Every conversation about AI-generated writing eventually lands on two words: perplexity and burstiness. The perplexity checker on this page exists to demystify both — not to score your text as "AI" or "not AI," but to show you the actual numbers language-model researchers use, translated into a browser-friendly proxy you can run on any paragraph in a few seconds.
This is a purely educational tool. It does not render a verdict, it does not compare you against a database of known AI samples, and it does not pretend to be a lie detector. It measures two properties of text — how predictable the word choices are, and how much the sentence structure varies — and hands you the raw numbers with a plain-English interpretation. Arb Digital built this alongside our other AI-writing tools because understanding the mechanics behind detection claims matters more than trusting a black-box score, and these two metrics are the actual foundation every detector — free or commercial — is built on.
What Perplexity and Burstiness Actually Mean
Perplexity, in the original machine-learning sense, measures how "surprised" a language model is by the next word in a sequence, given everything that came before it. A model trained on huge amounts of text gets very good at predicting common, expected word patterns — so text that closely follows those patterns produces low perplexity, while unusual word choices, unexpected turns of phrase, and idiosyncratic vocabulary produce high perplexity. Low perplexity text reads as smooth and predictable; high perplexity text reads as more surprising, more specific, or more personal. Because language models are literally optimized to generate low-perplexity continuations — that's what "sounds fluent" means to a model — their own output tends to score lower in perplexity than typical unedited human writing, which is fuller of quirks, tangents, and word choices a predictive model wouldn't have guessed.
Burstiness is a different axis entirely: it's about rhythm, not word choice. Human writers naturally alternate between long, complex sentences and short, blunt ones — a habit linguists call "burstiness" because the sentence-length pattern comes in bursts rather than a steady drumbeat. Ask someone to explain a decision and you'll typically get a short opening statement, a longer explanatory sentence, then maybe a two-word fragment for emphasis. Language models, left to their own devices, tend to produce a narrower, more even distribution of sentence lengths — technically fluent, but rhythmically flatter than how people naturally write and speak.
How to Use This Checker
- Paste a sample. Aim for at least 100–150 words; both metrics get noisier on very short snippets.
- Click Measure Text. The tool computes a vocabulary-diversity proxy for perplexity and a sentence-length-variance proxy for burstiness.
- Read the plain-English interpretation. The headline readout translates the raw numbers into a sentence like "Low perplexity, low burstiness — reads as highly predictable and structurally uniform."
- Compare samples. Run the same checker on a few paragraphs you know are human-written versus a few you know are AI-drafted, and watch how the numbers shift. That comparison teaches you more than any single score.
How the Proxy Metrics Are Calculated
True perplexity requires the actual probability distribution a language model assigns to each word, which isn't something a browser script can compute without the model itself. So this tool uses two well-established stand-ins that correlate with perplexity: type-token ratio (the number of unique words divided by the total word count — a measure of vocabulary diversity) and the density of extremely common function words ("the," "a," "is," "and," and similar). Text with a low type-token ratio and a high share of common words tends to be more predictable — closer to what a model would generate on its own — while text with a high type-token ratio and unusual word choices tends to score as higher perplexity. This mirrors the intuition behind real perplexity scoring, even though it isn't the same underlying math a GPT-style model uses internally. Google's documentation on language model fundamentals and Anthropic's published research both describe how next-token prediction underlies the fluency of modern models — the same mechanism that makes low-perplexity text easy for a model to produce and, often, easy for a detector to flag.
Burstiness is measured directly and honestly: the tool splits your text into sentences, counts words per sentence, and calculates the statistical variance of those lengths. A high variance — some four-word sentences mixed with some thirty-word sentences — produces a high burstiness reading. A tight cluster of similarly-sized sentences produces a low one. We also surface the single longest and shortest sentence in your sample so you can see exactly which lines are driving the number, and we count repeated three-word-and-longer phrases, since AI drafts often reuse the same transitional phrasing ("it is important to," "in addition to this") multiple times in one passage.
Why Every AI Detector Is Really Measuring These Two Things
Strip away the branding on any commercial AI-detection product and you'll find perplexity and burstiness doing most of the work, often combined with a handful of other signals like word-frequency profiles and formatting patterns. The core intuition is straightforward: a language model generates the statistically most likely next word most of the time, which produces text that is simultaneously low in perplexity (predictable word choices) and low in burstiness (even sentence rhythm). Human writing, especially unedited first-draft writing, tends to be messier on both axes — we reach for unexpected words, we ramble in one sentence and clip the next, we repeat ourselves inconsistently rather than uniformly.
