AntiGPT

How GPTZero works, what its scores mean, and what to do if it flags your writing

By Vitalik Hakim · Published 20 September 2026 · Updated 18 October 2026 · Facts checked 18 October 2026

GPTZero is an AI detector that classifies a text as written by a human, by AI or by a mix of the two, and its percentage is the model's confidence in that verdict, not the share of your text that AI wrote. GPTZero itself says its results "should not be used to punish students", so if it flags your work, your strongest answer is evidence of how you wrote it, such as a version history or GPTZero's own Writing Report.

We don't run GPTZero, so we publish no pass rate for it. The detector we do measure is here, with dates and raw data.

What GPTZero is

GPTZero was built by Edward Tian, then a computer science senior at Princeton, who released the first version on 2 January 2023, a month after ChatGPT launched. He runs it with co-founder Alex Cui, and it now claims more than 20 million users. It was made for education, for both teachers and students, and also sells to publishers, recruiters and businesses. In June 2026 Superhuman, the company formerly called Grammarly, announced that it is acquiring GPTZero.

Besides the detector, GPTZero offers an Advanced Scan, a plagiarism checker, a Writing Report that replays how a document was written, a Chrome extension for Google Docs, integrations with Canvas and Google Classroom, and an API. It is not the same tool as ZeroGPT (zerogpt.com), which is a different detector run by a different company despite the similar name.

How GPTZero works now

GPTZero became famous for two measures: perplexity (how predictable each word is to a language model) and burstiness (how much that predictability varies from sentence to sentence). How AI detectors work explains both. Many guides still describe GPTZero this way, but its support pages say that "as of autumn 2023, GPTZero no longer uses perplexity and burstiness for its AI detection because we migrated to a deep-learning based architecture." Today it describes a model with seven components, trained on writing from the web, from education and from many AI models.

The current model, GPTZero 4o, became the default for everyone on 20 September 2026. GPTZero says it rebuilt the architecture to be "robust to paraphrasing", added an "AI Patterns" view to Advanced Scan, and leaves headings out of the analysis. It also says the model is "designed to be indifferent to punctuation changes", including em dashes, so the popular belief that one punctuation mark gives AI away does not apply to it.

Advanced Scan shows which sentences and phrases are pushing the overall score towards AI or towards human. That is useful for a writer, but it shows where the model reacted, not proof of where AI was used.

How to read a GPTZero result

Since January 2024 GPTZero sorts text into three classes: written entirely by a human, written entirely by AI, or written by a mix of human and AI. The result reads like "We are highly confident this text was AI generated" or "We are highly confident this text is entirely human", with a percentage beside it.

That percentage is a probability, not a measurement of how much AI is in the text. GPTZero is explicit: "the percent you're seeing is NOT saying that '6% of your document contains AI.'" The confidence wording matters as much as the number. GPTZero labels each verdict uncertain, moderately confident or highly confident, and its own published error rates differ sharply between them:

  • Highly confident: an error rate below 1% on its technology page, or below 2% in a 2024 explainer.
  • Moderately confident: an error rate of around 10%.
  • Low confidence: an error rate of 14% or higher.

The labels for edited and paraphrased text

Two newer labels catch people out. Since June 2025, "AI Polished" marks text that GPTZero thinks a human wrote and AI later improved; it appears under the mixed class, with the edited sentences highlighted in light green. If you run your own writing through an AI editing tool, expect that label. Since November 2024 GPTZero also flags "possible AI paraphrasing detected", from a feature it says detects text run through AI paraphrasers and humanizers. That feature is not part of the free plan.

GPTZero sets no minimum length, but says "accuracy improves with longer inputs", and independent researchers found detectors in general losing accuracy below 50 words. It fully supports English, German, Portuguese, French and Spanish, and performs best on long English prose.

GPTZero and humanizer tools

AntiGPT is a humanizer, so we will be direct. GPTZero says it detects text from paraphrasers and humanizers, and its 2026 model was rebuilt with paraphrasing in mind. We do not run GPTZero, we publish no figures for it, and we cannot tell you how it will score any text, including text rewritten with AntiGPT. If your teacher or your course does not allow AI, rewriting AI text with a tool does not change the fact that you used AI.

Where AI help is allowed, the useful question is not what GPTZero will say but whether the text is yours: your argument, your examples, your structure. The habits that make writing read as human to a person are the ones described in why real writing gets flagged and in our guide to essays and coursework.

Why GPTZero flags human writing

GPTZero's own figures have improved a great deal. It says it keeps its false-positive rate "at no more than 1%", that it reduced false positives on essays by non-native English speakers to 1.1%, and that GPTZero 4o has a false-positive rate of "1 in 10,402" on student writing. It also says, fairly, that "there always exist edge cases" where human writing is classified as AI.

The edge cases are well documented. In 2023 GPTZero labelled the US Constitution as "likely to be written entirely by AI" when Ars Technica tested it; Tian's explanation was that the Constitution "is a text fed repeatedly into the training data of many large language models". The same year a University of California, Davis student was cleared of AI use after showing his Google Docs edit history, and a 2023 Stanford study found seven detectors, GPTZero among them, labelling on average 61% of TOEFL essays by non-native English speakers as AI. In October 2024 Bloomberg Businessweek ran 500 essays written at Texas A&M in summer 2022, before ChatGPT existed, through GPTZero and another detector: about 1 to 2% were wrongly flagged as likely AI, some with near-total confidence.

