How GPTZero Computes Its Score: Perplexity, Burstiness, and the Sentence Highlighter

Paste a paragraph into GPTZero and you get a verdict in seconds: a percentage, a label, and a wash of highlighted sentences. It feels like the tool *read* your writing. It didn’t. It measured two statistical properties of your text and compared them to patterns it has seen before.

Once you understand those two measurements, the score stops being a black box. You can see why it landed where it did, and you can tell when it’s wrong.

Key takeaways

• GPTZero leans on two statistical signals: perplexity (how predictable your word choices are to a language model) and burstiness (how much your sentence length and complexity vary).
• Low perplexity plus low burstiness pushes the score toward “AI.” Predictable wording and uniform sentences are the pattern detectors associate with machine text.
• GPTZero reports at two levels at once: a document-level verdict and a sentence-level highlighter that marks which specific sentences look machine-written.
• The output is a probability and a classification, not proof. Human writing can score high; edited AI can score low.
• You can lower a high score honestly, by varying sentence length, adding concrete specifics, and writing in your own voice, not by chasing a number with tricks.

The two numbers under the hood

GPTZero’s score comes mostly from two ideas. Neither one is magic, and both have plain-English versions.

Perplexity: how surprised a model is by your next word

Perplexity asks a simple question: given everything written so far, how predictable is the next word? A language model assigns a probability to each possible next token. When your actual word is one the model expected, perplexity is low. When you reach for something the model didn’t see coming, perplexity goes up.

Here’s the catch for AI detection. Large language models are *built* to produce the most probable next word. That’s their whole job. So text generated by a model tends to have low perplexity, because the model is essentially picking the words it already finds most likely. Human writing wanders more. We pick odd verbs, drop in a slang term, start a sentence in a way a model would rate as unlikely.

A quick illustration. Compare these two openers:

> “The Industrial Revolution was a period of significant technological advancement that transformed society.”

> “Steam ruined Sunday. Factories didn’t care what day it was.”

The first sentence is fluent and correct, and almost every word is the obvious choice given the one before it. Low perplexity. The second is jagged and specific, full of words a model wouldn’t rank first. Higher perplexity. To GPTZero, the second *looks* more human, even though a person could easily have written both.

Burstiness: the rhythm of your sentences

Perplexity works at the word level. Burstiness works at the sentence level. It measures variation, the spread of sentence lengths and complexity across a passage.

Human writing is bursty. We write a long, clause-heavy sentence that builds and qualifies and circles back, and then we stop. One word. We mix it up without planning to. AI drafts, especially straight out of a chatbot, tend to be smooth: sentence after sentence of similar length, similar structure, similar grammatical shape. Low burstiness.

Think of it as a heart-rate monitor. A human document spikes and dips. A machine document looks more like a flat, even pulse. GPTZero treats that flatness as a signal.

For the underlying math, including how these are actually computed, see our deeper dive on the statistics behind perplexity and burstiness.

How the two combine into one score

GPTZero doesn’t just average perplexity and burstiness. It uses them together, alongside other model-derived features, to classify text and output a probability that the passage is AI-generated.

The intuition is that the two signals reinforce each other. Text that is *both* predictable word-to-word *and* uniform sentence-to-sentence is the classic machine fingerprint. Text that’s predictable but bursty, or surprising but uniform, is more ambiguous, and the verdict gets less confident.

That probability then maps to a label you see in the interface, something like “likely human,” “likely AI,” or “mixed.” The exact thresholds are GPTZero’s to set, and they’ve changed over versions. What matters for you is that the words on screen are a translation of a probability, not a binary fact the tool discovered.

The sentence highlighter: scoring at two levels

The feature people notice most is the highlighter. GPTZero doesn’t only score your whole document; it runs a classification at the sentence level and shades the sentences it judges most likely to be machine-written.

This is why you’ll sometimes see a document rated “likely human” overall, with two or three sentences glowing anyway, or the reverse. The document score and the sentence highlights come from related but separate passes over your text.

A mini-scenario

Say you’re a student who drafted an essay yourself, then asked a chatbot to “smooth out” your conclusion. You paste the whole thing into GPTZero.

