Is GPTZero accurate?
Recent results put GPTZero's false positive rate near 0.7% and its benchmark detection rate at 97.2%, but older academic tests scored it poorly and humanized or very short text still trips it.
Short answer
GPTZero is mostly reliable on unedited text. A 2025 study found it wrongly flagged about 0.7% of human passages, and it reached 97.2% detection on the RAID benchmark with vendor-submitted results. Older versions scored only 54% in 2023, and humanized text still fools it.
Is GPTZero accurate? It is much better than it was. In the first wave of academic tests, GPTZero sat near the bottom of the table. The newer evidence is stronger: a low false positive rate in a large 2025 study and a high score on a public benchmark. Its weak spots are text run through humanizer tools and very short passages. Every figure below is dated and linked, and the vendor's own number is labelled as a vendor claim. For how we weigh sources, see how we score AI accuracy.
GPTZero test results
- Independent study
- Benchmark
- Vendor claim
| Test | What was measured | Result | Date | Source |
|---|---|---|---|---|
| GPTZero homepageVendor's own figure, not independently checked | Detection rate (vendor's headline figure, own demo dataset) | 99.9%, with 0 humans flagged and 0 AI missed | Source 1Vendor claim | |
| Jabarian and Imas 2025, 1,992 passages in six genres | False positive rate on human text | About 0.7% | Source 2Independent study | |
| RAID leaderboard | Detection rate at a 5% false positive rate | 97.2% (all), 98.4% (no adversarial attacks) | Source 3Public benchmark (vendor-submitted results) | |
| RAID leaderboard | Detection rate at a 1% false positive rate | 92.9% | Source 3Public benchmark (vendor-submitted results) | |
| Weber-Wulff et al. 2023, 54 documents | Detection rate (binary accuracy) | 54%, 8th of 14 tools | Source 4Independent study | |
| Perkins et al. 2024 | Detection rate on unedited AI text | 26.3%, 7th of 7 detectors | Source 5Independent study | |
| Liang et al. 2023, 91 TOEFL essays | False positive rate on non-native English essays (average of seven detectors, GPTZero included) | 61.22% | Source 6Independent study |
What the GPTZero accuracy numbers mean
The numbers split cleanly by date. Before 2025, GPTZero did badly. Weber-Wulff and colleagues scored it 54% in 2023, and it classified only 6 of 9 fully human texts correctly in their binary scheme. Perkins and colleagues found it caught 26.3% of unedited AI text in 2024, the lowest of seven detectors. Both papers tested older versions and warn their results may not match current tools.
The 2025 results are far better. Jabarian and Imas, in a working paper from the University of Chicago's Becker Friedman Institute, ran 1,992 human passages across six genres and found GPTZero flagged about 0.7% of them as AI. That rate stayed close to 0.7% across thresholds. The same paper placed GPTZero in a second tier behind Pangram.
On the RAID leaderboard, GPTZero caught 97.2% of AI text when held to a 5% false positive rate, and 92.9% at a stricter 1% rate. RAID is a public benchmark, but GPTZero ran its own detector and submitted the predictions.
GPTZero's homepage shows a headline figure of 99.9%, with no humans flagged and no AI missed. That result comes from GPTZero's own demo dataset, under conditions GPTZero set.
Where GPTZero goes wrong
Humanized text
The Jabarian and Imas paper found GPTZero vulnerable to humanizer tools that rewrite AI output.
Very short passages
The same paper flags short texts as a weak spot. A paragraph or two gives the detector little to work with.
Attack-free versus real-world text
On RAID its rate fell from 98.4% without adversarial attacks to 97.2% with them, and to 92.9% at a 1% false positive limit.
Non-native writers
GPTZero 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.
How GPTZero compares in the same tests
Only compare detectors inside one study.
| Test | GPTZero | Comparison |
|---|---|---|
| RAID leaderboard, 5% false positive rate | 97.2% | Grammarly 99.5% |
| RAID leaderboard, 1% false positive rate | 92.9% | Grammarly 98.2% |
| Weber-Wulff et al. 2023, binary accuracy | 54% | Turnitin 76%, ZeroGPT 59% |
| Perkins et al. 2024, unedited AI text | 26.3% | Turnitin 61%, ZeroGPT 46.1% |
On the RAID leaderboard, GPTZero trails Grammarly's detector at both settings, and both sets of results were submitted by the vendors. In the older academic tests, GPTZero ranked behind both Turnitin and ZeroGPT. Those older tests are the only ones that put all three side by side. For the full picture, see are AI detectors accurate.
Non-native English writers and GPTZero
GPTZero 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 detector. US eighth graders' essays drew almost no false flags. The authors tie the bias to simpler vocabulary. The study used 2023 versions, and GPTZero's 2025 false positive rate is far lower. If English is your second language, a GPTZero flag still deserves a second look.
What a GPTZero score can and cannot prove
No detector score should be the only evidence in a misconduct case.
It can show
- A GPTZero score can tell you that a passage reads like model output to a detector with a low recent false positive rate. That makes it a better signal than many free tools.
It cannot prove
- A GPTZero score cannot tell you who wrote the text, how much AI was used, or whether a humanizer tool or heavy editing hid AI use. A false positive rate near 0.7% is low, but across the thousands of essays a school scans each term it still means real students get flagged.
What to do if GPTZero flags your work
Write down the score, the date, and the sentences GPTZero highlighted, and save a screenshot.
Collect your drafts, outline, notes and version history from Google Docs or Word.
Ask whether the flagged text is short; GPTZero is weaker on very short passages, per the 2025 study.
If English is not your first language, point to the Liang et al. study, which included GPTZero.
Offer to explain 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
- GPTZero results file, released 2025-10-07, and Grammarly results fileBenchmark
- Independent study
- Independent study
- Independent study
Questions
How accurate is GPTZero?
Recent evidence says fairly accurate on unedited text. GPTZero reached 97.2% detection on the RAID leaderboard at a 5% false positive rate, with results it submitted itself. Its own homepage shows a headline figure of 99.9% on its demo dataset. Older academic tests scored earlier versions at 54% in 2023 and 26.3% in 2024.
Does GPTZero give false positives?
Yes, but rarely in the latest study. Jabarian and Imas found GPTZero flagged about 0.7% of 1,992 human passages as AI in 2025. In 2023, Weber-Wulff and colleagues found it classified only 6 of 9 fully human texts correctly. GPTZero was also part of the Liang study on non-native English writers.
Is GPTZero better than Turnitin?
In the 2023 and 2024 academic tests, no: Turnitin scored 76% against 54% and 61% against 26.3%. Those were older versions. GPTZero's newer results, 97.2% on RAID and about 0.7% false positives in 2025, come from different tests, so they cannot be ranked against Turnitin's older figures.
Can GPTZero detect humanized or paraphrased AI text?
Not reliably. The 2025 Jabarian and Imas paper found GPTZero vulnerable to humanizer tools. On the RAID leaderboard its detection rate dropped from 98.4% on text without adversarial attacks to 97.2% across all text. A low score on heavily edited text deserves the same caution as a high one: neither shows who wrote it.
Can GPTZero check short text?
It is weaker on very short passages, according to the 2025 Jabarian and Imas study. A few sentences give a detector little evidence to work with, so a short-text score carries less weight than one on a full essay. Run longer samples where you can, and never treat a short-text flag as proof.
