How Does GPTZero Detect AI Writing? The Full Breakdown

Table of Contents

• How does GPTZero detect AI writing today?

• What does a GPTZero result actually tell you?

• What does GPTZero ignore?

• How has GPTZero's model changed in 2026?

• How did GPTZero work before? (Perplexity and burstiness)

• Where does GPTZero still get it wrong?

• FAQ

Ask most people how GPTZero works and you'll get an answer that was true for a few months in 2023. The tool has changed a great deal since then, and it hasn't stopped. Its release notes list thirteen model versions in 2026 alone, and the newest, model 5.0b, shipped on November 4. The prior release, model 4.11b, shipped on October 7.

What follows is how GPTZero decides a piece of writing is AI today, taken from the company's own pages: its technology page, its release notes, its model announcements and a technical report its team posted to arXiv. Where a number comes from somebody else, the post says who. Everything here was checked on November 4, 2026. If you'd rather have the general theory of detectors than the workings of this one, that's in how AI detectors work.

How does GPTZero detect AI writing today?

GPTZero describes its method plainly. "We employ an end-to-end deep learning approach, trained on text datasets from the web, education, and AI-generated from a range of LLMs," its technology page says. On top of that sits what it calls a sentence classifier: "A sentence-by-sentence classification model determines the probability and confidence that a text was created by AI."

Put simply, nobody sat down and wrote a rule that says AI text looks like this. The model was shown a very large amount of writing whose origin was already known, and it learned to tell the two kinds apart. The technical report breaks the training data down by domain, and the numbers are large: 16 million news documents, 8 million web articles, 2.8 million encyclopedia entries and 1.25 million academic texts, among others.

The report also describes the structure, which it calls "a hierarchical, multi-task architecture." At the top level the model sorts a document into human, AI or mixed. Under AI it makes a finer guess: pure AI, polished (human writing that an AI cleaned up) or AI paraphrased (AI writing that an AI then reworded). Alongside that, it scores every sentence on its own for how likely it is to be AI. Both answers come out of what the report calls "a single forward pass," so the verdict on the document and the highlights on the sentences are produced together, not one worked out after the other.

Long documents are cut into windows the model can handle, scored piece by piece and stitched back together. A last remapping step adjusts the overall prediction, which the report says "improves calibration and reduces false-positive predictions." That stitching is one of the things a recent model touched. GPTZero's release notes list "Model 4.11b (aka 4.1o)," released October 7, 2026, with "Improved document chunking and aggregation" among its changes, next to "Improved recall on sub-classes AI paraphrased, mixed and polished." Model 5.0b, released November 4, 2026, followed with further gains on polished and AI-paraphrased text, lower false-positive rates, and a larger detection architecture.

What does a GPTZero result actually tell you?

A GPTZero result is a label, a set of probabilities and a confidence level, all at once. In the API, according to the technology page, the label sits in a field called document_classification and reads HUMAN_ONLY, MIXED or AI_ONLY. The probabilities for each class are in class_probabilities, and confidence_category reads high, medium or low. GPTZero says that when confidence is high, "error rates are below 1%."

The mixed label matters more than it first appears to. GPTZero added it on December 11, 2023, and the release notes explain the reason. A result of "50% AI" could mean the model can't decide whether the whole document is AI, or that it's quite sure half of it is. For the person who wrote it, those are very different situations, and one percentage can't separate them. A month later GPTZero tuned its probabilities so that, in its words, "when we provide an AI score of 90%, there is a 90% chance that the text is entirely written by an AI."

Then there are the highlights, which show the sentences that pushed the result one way or the other. For most students and teachers they're the first thing the eye goes to. Advanced Scan goes a step further. Its AI Patterns feature, introduced with GPTZero 4o, names habits typical of chatbot writing. There are ten in this release, among them "arranging ideas in neat sets of threes" and "framing claims using a 'Not X, but Y' structure."

If you're looking at a high score and wondering what to do about it, that's a separate question, and we've answered it in how to lower your GPTZero score.

What does GPTZero ignore?

Two things, by its own account. The first is headers. The release notes for model 4.8b, released August 1, 2026, say "headers are now excluded from our model's input." The 4o announcement later presented this as a feature called Masked Detection. The detector "does not consider headings as a feature," it says, and a new gray highlight marks the text that was left out. So if a newer result shows gray headings and an older one doesn't, the two weren't produced the same way.

The second is punctuation. The 4o announcement says GPTZero's models "are designed to be indifferent to punctuation changes," and it gives the reason in so many words: "we know many write with em dashes, and we don't want to be biased against that." Adding or removing a semicolon, it says, "is not going to determine" the outcome. Both of these are GPTZero describing its own design. We haven't tested them, and they're best read as the company's claims.

