How Does GPTZero Work? What It Looks For, Step by Step

GPTZero works by running your text through a classifier trained on a large corpus of human and machine writing, scoring how predictable each sentence is and how much the rhythm varies across the document, then returning a probability and one of three labels: human, mixed or AI. It is a statistical guess about style, not a record of what you did. That single fact explains most of what people find confusing about it.

Below is what GPTZero actually measures, what the mixed label means, how its writing report differs from detection, why paraphrasing sometimes slips through, and how much weight the output deserves.

How does GPTZero detect AI writing?

GPTZero reads two statistical properties of your prose. The first is how surprising your word choices are, sentence by sentence. Language models pick high-probability continuations, so machine text tends to be smooth and unsurprising. The second is how much your sentence length and cadence vary across a passage. People write unevenly, with a long winding sentence followed by a short one. Models hold a steadier beat.

The model combines those signals across the whole submission and produces a probability, then highlights the individual sentences that pushed the score one way or the other. GPTZero is explicit that accuracy is highest at document level, lower at paragraph level, and lowest sentence by sentence. That ordering is worth remembering, because the sentence highlights are the part people fixate on and the least reliable part of the report.

What does GPTZero look for?

It is not looking for a watermark, a hidden marker, or anything the model left behind. There is nothing in ChatGPT or Claude output that identifies it as such. GPTZero is inferring authorship from writing style alone, which is why the signals it reads are all things a careful human writer might also produce.

Read that table again from the other direction and you can see the failure mode. Formal academic prose, technical documentation, legal writing and English written by someone who learned it as a second language all tend toward the left column, and none of them involved a model.

How does GPTZero know it's AI?

It does not know. It estimates. GPTZero says its model was trained on a large, diverse corpus of human-written and AI-generated text, so what it produces is a similarity judgment: how closely does this text resemble the machine-written examples it learned from, compared with the human ones. The company publishes that at high confidence, 99.1 percent of human articles are classified as human and 98.4 percent of AI articles are classified as AI, and states plainly that edge cases exist in both directions.

The practical consequence is that a GPTZero result is evidence to review rather than a finding. There is no threshold that means guilty, and the tool does not produce one. Anyone treating a percentage as a verdict is adding certainty the software never claimed.

What does GPTZero mixed mean?

Mixed means the classifier read some passages as human and others as machine within the same document. GPTZero labels submissions as human only, mixed, or AI only, and mixed is by far the most common result on real working drafts, because most real drafts are a blend of typing, editing, pasting and revising.

Mixed is also the label people misread most. It does not mean the tool caught a specific percentage of cheating. It means the statistical texture of your document is not consistent, which happens when you write two sections on different days, when you quote source material at length, when you paste in your own notes, or when you edit one half of a piece harder than the other. If you rewrote only the paragraphs that were flagged, mixed is exactly what you will get, and a patchy document reads worse to a human reviewer than a consistent one does.

How does GPTZero writing replay work?

This is a separate product from the detector, and the distinction matters more than anything else on this page. GPTZero offers writing reports through a Google Docs integration that replays how a document was composed over time, so a reader can see whether the text accumulated through drafting or arrived in a few large pastes.

That is provenance, not statistics. It does not analyze your style at all, it looks at the edit history of the file. Two things follow. First, a writing report is much stronger evidence than a detection percentage, in both directions: it can clear you when a detector flags you, and it can contradict you when a detector does not. Second, no rewriting tool touches it, because the record lives in the document, not the prose. Grammarly Authorship and Word revision history work on the same principle. If there is any chance you will be asked how a piece was written, draft it inside the document, keep your notes, and let the history accumulate.

How does GPTZero detect AI paraphrasing?

Paraphrasing tools mostly substitute words while leaving sentence construction alone, and construction is half of what GPTZero measures. Swapping "important" for "salient" changes the vocabulary but not the rhythm, the paragraph shape or the predictability of the underlying sentence pattern, so a light paraphrase often lands in the same band as the original. Sometimes it scores worse, because rare synonyms in an otherwise smooth sentence read as odd rather than as human.

What does move the number is a rewrite that rebuilds sentences: different lengths, different clause order, different entry points into each idea. That is the difference between paraphrasing and genuinely rewriting, and it is also why the humanizer market splits so sharply on quality. If you are comparing tools for this specifically, we put six of them side by side on our GPTZero humanizer comparison, with prices and where each one loses.

Does GPTZero give false positives?

Yes, and GPTZero acknowledges the edge cases itself. Independent testing on real-world writing puts its accuracy near 85 to 90 percent with false positives in the 8 to 15 percent range, well off the vendor benchmark figures, and one widely cited study found 61.3 percent of essays by non-native English writers flagged as AI. In a class of 200, a 10 percent false-positive rate is roughly 20 honest papers with a mark against them.

We went through the numbers in detail in our breakdown of how accurate GPTZero really is, including where the vendor benchmark and the classroom results diverge and why. The short version: it is good on untouched model output and unreliable on careful human prose, which is the opposite of what you would want.

How does GPTZero compare to Turnitin?

They measure the same underlying properties and land in a similar accuracy range, so the meaningful differences are practical rather than technical. GPTZero is public, cheap to run and something you can check yourself on any text in seconds. Turnitin runs inside your institution against a submission corpus nobody outside the company can query, gives you no way to pre-check, and suppresses any AI score between 1 and 19 percent because it treats low readings as too unreliable to report.

For a wider view across the five detectors that matter, including Originality.ai and Copyleaks, see our comparison of which AI detector is most accurate, and our breakdown of how Turnitin detects AI.

Who actually runs GPTZero on your writing

Understanding the tool means knowing who is holding it, because the same percentage means different things in different hands. Teachers and instructors are the largest group, usually running a paste of an assignment out of suspicion rather than as policy. Journal editors and integrity teams use detection as one input inside a larger screening pipeline. Hiring managers increasingly paste cover letters in, which is a coin flip applied to writing that is formal by convention.

The fourth group is editorial. Content teams and agencies now screen freelancer drafts before publishing, partly for quality and partly because a client asked them to, and GPTZero sells an API for exactly that volume of checking. If you run that kind of desk, the check is the cheap part of the job. The expensive part is going back to improve the pages you already published, where the traffic actually lives.

What to do if GPTZero flags your writing

If you wrote it yourself, do not start editing in a panic, because edits made after the accusation are worth nothing as evidence and can look like tampering. Pull your version history first, whether that is Google Docs revisions, Word revision data or dated notes, since a record of real drafting is the strongest answer to a false positive on genuine work. Then run the same text through two other detectors. Disagreement between detectors is itself useful, and it undercuts the idea that any one score is definitive.

If you drafted with AI and want the writing to read as your own, the fix is depth of rewrite rather than word swapping: rebuild the sentences, add specifics only you know, and vary the pacing. A tool built for this can do the structural part, and you can confirm the result yourself because GPTZero is free to run. Paste the output back in, check the label rather than only the percentage, and make sure your numbers and citations survived the pass. Our GPTZero humanizer shows a live gauge and a own AI-pattern score for that reason, and GPTZero is free to run, so hold any vendor to the same test.

The short version

GPTZero infers authorship from writing style, returns a probability with a human, mixed or AI label, and is most reliable at document level on untouched model output. Its writing report is a different and much stronger kind of evidence, because it reads the document's history instead of its prose. Treat the percentage as a signal, keep a record of how you write, and check any tool's claims yourself instead of trusting a message that says the job is done.