Is ZeroGPT Accurate? What Our 60-Text Test Found
We ran 60 AI-written texts through ZeroGPT before and after rewriting. It caught 58 of the unedited texts, missed two short emails, and stopped flagging text once it was rewritten in depth. Here is the test, what ZeroGPT claims, and how to read its score.
Short answer: on unedited AI text, ZeroGPT was accurate in our test. It flagged 58 of 60 AI-written texts. It was much less reliable once the text had been rewritten: after an in-depth rewrite it labeled all 60 as human-written. Our test had no human-written texts; independent studies report false positive rates from none to 48%, depending on the writing.
Our test
In October 2026 we asked Claude and DeepSeek to write 60 texts of about 250 words (199–284), 30 each, across six kinds of writing: academic essays, research abstracts, work emails, business reports, blog posts and personal statements. Every text was scored through ZeroGPT's detection API, the same call our own AI Detector uses: once unedited, once after DeepSeek rewrote it with a one-line "sound human" prompt, and once after a full rewrite with RewriteMax in Balanced mode.
| ZeroGPT result | Unedited AI text | One-line prompt rewrite | Full rewrite (RewriteMax Balanced) |
|---|---|---|---|
| Average AI score | 89% | 77% | 12% |
| Median AI score | 93% | 83% | 8% |
| Labeled human-written | 2 of 60 | 7 of 60 | 60 of 60 |
| Scored under 20% AI | 1 of 60 | 3 of 60 | 44 of 60 |
What this says about ZeroGPT's accuracy
On unedited AI text it rarely missed. ZeroGPT flagged 29 of 30 Claude texts and 29 of 30 DeepSeek texts. Both misses were short work emails; one scored 0%. Emails averaged 68% AI, against 90–96% for essays, abstracts, reports, blog posts and personal statements, so short, plain business writing is where it was least sure.
On rewritten text, the verdict follows how much the wording changed. A one-line rewrite moved only five more texts to a human-written label. A full rewrite moved all of them. That is a limit of pattern-based detection rather than of ZeroGPT alone: it measures how text is written, not who wrote it. It is also not a quality check. Our reviewers found that heavy automated rewriting often introduced grammar errors and changed details, and ZeroGPT still labeled that text human-written.
Labels and scores do not always line up. One text scored 45% AI and was labeled human-written. Read the percentage and the highlighted sentences, not only the label.
Human writing: not tested here. Our test contained no human-written texts, so it says nothing about false positives. ZeroGPT says its false positive rate is under 1% and its accuracy 98.4%, without publishing the test conditions behind those figures. It also says polished, formulaic writing can resemble AI.
What independent studies found
Several published studies have tested ZeroGPT by name. They used older versions and different texts from ours, so read them as context rather than as one score:
| Study | Tested | What it found for ZeroGPT |
|---|---|---|
| Liang et al. 2023, *Patterns* | March 2023 | Flagged 48% of 91 TOEFL essays by non-native English writers as AI, and none of 88 US 8th-grade essays |
| Weber-Wulff et al. 2023, *International Journal for Educational Integrity* | Spring 2023 | No false accusations on 18 human documents; 13 of 36 AI documents went undetected, mostly ones edited by a person or paraphrased with QuillBot |
| Dugan et al. 2024, RAID (ACL 2024) | 2024 | Could not be held to a 5% false positive rate: on human text it plateaued at 16.9%, and higher on recipes (52%) and Wikipedia text (48%). Paraphrasing lowered its accuracy to 47% |
| Pratama 2025, *PeerJ Computer Science* | 2025 | Accuracy 64% on 72 published abstracts and their AI versions, with a 17% false positive rate |
| Yeadon et al. 2024 (preprint) | 2024 | No false positives and one miss on 30 human and 30 GPT-4 physics essays |
So independent false positive rates range from none in small essay samples to 17% on abstracts and 48% on essays by non-native English writers. That is the gap to keep in mind next to ZeroGPT's own under-1% figure (Liang et al., Weber-Wulff et al., RAID, Pratama, Yeadon et al.).
What ZeroGPT says about itself
- Accuracy: 98.4%, with a false positive rate under 1% (ZeroGPT). No test set or method is published with these figures.
- Length: it recommends at least 150–200 words (FAQ).
- Rewritten text: it says paraphrasing can reduce some of the signals it looks for.
- Models: it lists GPT models, Gemini, Claude, DeepSeek, Grok and LLaMA among the families it covers.
How to use a ZeroGPT score
- Check at least 200 words. Short texts gave the least reliable results in our test and in ZeroGPT's own guidance.
- Treat it as a signal, not proof. A high score on your own writing is not evidence that you used AI, and a low score does not prove you did not.
- Look at the sentences it highlights. They tell you more than the overall percentage.
- Keep your drafts and version history. If any detector flags work you wrote, that history is your strongest evidence.
FAQ
Is ZeroGPT accurate? On unedited AI text, it was in our test: 58 of 60 flagged. On text rewritten in depth it was not, and its false positive rate on human writing has not been independently confirmed by our test.
Can ZeroGPT detect ChatGPT, Claude and DeepSeek? It flagged 29 of 30 Claude texts and 29 of 30 DeepSeek texts in our test. We did not include ChatGPT text in this test.
Does ZeroGPT give false positives? ZeroGPT says under 1%, without published test conditions. Independent studies found anything from none in small essay samples to 17% on published abstracts and 48% on essays by non-native English writers.
Is ZeroGPT more accurate than GPTZero? We have not tested them side by side, so we make no claim either way.