Is ZeroGPT accurate? Our short answer
On everyday text, mostly yes. In our test of 40 texts, ZeroGPT flagged all 20 raw AI drafts and cleared all 14 human texts written in recent years. Where it failed, it failed hard: it labeled all six famous older texts in the set as AI, from the Declaration of Independence to Pride and Prejudice, with scores between 73% and 100%.
So a ZeroGPT score is a signal worth reading, not a verdict. That holds for every AI detector, ours included, and it matches what published research says about detector accuracy. Below are our method, every number we got, what other researchers found, and what to do with your own result.
How we tested ZeroGPT
We built a set of 40 English texts, 130 to 360 words each, where we know who wrote every one. Each human text was written before ChatGPT came out on November 30, 2022, so none of it can be AI. The AI texts are raw model output that nobody edited.
| Texts | What they are | Written |
|---|---|---|
| 6 human | Famous classics: the Declaration of Independence, the US Constitution, the Gettysburg Address, Federalist No. 10, Walden and Pride and Prejudice | 1776 to 1863 |
| 9 human | Academic writing: passages from open-access research papers and from theses | 2018 to 2021 |
| 5 human | Casual writing: Reddit comments about science, history, travel and home repair | Before 2022 |
| 12 AI | DeepSeek V4.1 Flash: essays, a business email, a cover letter, a news story, a how-to guide and more | 2026 |
| 8 AI | Claude Opus 5.5: school essays, a book report, a lab report, a scholarship essay and more | 2026 |
On October 2, 2026, we pasted each text into the free checker on ZeroGPT's website, one at a time, and wrote down its verdict and its "AI GPT" percentage. We counted a text as flagged when ZeroGPT scored it 50% AI or higher. In this run the cutoff never mattered: every score was either 0% or above 72%.
About 35 minutes later we ran all 40 texts through again. Every score and every verdict came back exactly the same, so the results below aren't a fluke of one scan.
Our ZeroGPT results
| Texts | Flagged as AI | ZeroGPT's scores |
|---|---|---|
| 20 AI drafts | 20 of 20 | 72% to 100% AI |
| 14 recent human texts | 0 of 14 | 0% AI each |
| 6 classic human texts | 6 of 6 | 73% to 100% AI |
That is 34 of 40 right. All six mistakes went the same way: human writing labeled as AI, a false positive. ZeroGPT missed no AI text.
The six classics ZeroGPT called AI
| Text | Year | ZeroGPT score |
|---|---|---|
| Declaration of Independence, opening | 1776 | 96.8% AI |
| US Constitution, Preamble and Article 1 | 1787 | 99.2% AI |
| James Madison, Federalist No. 10, opening | 1787 | 85% AI |
| Jane Austen, Pride and Prejudice, chapter 1 | 1813 | 73.4% AI |
| Henry David Thoreau, Walden | 1854 | 100% AI |
| Abraham Lincoln, Gettysburg Address | 1863 | 96.2% AI |
For all six, ZeroGPT showed the same verdict: "Your Text is AI/GPT Generated". The 14 recent human texts all got "Your Text is Human written" at 0%.
The AI side
The eight Claude texts scored 97% to 100% AI. The DeepSeek texts spread out more: a news article scored 72.4% and an essay on Macbeth 75.7%, both still well past the line.
Not every detector shares this blind spot. QuillBot's free AI detector scored the Declaration, the Constitution, the Gettysburg Address and Federalist No. 10 at 0% AI on the same day. Our own detector scored all 20 human texts at 0% AI and all 20 AI texts at 66% or higher. We make that one, so weigh its result accordingly.
Why ZeroGPT flags famous old texts
AI detectors score how predictable each word is to a language model. A text that a model saw many times during training is extremely predictable to it, and predictable text is what detectors read as AI. The Constitution and the Gettysburg Address are among the most copied texts on the web.
This isn't new. In July 2023, Ars Technica showed ZeroGPT calling part of the US Constitution "AI/GPT Generated" and flagging a passage from the Book of Genesis as 88.2% AI, and explained why AI detectors think the US Constitution was written by AI: the text is so ingrained in the models' training data that it looks machine-made. For the mechanics, read how AI detectors work.
