Burstiness and Perplexity Explained for Writers

Perplexity measures how predictable each word is given the words before it. Burstiness measures how much sentence length and structure vary across a passage. Detectors combine those two signals with others to estimate whether text was machine-generated. Neither metric proves authorship, and both can be changed by editing.

Key takeaways

• Perplexity is a predictability score: low perplexity means each next word was easy to guess from context.
• Burstiness is a variation score: low burstiness means your sentences are all roughly the same length and shape.
• Detectors read these as signals, not verdicts — a high score is a probability estimate, not proof.
• Sentence-length variation is the fastest lever you can pull, because it moves burstiness directly and shifts perplexity at the same time.
• Our own measurement on 2026-09-19 found hand-written DeAIze guides averaged an AI score of 27% (n = 46), while unedited model output averaged 35% (n = 12) — the overlap is the point.
• You cannot edit your way to a guaranteed number on any detector, and you should not try.

What is perplexity in AI detection?

In language modelling, perplexity is a measure of surprise. A model assigns a probability to each word based on everything before it. If the model expects “the” with high confidence, that word contributes little surprise. Perplexity aggregates those probabilities across a passage into a single number: low perplexity means the text was easy to predict, high perplexity means it was hard to predict.

A detector uses a language model — often a smaller, faster one than the model that wrote the text — to score your draft the same way. Text a model finds easy to predict tends to look machine-generated, because generative models are optimised to produce likely continuations. Human writing contains more local surprises: a sudden concrete noun, an aside, a clause that does not follow the expected path.

That is why a fluent, polished paragraph can score as machine-made. Fluency and predictability are close cousins. A writer who has internalised the rules of clean prose — vary nothing, connect everything, avoid surprises — can produce text with genuinely low perplexity without ever opening a chatbot.

Perplexity is also why short texts are unreliable. A two-sentence sample gives the model almost no context to be surprised by, so the estimate swings wildly. If you are checking a fragment, expect noise.

What is burstiness in AI detection?

Burstiness is about rhythm. Take the lengths of your sentences and look at how much they differ. A passage where every sentence runs 18 to 22 words has low burstiness. A passage that goes 6, 31, 12, 44, 9 has high burstiness.

Human writers are bursty by accident. We write a long, qualified sentence when an idea needs unpacking, then a short one to land it. We interrupt ourselves. We front-load a clause because it sounded better that way. Machine output tends toward a steady mid-length cadence, because the model is sampling from a distribution that rewards safe, complete sentences.

Burstiness is easier to see than perplexity because you can literally count it. Open your draft, put the sentence lengths in a column, and look at the shape. If the numbers cluster, you have a burstiness problem regardless of who or what wrote the text.

If you want to see burstiness on your own draft rather than guess at it, the AI detector highlights every sentence by AI probability, so you can see exactly which lines carry the signal before you change anything.

How does sentence-length variation move an AI score?

Sentence length is the single easiest variable to control, and it feeds both metrics.

When you alternate long and short sentences, you raise burstiness directly. You also raise perplexity indirectly: a very short sentence after a long one is harder for a language model to predict, because the expected continuation was another long sentence. The two metrics are not independent. Variation in structure produces surprise in word choice.

Here is a small worked illustration, not a real document:

Before (flat cadence):

The study examines the relationship between urban density and transit use across several metropolitan regions. The analysis draws on survey data collected over a multi-year period. The findings suggest that density alone does not predict ridership patterns. The authors recommend further research into local policy contexts.

Every sentence is a similar length and the same shape: subject, verb, object, qualifier. It reads like a summary because it is one.

After (varied cadence):

Density alone does not predict transit use. That is the finding, and it survives several metropolitan regions and years of survey data. What does predict ridership? Local policy. The authors want more research on that, which is a polite way of saying the density argument has been oversold.

Same claims. The second version asks a question, uses a three-word sentence, and lets one sentence run long. Burstiness is higher, and a model scoring it word by word has more places to be surprised.

Neither version is inherently more honest. The second is simply less predictable.

Why does unedited model output still score low sometimes?

This is the part most explanations skip. In our own measurement on 2026-09-19, we scored 12 unedited model passages against 46 hand-written DeAIze guides using the DeAIze detector at a 30% flag threshold. The unedited model output averaged an AI score of 35% (median 34%, range 25-49%). The human-written guides averaged 27% (median 25%, range 0-55%).

The distributions overlap. Some human guides scored above some model passages. That is the honest picture: perplexity and burstiness are signals, and signals overlap. A detector that reported certainty from them would be lying to you.