Your body paragraphs, full of specific examples, a misremembered date you later corrected, one sentence that runs too long, come back clean. Then the conclusion lights up. Every sentence in it is roughly the same length. The vocabulary is generic (“in conclusion, these factors demonstrate the importance of…”). Perplexity is low across the board, and burstiness inside that paragraph is nearly flat.

The highlighter just told you exactly which paragraph changed character. It didn’t read your intentions. It noticed that one block of text suddenly got smooth and predictable while the rest stayed jagged.

That’s genuinely useful information, as long as you read it correctly. The highlight isn’t proof that paragraph was AI-generated. A naturally tidy summary you wrote yourself could trigger the same pattern. It’s a flag, not a verdict.

What the score is, and what it isn’t

A GPTZero score is a probability and a classification. It is not evidence, and it is not a measurement of whether you personally typed the words.

Three things worth keeping in front of you:

• Detectors output probabilities. A “98% AI” reading means the model’s features placed your text deep in the range it associates with machine output, not that there’s a 98% chance you cheated.
• False positives are real. A 2023 Stanford study led by Weixin Liang found several detectors disproportionately flagged writing by non-native English speakers. In that study, the essays that got misclassified tended to have *lower* perplexity, the same plain, consistent vocabulary that trips the detectors. Clear, formulaic human writing can set off the same alarm.
• The field is humble about accuracy. OpenAI retired its own AI text classifier in July 2023, citing a low rate of accuracy. If the company that builds these models couldn’t reliably detect their output, it’s worth treating any detector’s verdict with caution.

For more context on how detection actually works across different tools, browse more on how AI detection works.

Writing that reads as human, honestly

If you understand the two signals, the fix for a high score isn’t a trick. It’s better writing.

• Vary your sentence length on purpose. Follow a long, winding sentence with a short one. Burstiness rewards real rhythm.
• Reach for the specific word. Names, numbers, places, small concrete details. Specificity raises perplexity in the way human writing naturally does, because precise details are surprising to a model trained on averages.
• Use your own voice. Contractions, the occasional fragment, an aside. The texture a model sands off is exactly the texture detectors read as human.
• Cut the filler openers. “It’s important to note that” and “in today’s world” are low-perplexity boilerplate. Delete them.

None of this is about gaming a number. It’s about writing that a human reader and a statistical model both recognize as written by a person, because it was.

Frequently asked questions

What perplexity score does GPTZero consider “human”?

GPTZero doesn’t publish a single cutoff, and you shouldn’t treat one as gospel. Perplexity is relative to the measuring model and to your topic; technical or formulaic writing runs low naturally, whether a person or a model wrote it. GPTZero blends perplexity with burstiness and other signals before producing a verdict, so no single value decides the outcome.

Why does GPTZero highlight some of my sentences but not others?

The highlighter runs a sentence-level pass on top of the document score. Short, predictable sentences that resemble their neighbors get flagged; sentences with surprising word choices or distinctive rhythm usually don’t. That’s why hand-written text can sit beside AI-drafted text and only part gets marked. It points to patterns, not to a confession.

Can GPTZero be wrong about my writing?

Yes. Detectors output probabilities, and false positives happen. The 2023 Stanford study found bias against non-native English writers, and OpenAI retired its own classifier in July 2023 for low accuracy. Treat any single score as one signal, not as evidence.

Does making sentences longer trick GPTZero?

Padding usually backfires. Burstiness measures genuine variation, the natural mix of short and long lines real writers produce. Uniformly long sentences don’t move it much and read worse to a human. Write with real rhythm and concrete detail instead.

Is GPTZero’s score the same as Turnitin’s?

No. Different detectors use different models and thresholds, so the same passage can score “likely AI” on one and “likely human” on another. GPTZero emphasizes perplexity, burstiness, and sentence highlighting; other tools weigh signals their own way.

The takeaway

GPTZero isn’t reading your mind. It’s measuring how predictable your words are and how much your sentences vary, then turning that into a probability. Knowing that, you can read a score critically, spot when it’s likely a false positive, and improve writing that genuinely reads flat, without resorting to tricks.

Want to see which sentences in your own draft look machine-written, and fix them while keeping your meaning intact? Run your draft through PaperBleach.