Is GPTZero Accurate? What the Studies Show
GPTZero is the most-tested AI detector, and the studies disagree. Here are GPTZero’s own figures, eight independent studies with their dates and samples, and what they mean for plain, human and humanized text.
Short answer: GPTZero is the most-studied AI detector, and the studies do not agree. On plain AI text it usually scores well. Its false positive rate on human writing has ranged from none to half of the human documents tested, depending on the study, the version and the kind of writing. Its newest figures come from GPTZero itself.
What GPTZero says about its accuracy
GPTZero's own pages give several figures, from different datasets, metrics and model versions:
- 99.9% accuracy, the homepage headline, covering human and AI samples (GPTZero).
- 95.7% of AI text detected at a 1% false positive rate on the RAID benchmark (accuracy post). This matches the RAID leaderboard entry, which GPTZero submitted itself.
- 99.76% accuracy and a 0.08% false positive rate for model 4.3b on its own benchmark, and 91.8% recall on text run through paraphrasers and bypass tools (benchmarking page).
- 1 false positive in 10,402 student essays for its newest model, GPTZero 4o, released October 2026 (4o post).
GPTZero also says results should not be used to punish anyone or as a final verdict, that it is strongest on English prose, and, in its FAQ, that it is not trained for AI text that has been heavily modified (FAQ). Newer pages claim the opposite for paraphrased text.
What independent studies found
| Study | Tested | What it found for GPTZero |
|---|---|---|
| Liang et al. 2023, *Patterns* | March 2023 | Flagged 52% of TOEFL essays by non-native English writers as AI, and none of the US 8th-grade essays |
| Weber-Wulff et al. 2023, *Int. J. Educational Integrity* | Spring 2023 | Flagged 9 of 18 human or machine-translated human documents, the most of 14 tools, but missed only 3 of 36 AI documents |
| Orenstrakh et al. 2023 (preprint) | 2023 | Flagged 52 of 114 computer science submissions written before ChatGPT existed |
| Dugan et al. 2024, RAID (ACL 2024) | January 2024 | 66.5% average accuracy at a 5% false positive rate; strong on chat models (about 98%), weak on some open-source model output (under 10%), and unusually robust to adversarial edits |
| Perkins et al. 2024 (preprint) | Autumn 2023 | Caught 26% of AI texts on average, the lowest of seven tools; flagged 1 of 10 human controls |
| Pratama 2025, *PeerJ Computer Science* | 2025 | 97% accuracy and no false positives on 72 published abstracts |
| Dik et al. 2025 (preprint) | 2025 | Caught all 28 AI essays; flagged 8 of 50 human essays (16%) |
| Karr et al. 2026 (preprint) | 2026 | Flagged 8.9% of abstracts published in 2023–25; after Undetectable AI humanization, 96% of AI abstracts went undetected |
Sources: Liang et al., Weber-Wulff et al., Orenstrakh et al., RAID, Perkins et al., Pratama, Dik et al., Karr et al..
What the evidence adds up to
- Plain AI text: usually caught, especially from chat models.
- Human writing: the older studies found high false positive rates, especially for non-native English writers. Newer versions claim far lower rates, but the newest figures are GPTZero's own.
- Edited or humanized text: the most recent independent study found that a humanizer hid almost all AI abstracts from it.
- Versions matter: GPTZero has changed models several times, so a 2023 result may not describe today's detector, and today's claims are not yet independently tested.
How to read a GPTZero result
- Check longer text: GPTZero says document-level results are more reliable than paragraph or sentence results.
- Treat the score as one signal. GPTZero itself says it should not be used to punish students.
- Keep your drafts and version history. If GPTZero flags work you wrote, they are your strongest evidence.
FAQ
Is GPTZero accurate? On unedited AI text, usually. On human writing and edited AI text, independent results vary widely, and its newest accuracy figures come from GPTZero itself.
Does GPTZero give false positives? Yes. Independent studies found rates from none to 16% of human essays, 46% of computer science submissions written before ChatGPT, half of human documents in one 2023 study, and 52% of essays by non-native English writers. GPTZero says its newest model has a far lower rate.
Can GPTZero detect humanized text? GPTZero says its newer models handle paraphrased and bypassed text, with 91.8% recall on its own test. An independent 2026 study found that 96% of AI abstracts went undetected after humanization.
Is GPTZero free? It scans up to 10,000 characters without an account, and a free account covers up to 10,000 words a month. Its Terms describe a 14-day trial that is charged automatically at the end (students page, terms).