How has GPTZero's model changed in 2026?

GPTZero ships new models often and numbers each one. Versions ending in "b" carry a model version tagged "base," and those ending in "m" are tagged "multilingual." Here is a selection from the release notes, each with the release date GPTZero gives.

The dates need one note. The 4o announcement carries October 18, 2026 at the top, but its body says that "as of 20 September 2026" the model was "the default model offered on GPTZero," which fits the release notes. The September date is the one that matters for anyone comparing results.

And comparing results is where these changes bite. A score from before one of these releases and a score from after it aren't quite measuring the same thing, even on identical text. We went through why in fall 2026 AI detector model changes.

How did GPTZero work before? (Perplexity and burstiness)

GPTZero went public in January 2023, and that first version worked very differently. Edward Tian, GPTZero's CEO, wrote in a March 2023 post that "the original GPTZero model applied a 'statistical approach,'" and that its statistical layer "is composed of a 'perplexity' and 'burstiness' calculation." That's the GPTZero perplexity and burstiness method most explainers still describe. What the two measures actually are is covered in how perplexity and burstiness work, so we won't go over it again here.

The switch happened in stages. On May 23, 2023, the release notes record a new deep learning model that "looks for more subtle signals of AI-generated writing," with a plain line about how it was used: "We combine the results from our statistical approach with this deep learning model." For a few months the two ran side by side. In July 2023 GPTZero even added a message for when they disagreed: "Our ensemble of detectors have predicted different results for this text." New deep learning models followed in August, September and October 2023, and on November 20, 2023 the notes say GPTZero "Consolidated ensemble predictions."

The March 2023 post now opens with an update: "As of autumn 2023, GPTZero no longer uses perplexity and burstiness for its AI detection because we migrated to a deep-learning based architecture." Lower down, the same page still says the old methods "act as one of the seven 'indicators'" of the upgraded model, an older line that sits awkwardly under the note above it. The update note, the technology page and the technical report all agree with each other, and this post follows them.

Where does GPTZero still get it wrong?

Start with what GPTZero itself concedes. Its technology page says "No AI detector is 100% accurate," that GPTZero "performs best of longer texts and English prose," and that it should be used "as a conversation starter, and not as the final verdict." The technical report is franker in places. It says in-distribution performance metrics "are overly optimistic," and that being robust to users trying to bypass the detector "remains a priority."

On false positives, GPTZero's headline claim for 4o is a rate of "1 in 10,402" on student writing. The figure comes from the PERSUADE set of student essays, where GPTZero says it "found only 1 false positive out of 10,402 essays." It's the company's own test of its own model.

Outside figures look different, though they measure something else. HumanizerEval, a third-party benchmark site, ran its October 2026 review on October 22 against GPTZero 2026-10-07-base, which is model 4.11b. Eleven humanizer tools rewrote the same 50 AI-written texts, and on average their outputs passed GPTZero 46.7% of the time, meaning GPTZero gave them a human score above 50%. HumanizerEval is careful to add that this "is not a false-positive rate," because there is no human-written control set. Originality.ai, which sells a competing detector, cites another third party in its October 25 launch post: HumanizerBench's October 2026 results, which put GPTZero 4o at 35.95% detection on 420 humanized AI outputs. That's a competitor quoting someone else's test, and it should be weighed that way.

We make a humanizer, StealthGPT, so we have a stake in this, and it's fair to keep that in mind. What the numbers show is fairly narrow. In the tests GPTZero reports, it rarely flags plain human writing. Text that has been deliberately rewritten is where outside tests and GPTZero's own claims sit furthest apart. One score on one essay is a reason for a conversation. It isn't proof of anything.

FAQ

Does GPTZero still use perplexity and burstiness?

No. GPTZero says it stopped in autumn 2023, when it moved to a deep-learning architecture.

What model does GPTZero use now?

As of November 6, 2026, the newest entry in GPTZero's release notes is Model 5.0b, released November 4, 2026. The prior release was Model 4.11b, which GPTZero also calls "4.1o," released October 7, 2026.

How does GPTZero's sentence classifier work?

It scores each sentence for how likely it is to be AI. Those scores sit alongside a document label (human, mixed or AI) and a confidence level, and the model produces both at the same time.

Does GPTZero check headings?

No. Since model 4.8b (August 1, 2026), GPTZero excludes headers from its input, and newer results show them in gray.

Can GPTZero flag human writing?

It can. GPTZero reports a false-positive rate of 1 in 10,402 on student writing, which is the company's own figure, and any single score should be read as a signal, not proof.