More recent independent tests are kinder. A 2025 University of Chicago working paper found false-positive rates below 1%, and Epoch AI found GPTZero flagging none of 495 pre-2022 human passages in July 2026. Even so, 1 in 100 is a lot of students in a large course, and famous or formulaic text remains the weak spot. Our own benchmark shows the same failure in a different detector, for the same reason Tian gave.

What our own benchmark shows

Six passages written by people, long before AI writing tools existed, scored by the Hello-SimpleAI RoBERTa detector (open model) in our weekly benchmark. Run of 16 November 2026. "Likely AI" means a score of 60% or more.

PassageWordsScored as AIVerdict
Abraham Lincoln, the Gettysburg Address (1863)10299%Likely AI
Mark Twain, "Advice to Youth" (1882)7087%Likely AI
Frederick Douglass, Narrative of the Life of Frederick Douglass (1845)7430%Uncertain
Jane Austen, Pride and Prejudice (1813)702%Likely human
Henry David Thoreau, Walden (1854)990%Likely human
W. E. B. Du Bois, The Souls of Black Folk (1903)720%Likely human
2 of 6 human passages flagged as likely AI. The same passages have been flagged in every run since 24 September 2026: the error is consistent, not random. We do not run GPTZero, so this is not a measurement of GPTZero. It shows a failure that all detectors share. Method and raw data.

If GPTZero flags your work

GPTZero's advice to teachers is clear: "these results should not be used to punish students", and it recommends treating a result "as a conversation starter, and not as the final verdict." It tells educators to ask for "drafts, revision histories, or brainstorming notes" and to look for "a long-term pattern of AI use, as opposed to a single instance." Use that guidance:

  • Ask for the full result: the class (human, AI or mixed), the confidence label and the highlighted sentences, not only the percentage. An "uncertain" or "moderately confident" result has, by GPTZero's own figures, an error rate of 10% or more.
  • Show how you wrote it. Google Docs keeps version history automatically (Word does when the file is saved to OneDrive); GPTZero's Writing Report, available through its free Chrome extension, replays the writing of a Google Doc and has a share link for your teacher.
  • Point out anything that explains the score: a template you had to follow, quoted material, an AI grammar or editing tool, or a translation from your first language.
  • Offer to show your understanding: talk through your argument and sources, or write a short piece in front of your teacher, which GPTZero itself suggests to educators.
  • GPTZero has no appeal procedure of its own; the decision belongs to your school. If you are accused of using AI sets out the steps, and our guide for non-native English writers covers the extra risk if English is your second language.

Where AntiGPT fits, honestly

AntiGPT's detector is Hello-SimpleAI's open RoBERTa classifier, not GPTZero, so the two scores will not match and ours does not predict GPTZero's. We show it before and after a rewrite as feedback on how predictable your text reads, labelled as an estimate.

If AI help is allowed in your work, AntiGPT rewrites a draft so it reads in your own voice, with two versions to choose from and single sentences you can redo. Check the result the way a careful reader would: facts, quotations and citations intact, and an argument you can explain without notes. The free sentence variety checker shows one of the signals detectors read, on your own text and without an account.

Questions

Is GPTZero the same as ZeroGPT?

No. GPTZero (gptzero.me) was made by Edward Tian and Alex Cui; ZeroGPT (zerogpt.com) is a separate detector run by a different company. The famous 2023 screenshot of the US Constitution being flagged came from GPTZero.

What does the percentage on GPTZero mean?

It is GPTZero's confidence in its verdict, not the share of your text written by AI. GPTZero says a figure such as 6% does not mean 6% of the document contains AI. Read it together with the class (human, AI or mixed) and the confidence label.

Is GPTZero free?

There is a free plan with an account: 10,000 words a month and up to 10,000 characters per scan, without paraphrase detection or the plagiarism checker. Paid plans add larger allowances; GPTZero's pricing page showed different prices to different visitors when we checked, so check it directly.

Can GPTZero detect humanized or paraphrased AI text?

GPTZero says it can: it has flagged "possible AI paraphrasing" since November 2024, and its September 2026 model was rebuilt to resist paraphrasing. We do not run GPTZero and cannot say how it scores any particular text.

How accurate is GPTZero?

GPTZero reports a false-positive rate of at most 1%, and recent independent tests found similar or lower rates. Earlier tests found higher error rates, especially on essays by non-native English speakers and on famous texts such as the US Constitution.

Can my teacher fail me because of a GPTZero score?

That is your school's decision, but GPTZero itself says its results should not be used to punish students or as a final verdict. Ask for the full result and bring your drafts and version history to the conversation.

Sources

Facts about GPTZero are taken from its own published pages, last checked 18 October 2026. Vendors change their products without notice; tell us at support@antigpt.app if something here is out of date.

Check your own text

Paste a paragraph you drafted with AI help. You will see how a detector scores it, then you can rewrite it in your own voice. Free: 300 words a week, sign in with Google; your text is kept while you do.

0 words

Detector scores are estimates from third-party classifiers and change as those tools update. No text can be guaranteed to pass any detector. Use this to improve your writing, and follow your institution's or employer's rules on AI use.

All AI detector guides · how our own detector results are measured