What it is
A free web detector with an unusually generous free tier — 15,000 characters per check and a 1,250,000-word monthly allowance without a card. It returns a percentage and a verdict ranging from “Your text is Human written” to “Your text is AI/GPT Generated”, and it has expanded well beyond detection into a summariser, paraphraser, grammar checker, translator and its own chat product.
Its reach is the reason it matters. When the Virginia Tech researchers below chose which detectors to test, they picked ZeroGPT and GPTZero because those were, at the time, the second and first results on Google for “chat gpt text detector”. This is the tool a curious teacher finds in ten seconds, which makes its error profile a public concern rather than a product detail.
The one detailed measurement
Almost every accuracy number circulating about ZeroGPT traces back to a vendor or a content farm. There is one substantial independent test: Weichert and Dimobi at Virginia Tech, who ran 212 human-written undergraduate essays from the Michigan Corpus of Upper-Level Student Papers — all written between 2004 and 2009, so no LLM could have touched them — against 208 ChatGPT-generated essays on the same titles.
It is a preprint that began as a graduate course project, and it has not been peer-reviewed. We are citing it because it is the most detailed public data that exists on this tool, not because it settles anything. Its findings:
- 78.62% overall accuracy — against the “up to 98%” the ZeroGPT site claimed at the time, a gap the authors said led them “to seriously question the efficacy of the detector”.
- 24.64% false positive rate. Close to one in four essays written by humans, years before ChatGPT existed, were flagged as AI.
- 19.14% false negative rate on unaltered ChatGPT output.
The finding that should change how you read a score
The headline number is not the useful part. This is: ZeroGPT’s false-positive rate was not one rate. Broken down by discipline, across the same corpus of human essays, it was 3.70% for English, 23.53% for Political Science, 31.91% for Biology and 39.39% for Philosophy.
A tool that is wrong about 4% of English students and 39% of Philosophy students is not really measuring whether text is AI-generated. The authors offer a mechanism, and it holds up: they found a correlation of r = −0.60 between a human essay’s perplexity — how statistically surprising its word choices are — and the AI percentage ZeroGPT assigned it. Write plainly, in a discipline with settled vocabulary and a formal register, and the tool reads that as machine-like.
This is the same mechanism behind the bias against non-native English writers documented elsewhere in this section. Someone writing in a second language uses a smaller, more common set of words and simpler constructions. That is low perplexity. A perplexity-driven detector cannot tell the difference between a language model and a careful writer with a smaller vocabulary, because along the axis it measures, there is no difference.
Paraphrasing walks straight through it
In the same study, AI text run through a single ChatGPT paraphrasing prompt got past ZeroGPT 89.64% of the time — a 70.5 percentage point increase in its false-negative rate. Two other paraphrase prompts scored 88.75% and 75.00%.
That is the worst result of the three detectors examined, and it is predictable from the mechanism. If the tool is largely reading perplexity, and paraphrasing raises perplexity, then paraphrasing is not a clever exploit — it is just the thing the tool cannot see. The uncomfortable corollary is that a perplexity detector penalises the honest plain writer and waves through the text that has been deliberately worked over.
What the peer-reviewed record does and does not say
ZeroGPT was one of the fourteen tools in Weber-Wulff et al. (2023), whose conclusion covers the whole field rather than any one tool: the detectors were “neither accurate nor reliable”, all scored below 80% accuracy, and only five exceeded 70%.
One figure to be careful with. ZeroGPT was also among the seven detectors in Liang et al. (2023), the study that found 61.22% of human-written TOEFL essays misclassified as AI. That number is a seven-detector average, and the paper publishes no per-tool breakdown. It is regularly quoted as though it were one tool’s false-positive rate. It is not ZeroGPT’s rate, it is not GPTZero’s, and anyone citing it as such — including anyone citing it against us — is misreading the paper.
One more number you will meet: Scribbr scores ZeroGPT at 64% in its widely-cited comparison of twelve detectors. That is a competitor’s test of thirty texts in which the top three places all went to the tester’s own brands, which is worth knowing before you repeat the figure — we set out why on our Scribbr page.
Who you are actually dealing with
We would normally skip corporate paperwork. Here it is worth a paragraph, because the site makes a specific promise that the paperwork does not support.
The FAQ states: “Your data is encrypted at rest and our infrastructure is hosted in Germany and operated in accordance with the GDPR, ensuring strong data protection and confidentiality.” The privacy policy names no legal entity, no country, no data location and no governing law. The terms of use name OLIVE WORKS LLC, at a Casper, Wyoming address, and submit disputes to the courts of “Netherlands County, California” — a county that does not exist, in a state that does.
The most likely explanation is dull: an unedited contract template, which is extremely common in this category and not evidence of bad faith. But the practical position is that a European user choosing this tool for the GDPR assurance has that assurance in a marketing FAQ and nowhere in the documents that would bind anybody. Read the retention promise the same way: the FAQ’s “not saved, shared, published online, or used to train our AI detection model” is a good, clear commitment, and it is not repeated in the privacy policy.
What it costs
Free: $0, no card, 15,000 characters per detection, 1,250,000 words a month. PRO: $9.99/month or $119.88/year, 100,000 characters per check. PLUS: $16.99/month or $203.88/year, adding 35,000 words a month of plagiarism checking. MAX: $20.99/month or $251.88/year, 150,000 characters per check. There are EDU and EXPERT tiers above these, and a pay-as-you-go API from $0.034 per 1,000 words.
Note what the money buys: capacity, not a better detector. Every tier runs the same detection you get free, so paying does not reduce the false-positive rate above.
Where we stand
We sell a detector and a humanizer, so ZeroGPT is a competitor and you should weigh this page accordingly. We have not measured our own pass rate against it and will not publish one before our benchmark produces it under a method we have published in advance.
What we would say to a student: a ZeroGPT score is a weak signal at best, and if you write plainly or in a technical discipline it is weaker still. If one has been used against you, the discipline-by-discipline false-positive spread is a legitimate and specific thing to raise — see what to do next and how to prove you wrote it.
Related
- GPTZero, reviewed — tested in the same study
- Scribbr, reviewed
- Copyleaks, reviewed
- Why human writing gets flagged