Is ZeroGPT Accurate? What Its Own Figures and Three Published Studies Show

Is ZeroGPT accurate? It depends on the text you give it and which mistake worries you more. ZeroGPT's own page reports high figures without publishing the test behind them. Three published evaluations that name the tool found a mixed result: in a 2023 study it raised no false alarm on 18 human-written documents but missed more than a third of the AI-written ones. This page sets the vendor's claims beside those studies, explains what each number counts, and shows how to read a ZeroGPT result fairly.

What ZeroGPT says about its own accuracy

ZeroGPT's DeepAnalyse technology page, read on November 2, 2026, gives three headline figures: 98.4% detection accuracy, labeled a "reported performance metric"; a false positive rate under 1%, defined as "human text incorrectly identified as AI"; and coverage of more than 35 model families. The page defines detection accuracy as "the reported percentage of evaluated classifications that were correct in the referenced evaluation," and a note at the bottom calls all three numbers "ZeroGPT product metrics."

What the page does not give is the evaluation itself: no count of texts, no split between human and machine-written samples, no languages or dates. It says the exact training composition and model weights "are not disclosed publicly." Without the test set, nobody outside the company can tell whether 98.4% was measured on writing like yours, or on unedited chatbot essays that are easy to catch.

The same page says plainly that a high AI score does not prove AI was used: "No. AI detection is probabilistic."

Is ZeroGPT accurate in independent tests?

A vendor's accuracy figure describes the vendor's test. Your essay was not in it.

The most detailed academic test that names ZeroGPT is Weber-Wulff and colleagues' study in the International Journal for Educational Integrity (2023). Nine researchers prepared 54 English documents of known origin: 9 written by humans in English, 9 written by humans in other languages and machine-translated into English, and 36 generated by ChatGPT, of which 9 were then edited by hand and 9 run through a paraphrasing tool. All 54 went through 14 detectors; the free tools, ZeroGPT among them, were tested between March 7 and March 28, 2023.

Under the study's strictest scoring, ZeroGPT classified 32 of the 54 documents correctly, an accuracy of 59%, tied for sixth of the 14 tools. It produced no false accusations: none of its results on the 18 human-written documents would have led to one. Its errors ran the other way. It left 13 of the 36 AI-generated documents undetected (36.1%), and it classified only 2 of the 9 hand-edited AI texts and 1 of the 9 machine-paraphrased ones fully correctly. The authors concluded that the tools as a group were "neither accurate nor reliable."

Liang and colleagues' study of detector bias, published in Patterns in July 2023, tested seven widely used detectors, ZeroGPT and GPTZero among them, accessed on March 15, 2023. They used 91 TOEFL essays written by non-native English speakers and 88 essays by US eighth graders. The detectors were near-perfect on the eighth-grade essays but labeled more than half of the TOEFL essays AI-generated, an average false positive rate of 61.3% across the seven. The text reports averages, not a separate ZeroGPT rate.

The largest benchmark to include ZeroGPT is RAID, presented at the ACL 2024 conference by Dugan and colleagues. It holds over 6 million machine-generated texts from 11 models across 8 domains, alongside human-written documents. At common default thresholds, ZeroGPT flagged 1.42% of the human-written texts as machine-made at a threshold of 0.5 and 1.71% at 0.25: low in absolute terms, and the highest of the four commercial detectors (GPTZero, Originality and Winston were the others). When each detector was tuned to misclassify 5% of human texts, ZeroGPT was marked as unable to reach that target, and its accuracy ranged from 95.4% on text from open chat models with greedy decoding to 0.3% on open non-chat models with random sampling and a repetition penalty.

ZeroGPT false positives: who is most at risk

Read together, the studies say ZeroGPT rarely flagged ordinary human English: no false accusations in Weber-Wulff's 18 human documents, and a rate below 2% on RAID's human texts. Two cautions apply. A low rate is not zero at scale: if the 1.42% RAID rate held for 1,000 human-written texts, about 14 would be flagged, and RAID's texts are not student essays. And the risk is not spread evenly. Liang's study found that detectors as a group penalized the plainer, more predictable vocabulary common in non-native English, and Weber-Wulff's team found that machine-translated human writing lowered the tools' overall accuracy on human text. Marking suspected AI writing fairly sets out how to raise a concern without turning a score into an accusation, and what AI text detectors can and cannot tell you explains why these errors happen.

How to read a ZeroGPT result fairly

1. Check what was submitted

   Check the complete final version, not a fragment with headings and references mixed in.

2. Read the highlighted sentences

   ZeroGPT highlights the sentences it scored as AI-like. Read them, and ask whether they are generic or simply clear.

3. Weigh the kind of writing

   Published studies found the most trouble with non-native and machine-translated English. Lower your confidence in a flag on either.

4. Look for the writing process

   Drafts, notes, outlines and version history show how a text was built. A score cannot. Using version history as an authorship record explains what to look for.

5. Decide on evidence, not on the number

   A percentage is one input. Talk with the writer and keep any decision tied to what you can show.

Percentages from different detectors are not on one shared scale. How AI detection scores work, and what a percentage measures covers that in detail.

A worked example: one flagged paragraph

Here is an example paragraph written for this page, from a short student reflection. Suppose a detector highlights the first and third sentences.

• Sentence 1 (highlighted, example): "Climate change is one of the most important issues facing the world today."
• Sentence 2 (example): "Last July my grandmother's well in Fresno ran dry for the first time in forty years."
• Sentence 3 (highlighted, example): "It is important for governments and individuals to work together to find solutions."
• Sentence 4 (example): "We now drive to my uncle's house with plastic jugs every Sunday."

The highlighted sentences are the predictable ones: stock lines that countless human writers also produce. The others carry a date, a place and a family detail. A fair reader concludes nothing about authorship from this pattern. The useful response is an editing note: keep the specific sentences, cut the generic ones. If authorship really is in question, ask for drafts rather than rerunning the detector.

Checking a second detector

If you are checking your own draft before you submit it, a second tool shows whether a flag is specific to ZeroGPT. You can check the same draft with a second detector such as Walter Writes AI, an AI humanizer with a built-in AI detector that supports over 80 languages; every plan includes voice matching and the detector. The usual limit applies: no detector, Walter's included, can guarantee 100% accuracy, and detectors often disagree about the same paragraph.

Teachers choosing a tool for a class can use the comparison of AI detectors for teachers, and the sibling page on whether GPTZero is accurate covers the same ground for that tool.

Pitfalls

• Quoting 98.4% as if it applied to your text. It is a vendor metric from an unpublished evaluation.
• Reading a low score as proof of human writing. Weber-Wulff's study found ZeroGPT missed more than a third of AI texts, most of them edited or paraphrased.
• Reading a high score as proof of AI use. ZeroGPT's own page says a high score does not prove AI was used.

Expert tips

• Ask the vendor question every time. For any accuracy claim, ask on what texts, how many, and when. If the page cannot answer, give the number little weight.
• Prefer error rates to accuracy. The false positive rate on human writing matters more to a student than an overall accuracy figure.

For a fourth checker read the same way, see the QuillBot AI detector and its published results.

So, is ZeroGPT accurate? On plain native English in published tests, it rarely accused human writers. On edited or paraphrased AI text it missed a great deal, and its vendor figure comes from an evaluation nobody outside the company can inspect. Use it to find sentences worth rereading, and look for drafts before you conclude anything. More guides to checking and editing AI writing are on the AI writing hub.