What Perplexity and Burstiness Mean in AI Detection

When an AI detector says your text "reads like ChatGPT wrote it," it's not running a magic spell. It's measuring two specific statistical properties of your prose: perplexity and burstiness. If you understand what those two things actually are, you can write (or revise) copy that doesn't trigger false positives. You'll also understand exactly why some rewrites work while others don't.

Here's the short version: perplexity measures how predictable each word choice is. Burstiness measures whether your sentence lengths vary the way a human's naturally would. Detectors flag text that scores low on both.

At a Glance

• Perplexity measures how surprised a language model is by your word choices. Predictable writing scores low; distinctive, specific writing scores higher.
• Burstiness measures how much your sentence length varies across a passage. A steady 18-25 word rhythm scores low; a mix of very short and very long sentences scores high.
• AI detectors flag the combination of low perplexity and low burstiness together, not either signal alone.
• Both metrics have real research behind them: GLTR visualizes token predictability, DetectGPT uses probability curvature, and the underlying idea of burstiness in word distributions is documented in peer-reviewed linguistics research.
• Neither score is a verdict. Specialized, technical, or non-native writing can score "AI-like" on both metrics without a model ever touching it.
• The fix isn't randomness for its own sake. It's writing specific claims in your own words and letting sentence length vary the way it does when you actually mean what you're saying.

What perplexity measures

In linguistics and machine learning, perplexity is a measurement of how surprised a language model is by a sequence of words. A lower perplexity score means the text was very easy to predict: each word followed naturally from the one before it. A higher perplexity score means the text surprised the model, which signals more original word choices.

When GPT-4 generates text, it picks the statistically safest word at each step. That keeps output coherent, but it also makes it easy to predict in hindsight. A detector trained on AI-generated text learns to recognize this pattern: the whole paragraph feels like it could only go one way.

Here's a concrete example. Compare these two sentences:

Version A: "It is important to note that artificial intelligence has transformed the way businesses operate in today's digital landscape."

Version B: "Most marketing teams now use AI for copy, and a lot of them can't remember what the workflow looked like before."

A language model would assign a much lower perplexity score to Version A. Every word is a safe, expected choice. Version B uses a contraction and an informal aside that a model is less likely to generate on its own. The word choices are slightly less predictable, so the perplexity score goes up.

This isn't a new idea invented for AI detection. One of the earliest public tools built on it, GLTR (Giant Language model Test Room), color-codes each word in a passage by how likely a model thought it was at that point in the sentence. In the researchers' own study, participants who used GLTR's visual annotations improved their ability to spot generated text from 54% accuracy to 72%, without any prior training on the tool. That's the practical version of what "perplexity" is measuring: a map of which words a model saw coming from a mile away.

If you want a deeper look at how detectors actually use this under the hood, this breakdown of how AI content detectors actually work explains the model architecture involved.

What burstiness measures

Burstiness describes how much sentence length varies within a passage. Humans don't write in steady, even pulses. We write one long, meandering sentence that tries to get everything in, then we stop. Short one. Then another long one with a subordinate clause or two.

Language models tend to produce sentences that cluster around a comfortable middle length. They're rarely very short. They're rarely very long. The output is rhythmically flat, even when the content is interesting.

Burstiness is typically measured by looking at the variance in sentence length across a paragraph or page. High variance (a mix of short and long sentences) looks human. Low variance (sentences all running 18-25 words) looks generated.

The term itself predates AI detection by decades. Researchers studying how words cluster within documents found that content words don't appear at steady intervals. They arrive in clusters, long stretches with none of a given word followed by a burst of repeated use, a pattern formally documented in peer-reviewed corpus linguistics research on the temporal distribution of words in text. Sentence-length burstiness in AI detection is a more specific, modern application of the same underlying idea: real language clusters and varies, and enforced uniformity is itself a signal.

Why flat rhythm is so detectable

The problem isn't that AI sentences are "bad." It's that they're consistent in a way real writers aren't. A human drafting a blog post might get impatient and fire off three words. Or they get on a roll and write a sentence that keeps going through a dependent clause and a parenthetical before finally landing. AI rarely does either.

This is also why simple paraphrasing often doesn't help. If you ask an AI to rewrite its own output, it typically produces similar sentence lengths the second time around. The rhythm stays flat because the underlying generation process hasn't changed.

A worked example

Take a 100-word paragraph. An AI-generated version might have five sentences averaging 20 words each, with the shortest at 17 and the longest at 24, a spread of just 7 words. A human-edited version covering the same ground might run six sentences: one at 4 words, one at 38, and the rest scattered between 9 and 22, a spread of 34 words. Same word count, same information, very different variance. That spread, not the word count, is what a burstiness score is actually computing.

Why both signals are used together

Neither perplexity nor burstiness is reliable on its own.

High perplexity could just mean someone used unusual vocabulary or wrote in a niche domain. A legal brief or a medical abstract might score high perplexity simply because the terminology is specialized, not because a human wrote it. Low burstiness could mean someone wrote in a deliberate, formal style, which is a legitimate choice for certain audiences.

