How GPTZero Works: Perplexity, Burstiness, and What Its Score Really Measures

GPTZero is probably the AI detector you’ve heard of even if you’ve never used it. A student named Edward Tian built the first version over a winter break in 2022, and since then teachers, editors, and hiring managers have leaned on its score to answer one question: did a person write this? The catch is that GPTZero never actually answers that question. It estimates it.

This is a look under the hood. What the tool measures, how to read the number it spits out, what the highlighted sentences are telling you, and where the whole thing is more fragile than its confident percentage suggests.

Key takeaways

• GPTZero runs your text through its own language models and scores how predictable and how varied the writing is, then converts that into a probability that the text is AI-generated.
• Perplexity and burstiness are the starting point, but the current product layers a trained classifier and sentence-level analysis on top.
• The big number is a likelihood, not a verdict. GPTZero itself frames results as probabilities and warns against using them as the sole basis for punishment.
• The sentence highlights are often more useful than the headline percentage, because they show you exactly which lines read as predictable.
• Plain, formulaic, or non-native English writing can score high without any AI involved, so a flag is a starting point for a conversation, not proof.

What GPTZero is actually doing

Here’s the thing people get wrong first. GPTZero isn’t comparing your essay against a giant library of known ChatGPT outputs. There’s no fingerprint database. Instead, it uses its own language models to ask, over and over, a single question: how surprising is this writing?

That sounds vague, so let’s make it concrete. A language model reads text one chunk at a time and constantly predicts what comes next. Feed it “the data clearly shows that more research is” and it expects “needed” with high confidence. When the actual next word matches what the model expected, that’s low surprise. When you write something it didn’t see coming, that’s high surprise. GPTZero rolls all of that up into a score.

The reason this works at all is that current AI models tend to pick safe, high-probability words. Left to its own devices, a chatbot writes the unsurprising version of every sentence. So when GPTZero sees text its models could have produced almost word for word, it leans toward “AI.”

Perplexity and burstiness, briefly

Two terms do most of the heavy lifting, and GPTZero has used both publicly to explain itself.

Perplexity is the surprise measurement. Low perplexity means the writing was easy to predict. High perplexity means it kept the model guessing. GPTZero has described perplexity as roughly how likely its model would have been to choose the same words you did. Predictable writing scores low, and low reads as machine-made.

Burstiness measures how much that predictability bounces around within a single piece. Humans are uneven. We write a long, clause-stuffed sentence, then stop short. One word. Then we recover with something medium. Plotted out, our writing looks jagged. Models tend to hold a steady rhythm, so their output graphs as a flat line. High burstiness looks human; flat looks automated.

If you want the deep, math-free version of these two ideas with worked examples, we wrote a whole piece on the statistics behind AI detectors. For this article, just hold onto the gist: GPTZero is measuring how predictable and how varied your writing is, then turning that into a probability.

What’s been bolted on top

The early GPTZero leaned heavily on those two raw stats. The current product is more elaborate. GPTZero now describes its detector as using several components, including a trained classifier that weighs many features at once and models tuned for different kinds of writing. It also analyzes your document at the sentence level, not just as one blob.

That’s genuinely more sophisticated than perplexity alone. It does not, however, change the fundamental move. Every layer is still some version of “how predictable is this, and does that predictability vary the way a human’s would?” The output is still a probability.

How to read a GPTZero result

Open a result and you’ll usually see three things. Knowing what each one means keeps you from over-reading the headline.

The big percentage

This is the part everyone screenshots, and it’s the part most often misunderstood. A result like “92% probability AI” does not mean 92% of your words came from a bot. It means GPTZero’s model estimates a 92% likelihood that the passage is AI-generated, based on how the writing reads statistically. It’s a confidence level about a guess, not a measured fraction.

GPTZero typically buckets this into something like human, mixed, or AI rather than asking you to parse the raw decimal. That’s a reasonable simplification, but the bucket inherits all the uncertainty of the number underneath it.

The sentence highlights

This is the most useful and least screenshotted part. GPTZero highlights the specific sentences it considers likely AI. If you actually want to improve a draft, or understand why it got flagged, the highlights tell you far more than the percentage.

Watch what gets lit up and a pattern emerges fast. The flagged sentences are almost always the flattest ones: generic claims, textbook transitions, even-length lines with safe word choices. The sentences left alone tend to be the specific, oddly phrased, personal ones. That’s perplexity and burstiness made visible, one line at a time.

The mini-scenario

Picture a student, Maya, who wrote her own history essay but writes very plainly because she was taught to keep sentences clear and short. She pastes it into GPTZero and gets back 78% AI. Panic.