We also counted AI-tell frequency in the same run: 3 per 100 words in the human sample versus 1 per 100 words in the unedited model output. The human writers used more of the stock phrases people associate with machine writing — “moreover”, “it is worth noting”, connective scaffolding — than the model did. Tells are a weak signal on their own, and they can point the wrong way.

If you want the broader picture of how these signals combine with others, how AI detection works covers the full stack.

Concrete editing moves that change perplexity and burstiness

These are the moves that actually move the numbers. Do them on your own draft.

1. Count your sentence lengths. Put them in a list. If the spread is narrow, that is your first fix.
2. Break one long sentence into three. Not into two — three, so one of them is very short.
3. Add a question where the text is explaining. A direct question raises perplexity because the expected continuation was a statement.
4. Replace one abstract noun with a specific one. “Several factors” becomes “parking minimums”. Specificity is unpredictable.
5. Cut one connective. “Moreover”, “furthermore” and “additionally” are the most predictable words in any paragraph.
6. Vary your openings. If three consecutive sentences start with the subject, invert one.
7. Read it aloud. Your ear catches flat cadence faster than your eye.

Edit	Effect on burstiness	Effect on perplexity
Split a long sentence into three	Large increase	Moderate increase
Add a direct question	Moderate increase	Large increase
Replace abstract noun with concrete one	Small	Moderate increase
Delete a connective	Small	Small increase
Front a clause for emphasis	Moderate increase	Moderate increase
Rewrite every sentence to the same length	Large decrease	Large decrease

Notice the last row. You can lower your score’s readability and raise your AI score at the same time. Flat prose is not automatically human-looking; it just looks flat.

What are the limitations of these two metrics?

Perplexity and burstiness are useful because they are inspectable. You can count sentence lengths. You can look at where a model was surprised. That is more than you get from a black-box percentage.

But they are not a verdict, and treating them as one causes real harm.

They penalise clean writing. A careful editor who standardises sentence length for a house style is producing low-burstiness text. So is a non-native English writer applying textbook rules. That is a known failure mode, and it is worth reading why an AI detector flags non native English writers if that describes you.

They are unstable on short text. A 200-word sample cannot support a confident estimate. Neither can a single paragraph.

They do not know who wrote the words. Perplexity measures predictability. A predictable paragraph written by a human is still predictable.

Detectors disagree with each other. Different models, different thresholds, different training data. If two tools give you different readings, that is expected behaviour, not evidence that one is broken. Are AI detectors accurate? goes through why two tools disagree on the same essay.

Editing for the metric is not the same as editing for the reader. Chasing burstiness can produce choppy, self-conscious prose. If a sentence reads better long, leave it long.

A practical routine for checking your own draft

Here is a sequence that takes about ten minutes on a normal article.

1. Run the draft through a detector that shows sentence-level highlighting, not just a document score. The AI detector does this: one score for the whole document, plus a per-sentence breakdown.
2. Look at which sentences carry the signal. Usually it is a run of similar-length statements in the middle of an explanatory section.
3. Rewrite only those sentences. Leave everything else alone — you are trying to fix a cadence problem, not launder an entire document.
4. Re-score. If the flagged run is gone, stop. Do not chase a number.
5. Check that the meaning survived. This is the step people skip. A rewrite that changes your claim is worse than a flag.

That last point matters more than any metric. If you used AI to draft a section, the ideas, data and conclusions still have to be yours. Rewriting for burstiness does not make someone else’s argument your own, and no detector score changes that.

If your concern is a formal process rather than a draft, the academic integrity policy guide explains what institutions typically assess.

Where burstiness and perplexity fit in the bigger picture

Think of them as two dials on the same panel. Perplexity tracks word-level predictability. Burstiness tracks structural variation. Detectors combine them with token patterns, repetition measures and other features, then output a probability.

That probability is an estimate. It is not proof of authorship, and no responsible tool should present it as one. DeAIze reports an average AI-score reduction of 60%+ after humanizing on our own predicted scores, with a best case of 71%, and on the example paragraph used on the site a 92% AI score becomes 4% AI. Those are our predicted scores, not an official verdict from any third-party detector, and they will not match what every tool reports.

What you can rely on is the mechanism. If you understand why a flat paragraph scores high, you can fix it deliberately, on your own terms, without guessing.

Ready to see which sentences read machine-made?

Reading about perplexity is not the same as seeing it on your draft. The fastest way to learn the mechanism is to watch it operate on text you wrote.

Start with the AI detector to get a sentence-level breakdown, then use the draft editor to rewrite only the flagged lines and re-score. If you want to see how the loop works before committing, pricing lists the pay-as-you-go tiers with credits that never expire, and you can register for 200 free words with no credit card.

Run it on one paragraph. If the highlighted sentences are the ones you already suspected were flat, you have learned the mechanism. That is the point of the exercise.