Burstiness vs Perplexity: Two Different Signals AI Detectors Measure (and Why They Disagree)

People throw “perplexity” and “burstiness” around like they’re the same idea wearing two names. They’re not. One measures your word choices; the other measures the shape of your whole piece, and on the very same paragraph they can point in opposite directions.

That disagreement isn’t a bug. It’s the most useful thing about them, because when the two signals split, they’re telling you something specific about what your writing needs.

Key takeaways

• Perplexity is a word-level signal. It asks, for each next word, how surprised a language model is to see it.
• Burstiness is a document-level signal. It asks how much that surprise varies from sentence to sentence across the whole piece.
• Because they measure at different scales, they can disagree on the same text. Human-sounding word choices don’t guarantee a human-sounding rhythm, and vice versa.
• Most modern detectors fold both, plus other features, into a single probability. A score is a blend, not a verdict.
• When the two split, the conflict points you to the fix: vocabulary and specificity for perplexity, sentence variation for burstiness.

The one distinction that explains everything

If you only remember one thing, make it this: perplexity is local, burstiness is global.

Perplexity looks at your text one word at a time. After “the cat sat on the,” a model expects “mat” and is barely surprised. After “the cat sat on the,” if your next word is “fax,” the model is shocked, and perplexity for that spot spikes. Add up that surprise across every word and you get the perplexity of the passage. GPTZero, which popularized the terms, describes perplexity as a measure of how likely a model would have chosen the same words. Low perplexity reads as machine-like, because models gravitate toward the predictable word.

Burstiness doesn’t care about any single word. It zooms out and asks: as you move sentence to sentence, does the predictability stay flat, or does it bounce around? Humans bounce. We write a long, clause-stacked sentence, then a three-word one. A model left on autopilot tends to hum along at one steady level. Burstiness is the size of that bounce.

So perplexity is texture. Burstiness is rhythm. You can have great texture and dead rhythm. You can have a lively rhythm built out of bland words. They are not the same measurement, and pretending they are is where most confusion starts. If you want the longer foundation on both, we wrote the full plain-English breakdown of perplexity and burstiness separately. This piece is about what happens when they fight.

Side by side

	Perplexity	Burstiness
Scale	Per word	Whole document
Question it asks	How predictable is this word?	How much does predictability vary?
What raises it	Surprising, specific, uncommon word choices	Mixing short and long, simple and complex sentences
What a low score looks like	Common, expected vocabulary	Even, metronome-like sentences
What machines tend to do	Pick the likely word (low)	Hold a steady rhythm (low)

Read the bottom row closely. Models tend to score low on *both* by default, which is why people lump the signals together. But “they usually move the same direction in raw AI output” is not the same as “they’re the same thing.” The moment a human edits, or a tool rewrites, the two can come apart fast.

Where they disagree: three mini-scenarios

Scenario 1: high perplexity, low burstiness

A student runs a ChatGPT draft through a synonym-swapper. “Important” becomes “salient.” “Shows” becomes “elucidates.” “Use” becomes “leverage.” Every word is now less predictable, so perplexity climbs. The detector’s word-level signal relaxes.

But the swapper changed nothing about structure. Every sentence is still 18 to 22 words, subject-verb-object, one idea each, on and on. Burstiness stays flat as a parking lot. A detector that weighs variance sees a smooth, even hum dressed up in expensive vocabulary, and still leans toward “AI.” The two signals disagree, and the structural one wins.

This is the most common real-world split, and it’s why thesaurus-humanizing rarely works on its own.

Scenario 2: low perplexity, high burstiness

Now flip it. A careful human writer keeps her vocabulary plain on purpose. She’s writing instructions, so the words are simple and predictable. Low perplexity. By the word-level signal alone, she looks machine-made.

But her rhythm is alive. A long step with a warning baked in. Then: “Stop here.” Then a medium sentence picking the thread back up. The sentence-to-sentence variance is high, so burstiness reads human. Here the signals disagree in her favor, and depending on how the detector weights them, she may pass or may get a confusingly mixed score.

Scenario 3: both agree, and they’re still wrong

A non-native English speaker writes a clean TOEFL essay. Conventional phrasing means low perplexity. Even, careful sentences mean low burstiness. Both signals agree: “AI.” Both are wrong.

