What Are Perplexity and Burstiness? Why They Matter for AI Detection

Two numbers decide whether Turnitin flags your paper. Perplexity tracks how predictable your word choices are. Burstiness captures whether your sentences march along at the same length or bounce around unpredictably. Get those two metrics wrong and every major detector, Turnitin included, will tag your writing as machine-made before your professor even opens the file.

Last updated: March 26, 2026

Perplexity: How Predictable Is Your Writing?

The name gives it away. Perplexity measures how perplexed a language model feels when it reads a string of text. High perplexity means the model keeps getting surprised by what comes next. Low perplexity means it saw every word coming from a mile off.

Mechanically, here's what happens. A reference model scans your text one word at a time, assigning a probability to each position based on everything before it. Words the model predicted easily contribute low perplexity. Weird, unexpected choices spike it higher. Average all those scores together and you get a document-level number.

Perplexity in Numbers

A January 2026 study out of the University of Maryland measured perplexity across 10,000 documents:

• Human-written academic essays: Average perplexity of 72, range of 55-120
• GPT-4o output: Average perplexity of 23, range of 15-35
• Claude 3.5 output: Average perplexity of 28, range of 18-40
• Gemini output: Average perplexity of 25, range of 16-37

That gap is staggering. Humans land at roughly 3x the perplexity of any major model. No wonder detectors latch onto this signal so aggressively.

Why AI Text Has Low Perplexity

Think about what an LLM actually does. At every token position it picks from a probability distribution, gravitating toward the statistically safest bet. Temperature and sampling shake things up a little, sure, but the output still clusters around likely word sequences. The model is essentially writing text that another model would find completely unsurprising. Because one did write it.

Humans don't work that way. We reach for odd words, chase tangents, throw in half-remembered phrases from a podcast we heard last Tuesday. We delete three sentences and start over. We get distracted by a text message and come back with a totally different angle. All that chaos produces word sequences no language model would confidently predict.

A Concrete Example

Low perplexity (AI-typical): "The study examined the relationship between social media usage and academic performance among college students. The results indicated a significant negative correlation between time spent on social platforms and grade point average."

High perplexity (human-typical): "We looked at whether scrolling through TikTok for three hours a night tanks your GPA. Spoiler: it does, but not for the reasons you'd think. The sleep deprivation wrecks you, not the distraction itself, and that's what actually correlates with lower grades."

Same information, wildly different statistical profiles. The second version uses unpredictable phrasing ("tanks your GPA"), shifts register mid-paragraph, and pivots to a surprise reveal. A reference model would stumble over half those word choices. That stumbling is perplexity doing its job.

For deeper technical detail, see our glossary entry on perplexity.

Burstiness: How Much Does Your Writing Rhythm Vary?

Borrowed from information theory, burstiness originally described data that arrives in clumps instead of a steady trickle. Applied to writing, it measures sentence-length variation across a document. Do your sentences range from 3 words to 40? That's bursty. Do they all hover between 15 and 22? Dead giveaway.

Burstiness in Numbers

Standard deviation of sentence word counts tells the story:

• Human writing: Standard deviation of 8-15 words per sentence
• AI writing: Standard deviation of 2-5 words per sentence

Real writers swing wildly. A 3-word gut punch, then a sprawling 45-word monster, then a 7-word question. LLMs? They cruise at a comfortable 15-25 words per sentence, barely deviating.

Why AI Text Has Low Burstiness

Models generate sequentially and settle into a groove. They never get worked up and fire off four punchy fragments in a row. They never lose the thread and produce a breathless 50-word run-on because they got excited about their own argument. The probability math just favors medium-length outputs, so that's what you get: a metronomic drumbeat of similarly-sized sentences.

Human thinking is messy. We punch a point home in five words. Then we spiral into an over-explained monster of a sentence, parentheses and all (exactly like this one, frankly), before catching ourselves and snapping back to something short. Is that efficient? Not really. But it's unmistakably human, and that ragged rhythm is exactly what detectors hunt for.

A Concrete Example

Low burstiness (AI-typical):

• "Climate change presents a significant challenge to global agriculture." (9 words)
• "Rising temperatures are expected to reduce crop yields in many regions." (10 words)
• "Water scarcity will further exacerbate the difficulties faced by farmers." (9 words)
• "Adaptation strategies will be essential to ensure long-term food security." (9 words)

Sentence lengths: 9, 10, 9, 9. Standard deviation: 0.5. Flatlined.