Look at the highlights, though, and the story changes. The flagged sentences are her topic sentences, which all follow the same “X was important because Y” shape, in roughly the same length, with no surprises. The paragraphs where she described a specific primary source in her own voice came back clean. Nothing about this means Maya cheated. It means a chunk of her writing happens to be predictable and even, which is exactly what the math punishes. The score is a description of style, and it’s quietly mistaking style for authorship.

Where GPTZero’s score gets shaky

A confident percentage hides a lot of fragility. Three things are worth knowing before you trust one.

Plain human writing scores like AI. The detector rewards surprise and variation. But mountains of legitimate human writing are predictable on purpose: technical docs, legal boilerplate, careful beginner writing, anyone trained to be clear and consistent. Low perplexity, flat burstiness, fully human. GPTZero can’t see the difference, because there isn’t a statistical one.

It can penalize non-native English writers. This is the failure that should give everyone pause. A 2023 study by Liang and colleagues, published in *Patterns*, found that AI detectors as a group were biased against non-native English writers: more than half of a sample of TOEFL essays got wrongly flagged as AI-generated, while essays by US eighth-graders were classified roughly correctly. The mechanism is exactly what you’d predict: a smaller working vocabulary and more conventional phrasing produce lower perplexity, and low perplexity reads as “machine.” The tool ends up penalizing a linguistic background, not catching misconduct.

Even the model-makers couldn’t make this reliable. OpenAI shipped its own AI-text classifier in early 2023 and pulled it that July, citing low accuracy. On its own challenge set, the tool correctly flagged only about 26% of AI-written text while mislabeling 9% of human text as AI. If the company that builds the models can’t reliably detect their output, that resets what you should expect from any third-party detector, GPTZero included.

To its credit, GPTZero is upfront that results are probabilities and shouldn’t be the sole basis for an accusation. That honesty is the whole point. The number is a signal to look closer, not a conviction.

So how should you use the score?

If you’re checking your own writing, ignore the headline and work the highlights. The lit-up sentences are your flat spots. Vary your sentence lengths on purpose, swap a generic claim for a concrete detail only you would know, read the thing aloud and fix whatever drones. That genuinely lowers the score because it genuinely makes the writing less uniform, and it happens to make it better for human readers too.

If you’re an evaluator, treat a flag as the start of a conversation, not the end of one. Look at drafts, version history, and whether the writer can actually talk through their own argument. A probability paired with a real discussion tells you far more than a probability alone ever could.

Frequently asked questions

How does GPTZero decide if text is AI?

It feeds your text through its own language models and measures how predictable each stretch is (perplexity) and how much that predictability varies across the document (burstiness). A trained classifier then weighs those signals, plus others, and outputs a probability that the passage is AI-generated. It also flags individual sentences it finds suspicious. There’s no database of “known AI essays” being matched against; it’s a statistical judgment about how the writing reads.

What does a GPTZero score actually mean?

The headline percentage is GPTZero’s estimated likelihood that the text is AI-generated, not a measured fact. An 85% result means the model thinks the writing is statistically similar to AI output, not that 85% of the words came from a bot. GPTZero’s own guidance treats results as probabilities and cautions against using them as the only evidence in an academic case.

How accurate is GPTZero?

It catches a lot of straightforward AI text, but accuracy depends heavily on the writing. Highly formulaic human prose, technical writing, and essays from non-native English speakers can score high with no AI involved. A 2023 study by Liang and colleagues in *Patterns* found AI detectors as a category were biased against non-native English writers, wrongly flagging more than half of a sample of TOEFL essays while classifying native-speaker essays roughly correctly. No detector, GPTZero included, is reliable enough to treat as proof on its own.

Can GPTZero be wrong about human writing?

Yes, and routinely. The same low-perplexity, even-rhythm pattern that signals AI also describes a lot of plain, careful, or rule-following human writing. If you write in short uniform sentences with common word choices, you can score as AI without touching a chatbot. That’s a limitation of measuring style with statistics, not a sign you did anything wrong.

Should a teacher or editor rely on a GPTZero score alone?

No. GPTZero itself says results shouldn’t be the sole basis for an accusation. Treat the score as one signal alongside drafts, version history, the writer’s familiarity with their own argument, and a conversation. A probability is a prompt to look closer, not a conviction.

The short version

GPTZero works by measuring how predictable and how varied your writing is, then turning that into a probability and a set of sentence highlights. It’s a clever tool and a genuinely imperfect one, and the number it gives you is always a guess dressed up as a verdict. Use the highlights to find your flat spots, use judgment for everything else, and never let a percentage stand in for a real reader.

Curious what your own draft looks like through this lens? Run a draft through the free checker and watch which sentences light up. The flagged ones are almost always the flattest, which tells you exactly where to add variation and specificity. You can also browse the rest of the blog for more on how detection works, or compare plans if you want the paid tools.