Burstiness and Perplexity: How AI Detectors Actually Work

Every AI detector you've used — GPTZero, Originality.ai, Turnitin, aicheckr.io — leans on two statistical measures: perplexity and burstiness. Understanding what they actually measure tells you a lot about why detectors flag what they flag, why they sometimes get it wrong, and what kind of writing fools them.

Perplexity, in Plain English

Perplexity asks one question: How surprising is each word?

Imagine you've read 99 words of a sentence. You know the topic, the tone, the grammar around the next slot. A language model can give every possible 100th word a probability. If the actual 100th word is one of the most likely candidates, perplexity is low. If it's a rare or unexpected choice, perplexity is high.

Average perplexity across a whole document is the headline number. Low perplexity = predictable text. High perplexity = surprising text.

Why this is an AI signal

LLMs work by picking high-probability next tokens. By design, their output has low perplexity — it's the literal optimization target during training. Human writing tends to have higher perplexity because humans pick words for meaning, mood, and personal style, not just statistical fit.

So a passage with very low average perplexity is statistically unlikely to come from a human writer working freely. It's more likely to come from a model.

Where this breaks

Some humans write low-perplexity prose:

• Non-native English speakers using a smaller vocabulary range (see our piece on non-native speaker bias).
• Technical writers in highly conventional genres — legal briefs, scientific abstracts, regulatory filings.
• Anyone whose writing has been heavily edited by Grammarly or another smoothing tool.

Their writing looks "AI-ish" by perplexity alone. That's where false positives originate.

Burstiness, in Plain English

Burstiness asks: How much does sentence complexity vary?

Take a paragraph and measure each sentence's length and complexity. Now look at the spread. If sentences vary wildly — a 6-word punchline, a 35-word complex sentence, a fragment — that's high burstiness. If sentences cluster around the same length and structure, that's low burstiness.

Why this is an AI signal

LLMs trained on web text default to medium-length, syntactically similar sentences. Without explicit prompting, GPT-4 and earlier models produce remarkably uniform sentence rhythms. Humans don't. We start a paragraph with a long, careful sentence. Then we cut. Then we ramble for a bit, double back, contradict ourselves slightly. Real writing is bursty.

Where this breaks

Burstiness is much weaker as a signal than perplexity, for two reasons:

• Genre matters. Bullet-point lists and technical documentation are intentionally low-burstiness. They're not AI — they're just lists.
• GPT-5 fixed it. Newer models deliberately introduce sentence-length variation as part of RLHF training. See our GPT-5 detection guide for what changed.

How Detectors Combine the Two

The classic detector decision rule, simplified:

• Low perplexity + low burstiness → flag as AI.
• High perplexity + high burstiness → flag as human.
• Mixed signal → return a confidence score in the middle.

Modern detectors layer on additional features — token-level distribution analysis, watermark detection, cross-sentence consistency — but perplexity and burstiness are still the foundation. When a detector says "78% AI," that score is largely driven by these two measurements.

Why Sentence-Level Scoring Matters

Most detectors compute perplexity and burstiness across the whole document and return one average. That's a problem. A 1,000-word essay can have:

• Three GPT-pasted paragraphs in the middle (low perplexity).
• Five paragraphs of original human writing (normal perplexity).
• A whole-document score that says "32% AI" — vague enough to be useless either way.

Sentence-level detectors compute perplexity and burstiness per sentence. They can tell you which three paragraphs are flagging, which is what you actually need to know — whether you're an editor cleaning up a draft or a teacher investigating an essay.

This is where aicheckr.io diverges from whole-document tools: the highlighting shows you the exact sentences that are statistical outliers, so you can act on them rather than guess.

What Beats Perplexity-Based Detection

If you're trying to write text that passes detectors, the strategies that actually work are the ones that raise perplexity and burstiness:

• Replace common words with less-common synonyms (raises perplexity).
• Add specific details — names, dates, internal references — that the model wouldn't have generated (raises perplexity).
• Vary sentence length deliberately. Short. Then long, complex, wandering sentences. Then medium. (Raises burstiness.)
• Insert minor stylistic accidents: a contraction, an interrupted thought, a parenthetical aside.

Our humanizer does this systematically: it rewrites sentences to raise per-sentence perplexity and burstiness without changing meaning.

The Limits

Perplexity and burstiness are statistical, not magical. They give you a probability, not a verdict. A reasonable detection workflow:

1. Treat any single score as evidence, not proof.
2. Look at the distribution across sentences, not just the average.
3. Cross-check with a second detector that uses different methodology.
4. Always pair detection with corroborating evidence (draft history, in-class writing samples).

Bottom Line

If you understand perplexity and burstiness, you understand 80% of how AI detectors actually work. The remaining 20% is engineering polish. The two-signal foundation explains both why detectors are surprisingly good — modern LLMs really do produce statistically distinct text — and why they're fallible against well-edited human writing or newer models that deliberately add variation.