How AI detection works: perplexity, burstiness, and why your writing gets flagged

Imagine you're a teacher reading two essays. One is stiff, predictable, every sentence roughly the same length. The other is messy: a three-word sentence followed by a sprawling thirty-word one, with unusual word choices sprinkled in. You don't need a computer to tell you which one feels human.

AI detectors do something similar, but with math. They measure two things: perplexity and burstiness. Neither is complicated once you see how they work. And once you understand them, you'll know exactly why your perfectly human writing keeps getting flagged, and what to do about it.

This guide breaks down the technology behind every major AI detector, from GPTZero to Turnitin to Originality.ai. No jargon, no marketing claims. Just how the math works, where it fails, and what actually matters.

What is perplexity in AI detection

Perplexity measures how surprised a language model is by each word in your text. If the model can easily predict the next word, perplexity is low. If it keeps getting surprised, perplexity is high.

Here's a simple example. Low perplexity: For lunch, I had a bowl of soup. High perplexity: For lunch, I had a bowl of spiders. The word soup is predictable. Almost any language model would guess it after bowl of. The word spiders isn't. That surprise is what perplexity measures.

Now here's why this matters for detection: AI language models are built to pick the most probable next word. That's literally their job. So when an AI generates text, it produces consistently low-perplexity output. Every word is the statistically safest choice. Humans don't write like that. We choose weird words, interrupt ourselves, use idioms that make no literal sense. All of that creates higher perplexity. Detectors treat low perplexity as a signal of AI writing.

GPTZero built its early detection engine entirely around this idea. It calculates perplexity at the sentence level and flags any passage where the score stays uniformly low across the whole document. A human writer almost always has perplexity spikes. An AI almost never does.

What is burstiness in AI detection

If perplexity is about word choice, burstiness is about rhythm. It measures how much your sentence structure varies throughout a piece of writing. Humans write in bursts. We fire off a three-word sentence. Then a 40-word sentence with two clauses and a parenthetical. Then something short again. That variation is burstiness.

AI doesn't do this naturally. Language models tend to produce sentences in a narrow band, usually 15 to 20 words each, with remarkably consistent paragraph structures. The rhythm is flat. That flatness is what detectors call low burstiness. Under the hood, they calculate it by looking at the standard deviation of sentence lengths. Small deviation means all your sentences are roughly the same size. That's a flag.

This is why adding a few very short sentences to AI-generated text sometimes drops the detection score. You're not making the content better. You're just increasing the burstiness, which tricks the statistical model into reading the text as more human.

Classifier models: the third detection method

Perplexity and burstiness are statistical heuristics. Classifier models are something different: machine learning systems trained on millions of examples of human and AI text. They learn patterns that go beyond word predictability and sentence length. A classifier looks at word frequency distributions, syntactic structures, paragraph organization, transition patterns, even punctuation habits.

Turnitin uses a classifier trained specifically on academic writing. Originality.ai targets marketing and web content. Copyleaks runs an ensemble of classifiers combined with their plagiarism engine. The tradeoff: classifiers are more sophisticated than simple perplexity scoring, but they're only as good as their training data. Every new model release creates a gap where detection accuracy drops until the classifier is retrained.

Watermarking: the emerging fourth method

Some AI providers are working on a different approach: embedding invisible statistical watermarks into model output during generation. These watermarks work by subtly biasing token selection. The bias is invisible to a reader but detectable by the provider's own verification tools. OpenAI confirmed it has this capability but delayed rollout, citing concerns about impact on non-English speakers. Google DeepMind's SynthID is already live for some Gemini outputs.

Watermarks survive light editing but break down when text is substantially rewritten. And they only work if the provider implements them. Most API access and all open-source models don't include watermarks. For now, watermarking is a promising direction, not a deployed solution.

Where AI detectors fail

Here's the uncomfortable truth: the same statistical signals that flag AI writing also flag plenty of human writing. This isn't a rare edge case. It's a structural problem with how these tools work.

Non-native English writers are the most obvious victim. A 2023 Stanford study published in Patterns tested seven major detectors on 91 TOEFL essays written by non-native speakers. The result: 61% of those human-written essays were flagged as AI-generated. Non-native speakers tend toward simpler vocabulary and more uniform sentence structures, the exact patterns detectors are trained to catch.

Famous texts get flagged too. Perplexity-based detectors classify the US Declaration of Independence and passages from the Bible as AI-generated. The reason: these texts appear verbatim in AI training data thousands of times. Language models assign them extremely low perplexity, so they look indistinguishable from machine output. Formal writing gets caught for the same reason. Academic papers, legal documents, technical manuals, they all tend toward the low-perplexity, even-rhythm pattern that detectors flag.

OpenAI retired its own AI text classifier in July 2023, just six months after launch, citing a low rate of accuracy. If the company that builds the models couldn't build a detector that worked reliably, that tells you something about how hard this problem actually is.

We've written more about detector reliability and specific tool accuracy in our posts on whether AI detectors are reliable and AI detector false positive rates. The short version: none of these tools should be treated as a verdict.

How to avoid getting flagged by AI detectors

If you understand how detectors work, you understand how to stay ahead of them. The patterns they look for are mechanical. Here's what to change if your writing keeps getting flagged.

Vary your sentence length aggressively. Write a two-word sentence. Then a long one with nested clauses and a sudden turn. The wider your sentence length range, the higher your burstiness score. This alone can drop detection by 20 to 30 percentage points on GPTZero.

Use specific, unexpected words. AI picks the most probable word. You should pick the most precise one, even if it's unusual. Instead of important, say consequential. Instead of big, say sprawling. Instead of good, say whatever you'd actually say to a friend. The unexpected word choices spike your perplexity.

Add personal voice and break structural uniformity. Use contractions. Start sentences with And or But or So. Include specific opinions. Reference personal experience. Write a one-sentence paragraph. Then a long one. Skip the transition entirely sometimes and just move to the next point. The inconsistency is the signal. Classifiers are trained to spot the absence of personality, and AI writing tends toward even paragraph lengths with predictable transitions. Do the opposite.

Or use a tool. If you're generating content with AI and need it to pass detection, run it through a humanizer. These tools specifically target the statistical patterns detectors look for, increasing perplexity and burstiness while preserving meaning. Our AI humanizer accuracy comparison tested the major options side-by-side so you know which ones actually deliver.

Why understanding detection matters

AI detection isn't going away. Schools are embedding it into plagiarism workflows. Publishers are using it to screen submissions. Google is factoring it into content quality signals. Knowing how these tools work isn't a nice-to-have. It's a basic literacy skill for anyone who writes online in 2026.

But the deeper point is this: detectors don't measure whether writing is good. They measure whether it's statistically predictable. A detector flagging your text doesn't mean you sound like a robot. It means your word choices and sentence rhythms fall into a pattern that machines also fall into. Sometimes that's because a machine wrote it. Sometimes it's because you write clearly and simply. The detector can't tell the difference. The best defense is to understand the signals and write with them in mind. Not to trick the detector, but to write more like yourself. That's the whole point.

Frequently asked questions