Is ZeroGPT accurate?
Independent tests scored ZeroGPT far below its own 98.4% claim, and it wrongly flagged more human text than any other commercial detector in the RAID paper.
Short answer
ZeroGPT is not accurate enough to trust on its own. It claims 98.4% accuracy, but independent tests scored it 59% in 2023 and 46.1% in 2024, and the RAID paper found it wrongly flagged 1.71% of human texts, the highest false positive rate among the commercial detectors studied.
Is ZeroGPT accurate? Not by the published evidence. It is one of the most searched free AI detectors, and its homepage promises near-perfect results. The independent studies that included ZeroGPT tell a different story. Below are the results, the dates and the links, with the vendor's own figure marked as a vendor claim. For how we weigh sources, see how we score AI accuracy.
ZeroGPT test results
- Independent study
- Vendor claim
| Test | What was measured | Result | Date | Source |
|---|---|---|---|---|
| ZeroGPT homepageVendor's own figure, not independently checked | Detection rate (vendor's stated accuracy) | 98.4%, with false positives under 1% | Source 1Vendor claim | |
| Weber-Wulff et al. 2023, 54 documents | Detection rate (binary accuracy on human, translated, AI and obfuscated texts) | 59%, 6th of 14 tools | Source 2Independent study | |
| Perkins et al. 2024, 15 AI and 10 human samples | Detection rate on unedited AI text | 46.1%, 5th of 7 detectors | Source 3Independent study | |
| Perkins et al. 2024 | Detection rate after evasion edits | 17.3%, down from 31.3% | Source 3Independent study | |
| RAID paper (Dugan et al.), threshold 0.25 | False positive rate on human text | 1.71% | Source 4Independent study | |
| Liang et al. 2023, 91 TOEFL essays | False positive rate on non-native English essays (average of seven detectors, ZeroGPT included) | 61.22% | Source 5Independent study |
What the ZeroGPT accuracy numbers mean
The gap between the claim and the tests is the main story. ZeroGPT's homepage states 98.4% accuracy, a false positive rate under 1% and coverage of more than 35 model families. The page gives no method or dataset for those figures.
The independent numbers are much lower. Weber-Wulff and colleagues ran 54 documents through 14 tools in 2023, and ZeroGPT got 59% right. The paper's conclusion covered every tool: detectors are neither accurate nor reliable. In 2024, Perkins and colleagues found ZeroGPT caught 46.1% of unedited AI text, fifth of seven detectors.
The false positive picture matters even more for anyone who writes their own work. The RAID benchmark paper set ZeroGPT at a threshold of 0.25 and found it flagged 1.71% of human texts as AI. The researchers could not tune it down to their 5% false positive target in the usual way, and its rate was the highest among the commercial detectors in the paper.
Both academic studies used older versions of ZeroGPT, and both papers warn their results may not match today's tool. The vendor's own number is the most recent figure, and it is the one with no published method behind it.
Where ZeroGPT goes wrong
It flags human writing
At its default-like threshold, ZeroGPT had the highest false positive rate of the commercial detectors in the RAID paper.
Light edits get past it
After evasion edits in Perkins et al., its accuracy fell from 31.3% to 17.3%.
Non-native writers are at risk
ZeroGPT was one of the seven detectors in the Liang et al. study that flagged TOEFL essays as AI 61.22% of the time on average.
No public method
The homepage claim comes with no dataset you can check.
How ZeroGPT compares in the same tests
Only compare detectors inside one study. These rows come from the same tests, run on the same documents.
| Test | Turnitin | ZeroGPT | GPTZero |
|---|---|---|---|
| Weber-Wulff et al. 2023, binary accuracy | 76% | 59% | 54% |
| Perkins et al. 2024, unedited AI text | 61% | 46.1% | 26.3% |
After evasion edits in the same Perkins study, ZeroGPT fell from 31.3% to 17.3% while Turnitin dropped from 50% to 7.9%, so ZeroGPT held up better under rewriting, though neither result is usable.
In both academic tests, ZeroGPT sat behind Turnitin and ahead of an older version of GPTZero. Since then, GPTZero has posted much stronger results, including 97.2% on the RAID leaderboard in October 2025 and a false positive rate of about 0.7% in a 2025 study. Grammarly's detector reached 99.5% on the same leaderboard. For the full picture across tools, see are AI detectors accurate.
Non-native English writers and ZeroGPT
ZeroGPT was one of seven detectors in the Stanford study by Liang and colleagues. On 91 essays written for the TOEFL English test, the detectors flagged human work as AI 61.22% of the time on average (Liang et al., 2023), and 97.8% of the essays were flagged by at least one of them. Essays by US eighth graders drew almost no false flags. The authors tie the bias to simpler vocabulary, which reads as predictable to a detector. If English is your second language, a ZeroGPT score deserves extra caution.
What a ZeroGPT score can and cannot prove
No detector score should be the only evidence in a misconduct case.
It can show
- A ZeroGPT score can tell you that a passage reads like model output to this one tool. That can be a reason to look closer at a draft.
It cannot prove
- A ZeroGPT score cannot tell you who wrote the text, how much AI was used, or whether a careful human writer has simply been misread. Given its independent results, it is a weak signal even by detector standards.
What to do if ZeroGPT flags your work
Note the exact score, the date and the passages ZeroGPT highlighted, and save a screenshot.
Collect your drafts, outline, notes and version history from Google Docs or Word.
Share the independent results above, especially the 59% and 46.1% accuracy figures and the RAID false positive finding.
If English is not your first language, point to the Liang et al. study, which included ZeroGPT.
Offer to talk through your sources and argument, or to write a short sample under supervision.
Ask that any decision rest on more than one piece of evidence.
Another detector to try
Another option is the free AI detector from Walter Writes. Its page says checks are free and unlimited, that it detects text from ChatGPT and Gemini, that it gives a percentage score for every scan and that it works in more than 20 languages. You can try Walter's AI detector here. Treat its score like any other detector's: one signal, not proof.
Sources
- Vendor claim
- Independent study
- Independent study
- Independent study
- Independent study
Questions
How accurate is ZeroGPT?
Far less accurate than it claims. ZeroGPT states 98.4% accuracy on its homepage. Independent researchers measured 59% in 2023, sixth of 14 tools, and 46.1% on unedited AI text in 2024, fifth of seven. Both studies used older versions, and every independent result sits well below the vendor figure.
Does ZeroGPT give false positives?
Yes. The RAID benchmark paper found ZeroGPT flagged 1.71% of human texts as AI, the highest rate among the commercial detectors it studied. ZeroGPT was also one of seven detectors that flagged non-native English essays as AI 61.22% of the time on average in the Liang et al. study.
Is ZeroGPT more accurate than GPTZero?
In the 2023 and 2024 academic tests, ZeroGPT scored higher than the GPTZero versions of the time: 59% against 54% and 46.1% against 26.3%. Newer GPTZero results are much stronger, including 97.2% on the RAID leaderboard in October 2025 and a false positive rate of about 0.7% in a 2025 study.
Can ZeroGPT detect edited or paraphrased AI text?
Poorly. In Perkins et al. (2024), ZeroGPT's accuracy fell from 31.3% to 17.3% after simple evasion edits to AI text. The seven detectors in that study averaged 22.2% after the same edits, so ZeroGPT was mid-pack, but none of them handled rewritten AI text well.
Should a teacher rely on a ZeroGPT score?
No. A ZeroGPT score alone is weak evidence. Its independent accuracy results sit between 46.1% and 59%, it had the highest false positive rate of the commercial tools in the RAID paper, and it was part of a study showing heavy bias against non-native writers. Use drafts and a conversation instead.