High burstiness (human-typical):

• "Crops are dying." (3 words)
• "Not metaphorically. Literally withering in fields across sub-Saharan Africa, Southeast Asia, and the American Midwest, where 2025 temperatures exceeded historical averages by margins that would have been called impossible two decades ago." (33 words)
• "Farmers know this." (3 words)
• "The question isn't whether climate change is hurting agriculture but whether we can adapt fast enough to keep 8 billion people fed." (22 words)

Sentence lengths: 3, 33, 3, 22. Standard deviation: 13.9. Wild swings.

How Detectors Combine Both Signals

Neither metric nails it alone. Plenty of human writing has low perplexity (think lab reports, legal filings). And you can prompt an LLM to "vary sentence length," which nudges burstiness upward a bit. So detectors blend both signals with a stack of additional features before making a call.

Turnitin's classifier follows a pipeline roughly like this:

1. Break the document into sentences
2. Calculate perplexity for each sentence using a reference language model
3. Calculate burstiness across rolling windows of 5-10 sentences
4. Compute additional features (entropy, GLTR token distribution, proprietary signals)
5. Feed all features into a trained neural classifier
6. Output a sentence-level probability and aggregate to a document-level score

Low perplexity AND low burstiness together? Flagged with high confidence. One metric in the AI range but the other looking normal? The system hedges, placing it in a "mixed" or "uncertain" bucket. Both metrics in human territory? Clean pass.

The Overlap Zone (and Why It Matters)

Detection isn't a bright line. It's two bell curves that partially overlap. For perplexity, the fuzzy middle ground sits at roughly 35-55. Human technical writing sometimes dips that low; cranked-up-temperature AI text occasionally climbs that high. For burstiness, the overlap lives around a standard deviation of 4-7.

Land squarely in that overlap and even the best classifier struggles to decide. That's why Turnitin labels anything under 20% as "inconclusive." They're admitting their model can't separate the two distributions with enough confidence in that zone to risk an accusation.

Why the Popular "Tricks" Barely Move the Needle

Synonym swapping: Replacing "use" with "employ" sounds different to you, but both words sit in the top-10 most probable tokens at that position. Perplexity barely flinches. You'd have to pick something truly unusual, and then the sentence just reads strangely.

Dropping in filler words: Sprinkling "basically" or "essentially" throughout adds tokens without shifting distributions at all. Those words are themselves among the most predictable in English.

Chopping sentences in half: Splitting one 20-word AI sentence into two 10-word AI sentences doesn't create burstiness. You've just traded one medium monotone for two short monotones. Genuine burstiness needs a 5-word fragment crashing into a 40-word labyrinth.

Paraphrasing tools: QuillBot and its cousins swap surface words but keep the information scaffolding intact, so perplexity stays low. Turnitin specifically trained on paraphrased AI text after their 2025 update and catches it at 85-90%.

What Actually Shifts These Numbers

Only two approaches reliably work.

Write it yourself. Every person's natural writing carries a unique perplexity and burstiness fingerprint. Even working from an AI-generated outline, composing each sentence in your own voice produces human-range statistics. The catch: it takes roughly as long as writing without AI at all.

Run it through a purpose-built humanizer. Tools like Anti-Turnitin specifically reshape perplexity and burstiness distributions. They don't just trade one synonym for another. They rebuild sentence structure, inject statistically improbable phrasing, and engineer the kind of wild length variation that mirrors how people actually write. That's why they work where paraphrasers fall flat. For side-by-side results, check our AI humanizer comparison.

What This Means for You

Perplexity and burstiness sit at the core of every major AI detector on the market right now. One gauges whether your words are predictable; the other gauges whether your sentences march in lockstep. LLMs fail both tests by default because they're engineered to produce fluent, even, statistically safe output.

Knowing this explains more than just how detection works. It clarifies why most workarounds flop and which few approaches genuinely succeed. If a method doesn't touch the underlying statistical distributions, it won't fool a modern detector. Period.

Want text that scores in the human range on both metrics? Try Anti-Turnitin free and see for yourself.