Detectors look for the combination: text that is both very predictable at the word level and very uniform at the sentence level. That pattern rarely appears in human writing except by accident. In AI output, it appears consistently. More recent detection research has moved past raw perplexity scoring toward methods like DetectGPT, which checks whether a passage sits in a region of unusually flat, "negative curvature" probability space rather than just measuring a single perplexity number. On text generated by a 20-billion-parameter model, the DetectGPT paper reported detection accuracy (AUROC) of 0.95, against 0.81 for the strongest simple baseline, which is roughly the difference between catching the overwhelming majority of generated passages and catching about four in five.

It's also worth knowing that this combination creates real problems for false positives. Non-native English writers, people writing in a formal register, and anyone editing their work for clarity can all end up with text that scores low on both metrics. This is one of the reasons AI detectors flag human-written text more often than most people expect.

Perplexity and burstiness at a glance

Metric	What it measures	Low score looks like	High score looks like
Perplexity	How predictable each word choice is to a model	Safe, generic phrasing ("it is important to note")	Specific, informal, or unusual word choices
Burstiness	How much sentence length varies across a passage	Every sentence 18-25 words	A mix of very short and long sentences
Combined AI signal	Both flagged together	Smooth, uniform, "safe" prose throughout	Uneven, opinionated, specific prose

How to write with higher perplexity and burstiness

You don't need to make your writing weird. You need to make it yours.

On perplexity: Use words and phrases you'd actually say. Avoid the phrasing that sounds like a press release ("it is essential to consider," "plays a critical role in"). Be slightly specific where you'd otherwise be generic. Instead of "businesses can benefit from this approach," write "a three-person agency can set this up in an afternoon."

On burstiness: Read your draft aloud. If it has a metronomic beat, break it up. Add a two-word sentence somewhere. Let one sentence run longer than feels comfortable. The rhythm should feel a little uneven, because that's what human writers actually do.

On both at once: The fastest fix is usually to cut your most "AI-sounding" sentence and replace it with something direct and slightly informal. Not sloppy. Just written from a specific point of view rather than a neutral one.

The free humanizer prompt at /humanizer-prompt is built around these principles. It targets both dimensions systematically, which is faster than hand-editing each paragraph alone.

What the scores can and can't tell you

Perplexity and burstiness scores are probabilistic signals, not verdicts. A low combined score means "statistically similar to AI output." It does not mean "definitely written by AI."

Different detectors weight these signals differently. Some also layer in other features: repetition patterns, syntactic uniformity, topic coherence across paragraphs. A text can score poorly on perplexity but high on burstiness and still pass, depending on the tool's thresholds and training data.

The score is also sensitive to the domain. A test on a coding tutorial will produce different numbers than the same algorithm run on a personal essay. Detectors trained primarily on web content sometimes mis-flag academic writing, and vice versa. Before trusting any single score, consider whether the detector is actually reliable for your use case.

None of this means you should ignore the scores entirely. A score of 95% "AI-generated" across multiple independent tools is meaningful data. A single tool giving you 62% is much less so.

---

Frequently asked questions

Does "high perplexity" mean my writing is confusing?

No. In the context of AI detection, perplexity is a technical measurement about word predictability, not about clarity. High perplexity means your word choices were less statistically expected, which is a good thing from a detection standpoint. Clear, readable prose can still have high perplexity if you're making genuine word choices instead of defaulting to the most common phrase.

Can I increase my burstiness score by just adding short sentences randomly?

You can, but it tends to look forced. Detectors now include models trained on edited AI text, and "fake burstiness" (short sentences dropped in mechanically) doesn't always fool them. The better approach is to revise for content first. If a sentence is doing too much, split it. If a paragraph needs a quick transition, let it be quick. Burstiness that comes from genuine editing tends to look more authentic than burstiness you manufacture.

Why do some AI detectors give completely different scores for the same text?

Because they're measuring different things, or weighting them differently. One tool might focus heavily on perplexity using a smaller reference model. Another might weight syntactic patterns or paragraph-level features more heavily, and newer tools like DetectGPT compare probability curvature instead of a single perplexity number. There's no standard formula. This is why a single score from a single tool isn't worth much. Treating any detector result as a final answer is a mistake.

Does using AI to rewrite AI text fix the perplexity and burstiness problem?

Usually not. If you ask an AI to "make this sound more human," it tends to adjust word choice slightly but keeps the same rhythmic patterns. The underlying generation process produces similar sentence length distributions. You can sometimes get marginal improvements, but for consistent results, you need a human editor or a prompt specifically designed to target these signals (like the one at /humanizer-prompt).

Are there types of writing where low perplexity and low burstiness are expected and normal?

Yes. Legal contracts, technical specifications, and certain forms of academic writing are deliberately uniform and predictable. That's a feature, not a bug. Detectors calibrated on general web content often struggle with these genres. If your writing falls into one of these categories and gets flagged, the detector may simply be poorly suited to your text type, not evidence that your writing is AI-generated.

Do perplexity and burstiness measure the same thing as an em-dash or rule-of-three check?

No, and that's worth keeping straight. Em-dash overuse and rule-of-three triplets are surface-level style tells you can spot by eye. Perplexity and burstiness are statistical measurements a model computes, usually invisible to a human reader unless the writing is extreme in one direction. A passage can pass a manual style check and still score low on both statistical metrics, and vice versa. Fixing one doesn't automatically fix the other, which is why thorough editing checks both.