mean sentence length — a visual way of seeing exactly which sentences are contributing to a low or high burstiness score. A chart that looks like a fairly even picket fence, with most bars close to the same height, corresponds to low burstiness. A chart with a genuine mix of short and tall bars corresponds to high burstiness.

This visual is often more immediately convincing than the raw CV number, especially when explaining to someone else why a passage reads as unusually uniform — or why a "likely AI" flag on your own writing might be a false positive driven by one section of unusually consistent sentences.

What the repeated-words table tells you

The frequency table strips out common function words (the, a, is, and similar) and shows which remaining content words appear most often, along with what share of the total text they represent. Two things are worth watching for here: a small number of words dominating the text (often a sign of a narrow, possibly repetitive topic treatment — common in both weak human writing and some AI output that circles back to the same phrasing) and any of the words matching known AI-associated vocabulary (utilize, leverage, robust, and similar terms appear in the checker's internal common-word list and also count toward the perplexity proxy calculation).

Why genre changes what "normal" looks like

It's worth repeating a point that's easy to lose in a sea of percentages: these statistics behave very differently across genres, and there's no universal threshold that cleanly separates human from AI writing across all contexts. Legal contracts, technical documentation, and academic abstracts are naturally low-burstiness even when entirely human-written, because those genres reward consistency over stylistic variation. Personal essays, fiction, and casual blog posts are naturally high-burstiness. Before drawing any conclusion from a reading on this page, it's worth asking whether the genre itself explains the number, independent of who or what wrote it.

Using these numbers to sanity-check any detector

One of the most useful things you can do with this page is use it alongside a commercial AI detector's output. If a detector flags a passage as "likely AI" but you can see here that the burstiness (CV) is well above 0.5 and the vocabulary is genuinely diverse, that's a reasonable basis to push back on the detector's verdict rather than accept it at face value — you have the actual mechanism in front of you, not just a percentage you have to trust blindly.

Frequently asked questions