The short answer
It depends on the text, and nobody has one reliable number. ZeroGPT claims 98.4% detection accuracy. One small preprint on physics essays found 98%. A preprint on Arabic articles found 80%, with 38% of human articles flagged as AI. A study of 14 detectors that included ZeroGPT found none reached 80% overall.
What ZeroGPT claims
Its home page lists 98.4% detection accuracy and a false positive rate under 1%. It says it covers more than 35 families of AI models and uses a multi-stage method it calls DeepAnalyse. Its FAQ describes accuracy differently: it says it is pushing toward more than 98% on internal evaluations.
So the figure is internal. We found no test set, sample size or date with it on the pages we read. The home page FAQ has a question on the accuracy rate, but no answer showed in the page we read.
What published tests found
| Study | Texts | ZeroGPT result |
|---|---|---|
| Yeadon et al., 2024 (preprint) | 30 AI and 30 human short physics essays, in English | 98% accuracy. No human essay flagged, one AI essay missed. |
| Almohaimeed et al., 2025 (preprint) | Human and AI articles in Arabic | 80% accuracy. 98% of AI text found, 38% of human text flagged. |
| Weber-Wulff et al., 2023 (peer reviewed) | 54 test cases across 14 tools | Every tool scored under 80% overall. We could not read a separate ZeroGPT figure in the article text. |
These do not cancel out. The first is small, in English, and from a test of unedited essays. The second is in another language. The third is from 2023, and detectors have changed since. The honest reading is that the result moves a lot with the text. The second paper also found that human articles lightly polished by an AI model were often flagged, which matters if you use AI to correct your own writing.
What ZeroGPT says about limits
- Length. Its FAQ recommends at least 150 to 200 words. It says 500 to 1,000 or more improves stability.
- Human writing. It says highly polished, formulaic or low-entropy writing can resemble AI.
- Languages. It says all available languages are supported. It gives no accuracy by language on the pages we read.
- Your text. It says text is not saved, shared, published or used to train its model by default. We did not verify this.
Why a ZeroGPT score is not proof
ZeroGPT highlights sentences and shows a percentage. That looks exact, but it is a prediction. For how detectors decide, see how AI detectors work. For what a number means, see what percentage of AI is acceptable.
Three things hold for any AI detector, this one included:
- It reads the finished text. It does not see how the text was written, so it can only estimate.
- Its figures depend on the test. An accuracy number holds for the texts it was measured on. Your text may be shorter, edited or in another language.
- A small error rate is still real people. Liang and colleagues (2023) found seven detectors, which the paper does not name, flagged 61.3% of TOEFL essays by non-native speakers on average. Detectors have changed since, but plain, predictable writing remains the risk.
If your own writing is flagged
- Keep your evidence. Drafts, version history, notes and sources show how you worked. See how to prove you didn’t use AI.
- Read your course policy. It decides what counts as acceptable AI use, not a detector.
- Talk to your instructor. Offer to explain your work and your sources.
Our AI detector gives a separate estimate with a likelihood for every sentence. It can disagree with ZeroGPT. We publish no accuracy figure for it.
Sources
- ZeroGPT: home page and FAQ, read 1 November 2026.
- Studies: Yeadon et al., Almohaimeed et al. and Weber-Wulff et al.