How AI Detectors Actually Work: Perplexity and Burstiness Explained

Every AI detector, from the free ones to the ones your university pays for, is trying to answer the same question: does this text look like it came from a predictable pattern-matcher, or a person? The two oldest ideas for answering that are perplexity and burstiness. They're not the whole story anymore, but they're still the reason AI writing gets flagged in the first place, and understanding them explains a lot about why detectors get it wrong as often as they get it right.

What is perplexity?

Perplexity measures how surprised a language model is by the words in front of it. Feed a sentence through a model and ask it to predict each next word: if the actual word was the model's top guess almost every time, perplexity is low. If the writing keeps picking words the model didn't expect, perplexity is high. Low perplexity means predictable, safe word choices. High perplexity means the writing takes turns a statistical model wouldn't have guessed.

AI text scores low almost by definition. A language model works by picking a statistically probable next word, over and over; that's the whole machine. So when a detector reports low perplexity, it's usually reporting exactly the thing that makes a language model a language model: conservative, expected word choices, one after another. There's no hard-and-fast cutoff, but one commonly cited rule of thumb treats a perplexity above about 85 as human-like and anything below as more machine-like.

What is burstiness?

Burstiness is how that perplexity varies from sentence to sentence within a document. Human writing is never perfectly smooth. Some sentences are simple and direct, some ramble, and some meander before taking a sharp turn no one saw coming. That variation, usually measured as the standard deviation of sentence-level perplexity, is burstiness. Language models are built to be coherent and keep a predictable cadence, so their sentences tend to follow a steady rhythm with a constant level of predictability. Low burstiness, smooth and even, looks artificial. High burstiness, with some sentences mundane and others surprising, looks human.

Put the two together and you get four rough zones. Low perplexity with low burstiness, predictable wording and nothing surprising, is the most telltale sign of AI. High perplexity with high burstiness, unexpected words and wild variation, reads as human. In the middle ground, where most real writing lives, detectors become uncertain.

Where the method breaks down

That middle ground is where it falls apart. Perplexity doesn't detect whether a human wrote a text; it measures how well a particular language model could predict it. Run the Declaration of Independence through a perplexity checker and it may look artificial, because that exact phrasing appears so often in the training data that models have learned to predict it. The same goes for Wikipedia articles. Perplexity isn't finding AI. It's finding commonness.

Perplexity also depends on the model. A word choice can be predictable to one model and not another, and every new model release shifts what counts as predictable. A detector built around what last year's GPT would have predicted doesn't necessarily recognize what this year's GPT predicts. And most commercial models don't expose their word-level probabilities at all.

The most consequential failure is what this does to non-native English writers. A 2023 Stanford study ran 91 human-written TOEFL essays through seven widely used GPT detectors and found they misclassified the essays as AI-generated at an average rate of 61.22%. Almost 98% of the essays were flagged by at least one detector. The reason is structural, not incidental: non-native writers tend to use simpler sentence structures and a narrower vocabulary, which produces exactly the low-perplexity, low-burstiness signal detectors are built to catch, whether or not any AI was involved.

What detectors use now

Detector companies know this, which is why the leading tools have mostly moved on. GPTZero has said publicly that it stopped relying on perplexity and burstiness as its primary method in 2023, shifting to a deep-learning classifier trained directly on examples of AI and human writing, with perplexity kept only as one signal among several. Modern detectors generally work more like a spam filter than a formula: a model trained on huge sets of labeled human and AI text learns deeper patterns, such as word frequency distributions, sentence and paragraph structure and punctuation habits, that go well beyond a single predictability score.

Turnitin works the same way and, since August 2025, also flags AI text that was run through paraphrasing or humanizing tools. We explain how its scores work in our Turnitin AI checker guide, and how strict different detectors are in AI humanizers vs. AI detectors.

What this means for your writing

None of this means detectors are useless (we look at the evidence in how accurate AI detectors really are). It means a single number, taken out of context, is a hint, not grounds for an accusation, particularly for non-native English speakers, people who write in a plain literal style, or anyone writing about a topic that lends itself to standard phrasing. If your writing keeps coming out with the same even rhythm, the fix isn't to write worse. It's to write with more variation: mix short and long sentences, use concrete specifics, and cut stock phrases. That variation is exactly what burstiness is trying to measure.

You can see how your own text scores before anyone else does. HumanizerPRO's free AI detector rates text sentence by sentence, so you can see which lines read as predictable. AI humanizers such as HumanizerPRO work by raising perplexity and burstiness: they rewrite AI drafts with more varied sentence lengths and less predictable phrasing while keeping the meaning. We explain the process in what AI humanization is.

Frequently asked questions

Do AI detectors still use perplexity and burstiness?

Some do, but usually as one signal among several rather than the main method. GPTZero, one of the most widely used detectors, has said it moved to a deep-learning classifier as its primary approach in 2023, keeping the older statistical metrics as supporting signals.

Why do AI detectors flag writing by non-native English speakers?

Because non-native writers tend to use simpler sentence structures and a narrower vocabulary, which produces the same low-perplexity, low-burstiness pattern detectors associate with AI text. A 2023 Stanford study found detectors misclassified real TOEFL essays as AI-generated 61.22% of the time on average.

Can a low perplexity score alone prove text is AI-generated?

No. Perplexity measures how predictable text is to one specific language model, not who wrote it. Text that closely matches common phrasing, like well-known historical documents or Wikipedia-style writing, can score as low-perplexity even when it's entirely human.

What's the difference between perplexity and burstiness?

Perplexity measures how predictable individual word choices are. Burstiness measures how much that predictability varies from sentence to sentence across a whole piece. Low values in both are the strongest signal detectors associate with AI writing.

How can I check the perplexity of my own writing?

Most detectors don't show raw perplexity numbers, but a sentence-level AI detector gives you the same practical insight. HumanizerPRO's free AI detector highlights the sentences that read as most predictable, which are usually the ones worth rewriting.

Sources

• GPTZero, What is perplexity and burstiness for AI detection?
• Pangram Labs, Why perplexity and burstiness fail to detect AI
• Liang et al., Stanford, GPT detectors are biased against non-native English writers (2023)