This isn’t hypothetical. A 2023 study led by Weixin Liang at Stanford, published in *Patterns*, found GPT detectors biased against non-native English writers: across 91 TOEFL essays from non-native speakers, seven detectors produced an average false-positive rate of 61.3%, while correctly clearing more than 90% of essays from US eighth-graders. When the researchers used ChatGPT to enrich the vocabulary of those same essays, the false-positive rate dropped to 11.6%. Raising perplexity made the flags disappear, which tells you the tool was penalizing a linguistic background, not catching anything.

So agreement between the signals isn’t proof either. Two flawed estimators agreeing is just two flawed estimators agreeing.

Why a single score hides the disagreement

Here’s the catch: most detectors don’t hand you the two numbers separately. They blend perplexity, burstiness, and often a stack of other features into one probability. GPTZero says it has moved its detector past those two stats into a system with several components.

That blending is why a confusing score can come from a hidden tug-of-war. A “62% likely AI” reading might be one strong human signal and one strong machine signal canceling halfway. You can’t see which is which, so you’re left guessing what to fix. The split is real even when the interface won’t show it to you.

And none of it rises to proof. Every reputable detector outputs a likelihood. OpenAI’s own classifier caught only about 26% of AI text while flagging 9% of human text as AI, and the company retired it in July 2023 for low accuracy. If the people who built the model couldn’t reliably detect its output, treat any blended score as a signal, not a sentence.

What to do when the signals disagree

The disagreement is a diagnosis. Use it.

If your words look human but the rhythm’s flat (high perplexity, low burstiness), stop polishing vocabulary. Go break up your sentences. Follow a long one with a short one. Cut a 25-word sentence into a 12 and a 6. Read it aloud and fix anything that drones. You’re adding variance, which is the thing burstiness measures.

If your rhythm’s varied but your words are bland (low perplexity, high burstiness), add specificity. Swap a generic claim for a concrete example only you would think of. Name the thing. Use the odd-but-right word instead of the safe one. That’s real perplexity, not thesaurus perplexity.

If both are flat, you need both fixes, and that’s actually the easiest case, because varying sentence length while adding concrete detail tends to move both signals at once. Genuine variation is the only thing that satisfies both measurements honestly, because it’s the thing they’re both clumsily trying to detect.

Frequently asked questions

What’s the core difference between perplexity and burstiness?

Scale. Perplexity is local: it looks at each word and asks how predictable that word was, given everything before it. Burstiness is global: it looks across the whole document and asks how much that predictability bounces around from sentence to sentence. Perplexity is the texture of individual word choices; burstiness is the rhythm of the piece. That difference in scale is exactly why the two can disagree on the same text.

Can a piece of writing have high perplexity but low burstiness?

Yes, and it’s common in polished AI drafts that went through a thesaurus or word-swapping tool. Fancier words raise perplexity. But if every sentence is still the same length and shape, the variance stays flat, so burstiness stays low. The result looks human at the word level and machine-made at the structural level, and that mismatch is itself a tell.

If the two signals disagree, which one matters more?

It depends on the detector, and you usually can’t see the weights. Some tools treat sentence-level variance as the stronger structural signal because it’s harder to fake by accident; others lean more on perplexity. The practical move is to fix both: specific word choices and uneven sentence lengths. Genuinely varied writing scores well no matter how the tool balances them.

Do all AI detectors still use perplexity and burstiness?

Many use them as components, not as the whole engine. Early detectors leaned heavily on these two stats. Bigger tools have moved toward trained classifiers that weigh dozens of features at once, with perplexity and burstiness as part of the mix. The underlying idea hasn’t changed: they measure how predictable your words are and how much that predictability varies.

Does fixing one signal help with the other?

Sometimes, not automatically. Adding a vivid, specific example raises perplexity, and writing it as a sentence with a different length than its neighbors also nudges burstiness. But you can raise perplexity with fancy vocabulary while leaving the rhythm dead flat. The reliable fix is varying sentence length and adding concrete detail, which moves both at once because it makes the writing genuinely less uniform.

The short version

Perplexity grades your words. Burstiness grades your rhythm. They’re different rulers, so they can hand you different answers about the same paragraph, and the gap between them is a free diagnosis of what your draft actually needs. Two flat signals mean uniform writing. One flat signal points straight at the fix.

If you want to see how the signals land on something you wrote, and tighten the parts that read as machine-made, run a draft through the free checker. When you want more, you can browse the rest of our writing on AI detection or see the plans.