But — and this is the part detectors gloss over — both metrics are trivially moved by editing. Asking a model to "vary sentence length" or "use more specific vocabulary" measurably shifts burstiness and perplexity in the human direction within a single prompt. And plenty of human writing genres — legal contracts, technical manuals, children's readers, formulaic five-paragraph essays — are naturally low-perplexity and low-burstiness because that's what clarity or convention demands in that context. Low scores on both metrics are evidence of a writing style, not evidence of authorship. That's exactly why this tool stops at reporting the numbers and refuses to convert them into a verdict.
Reading Your Results
A sample with high type-token ratio, high sentence-length variance, and a wide gap between its longest and shortest sentence is displaying the hallmarks of naturally varied writing — the kind that's harder for a predictive model to have generated in one pass. A sample with low type-token ratio, low variance, and several repeated phrases is displaying the opposite pattern — smooth, even, predictable. Neither pattern is a courtroom verdict. A tired writer banging out a quick email will often produce low-burstiness text. A model prompted carefully and edited by a skilled human can produce highly bursty, high-perplexity text. Use these numbers the way a musician uses a metronome reading — as information about tempo and texture, not as a judgment about who's playing.
It also helps to run the checker on your own past writing before you judge anyone else's. Pull an old email, a college essay, or a work report you know you wrote yourself and measure it. Most people are surprised at how low their own natural burstiness or type-token ratio reads on a rushed or repetitive piece — a reminder that these numbers describe writing conditions and habits as much as they describe authorship. A first draft written in ten minutes under deadline pressure and a polished third draft revised over a week from the same person can produce meaningfully different perplexity and burstiness readings, even though the same human wrote both.
Where These Metrics Come From
Perplexity has been a standard evaluation metric in natural language processing for decades, long before consumer chatbots existed — researchers used it to compare how well different statistical language models predicted held-out text, with lower perplexity scores signaling a better-fitting model. When large language models arrived, the same metric got repurposed in reverse: instead of using perplexity to judge a model, people started using a model's own low-perplexity output as a fingerprint to try to detect that model's writing. Burstiness has a similarly old pedigree in linguistics and even network science, where it originally described the clustering pattern of events in time — email traffic, earthquake aftershocks, word usage — before writing researchers borrowed the concept to describe how sentence lengths cluster within a single author's text. Neither metric was designed as an AI-detection tool; both were repurposed for that job because they happened to correlate with a real, if imperfect, difference in output patterns.
Arb Digital's editorial team writes with real research and varied voice — the kind of texture that comes from actual expertise, not a prompt. Talk to us about content that reads like it was written by someone who knows the subject.
Explore Our Services All Free ToolsCommon Mistakes to Avoid
- Testing tiny samples. A single sentence can't produce a meaningful variance calculation — you need at least several sentences.
- Treating perplexity as a yes/no signal. It's a spectrum, and plenty of legitimate human genres sit at the "predictable" end.
- Confusing this with an accusation tool. This checker deliberately does not label anything "AI-written" — that's a different tool with a different, more cautious framing.
- Ignoring genre. A legal disclaimer and a personal blog post should not be judged by the same perplexity baseline.
- Assuming higher perplexity always means "better" writing. Extremely high perplexity can also just mean confusing, poorly structured prose.
Related Free Tools From Arb Digital
Want a likelihood estimate instead of raw metrics? Try the AI Content Detector. Editing your own AI-assisted draft to sound more natural? Use the AI Text Humanizer Checker. Choosing which model to write with in the first place? See the AI Model Comparison and estimate your usage with the ChatGPT Token Counter or the Words to Tokens converter. Browse everything in our free online tools hub.
Frequently Asked Questions
Perplexity measures how surprised a language model would be by the next word in a passage. Low perplexity means the text follows highly predictable patterns; high perplexity means the word choices are less expected.
Burstiness measures how much sentence length varies across a passage. Human writing tends to burst between short and long sentences, while AI output often settles into a more even rhythm.
No. This tool reports the raw perplexity and burstiness metrics with a plain-English interpretation only. It does not produce an AI-or-human verdict — for a likelihood estimate, use our AI Content Detector.
Not exactly. True perplexity requires the model's internal probability distribution. This tool uses vocabulary-diversity and common-word-density proxies that correlate with perplexity but are calculated entirely in your browser.
Because language models are optimized to produce predictable, fluent continuations, their output tends to score low on both perplexity and burstiness — a pattern detectors are built to spot, even though both metrics can be shifted by editing or genre.
Yes, frequently. Technical documentation, legal text, and formulaic writing genres are naturally low on both metrics regardless of who — or what — wrote them.