ZeroGPT doesn't publish the details of its model, so we can't prove this is the cause in its case. The pattern fits it, though: the famous, much-quoted texts failed and the recent ones passed. What it means for you: if your essay quotes long passages from famous sources, or follows them closely, those passages may push your score up. Mark quotes as quotes, cite them, and keep your own analysis in your own words.
What other tests of ZeroGPT found
ZeroGPT's homepage advertises 98.4% detection accuracy and a false positive rate under 1%. It doesn't publish the test data behind those figures. Independent studies that included ZeroGPT found a much wider range:
| Study | What was tested | ZeroGPT's result |
|---|---|---|
| Weber-Wulff et al., 2023, tested March 2023 | 54 documents: human, ChatGPT, edited and paraphrased | 59% correct. It flagged no human text but missed 36% of the AI cases. |
| Walters, 2023, tested mid-2023 | 126 undergraduate essays: 84 by ChatGPT, 42 by students | Caught 92% of the AI essays. Of the student essays, 79% were called human, 2% AI and the rest "uncertain". |
| Liu et al., 2024 | 150 medical articles: 50 originals, 50 by ChatGPT, 50 rephrased by AI | Caught 96% of the ChatGPT articles and 88% of the rephrased ones. About 16% of the originals were labeled AI or uncertain. |
| Pratama, 2025 | 72 human research abstracts, half by non-native English writers, plus AI versions | 64% correct. About 17% of the human abstracts were flagged or marked uncertain, and 45% of the AI abstracts were missed. |
Read together, the studies say ZeroGPT's results depend heavily on the texts. In the 2023 test of 14 tools it was cautious: no false positives, many AI texts missed. On medical writing and research abstracts it flagged more human work. In those two studies about one human text in six was flagged or marked uncertain, far from the under-1% false positive rate ZeroGPT advertises, and our six classics show how far off it can be on one kind of writing.
ZeroGPT itself is more careful in its terms of use, which say a detection score "does not constitute definitive proof" and should not be "the sole basis for a decision".
ZeroGPT vs GPTZero
The names are close, but they are separate tools from separate companies. ZeroGPT's terms name Olive Works LLC of Wyoming as the business behind it. GPTZero was built by Edward Tian, then a Princeton senior, who released it on January 2, 2023; its terms of use name GPTZero LLC of New York.
On accuracy, the studies above don't crown either one. Walters found the two performed much alike. In the abstracts study, GPTZero did far better, with 97% correct and no false positives. In the medical study, GPTZero labeled about 22% of the human originals as AI or uncertain, against about 16% for ZeroGPT. Both have called the US Constitution AI-written.
We haven't tested GPTZero ourselves yet. For the published evidence on it, read is GPTZero accurate.
How to read your ZeroGPT score
- A high score on a famous or much-quoted text means little. Our test shows ZeroGPT flags those.
- A high score on your own fresh writing is a reason to look at the highlighted sentences, not to panic. Check what they share: stock transitions, sentences of the same length, broad claims with no examples. Why your writing gets flagged as AI covers the usual causes.
- Get a second opinion. Detectors use different models. In our test, ZeroGPT and QuillBot gave the same four classics opposite verdicts.
- Keep your drafts and version history. They show how you wrote, which no score can.
- If you grade work, never treat one score as proof. ZeroGPT's own terms say the same. Talk to the student and look at the drafts first.
The limits of our test
- 40 texts is a small set. It shows a clear pattern, not a precise accuracy rate.
- The AI texts are unedited. Edited or paraphrased AI text is harder for every detector, and we didn't test it.
- There are no essays by students writing after 2022, and none by non-native English writers, the group published studies find gets flagged most often.
- Both runs were on one day. A repeat scan gives the same score, but detectors update their models, so we plan to run the same texts again in the coming weeks and report any change.
- We make an AI detector ourselves. We used ZeroGPT's own free checker and report every result, including everything it got right.