How accurate is GPTZero? Its own claims and what independent tests found

How accurate is GPTZero? On plain, unedited AI text it is quite accurate, and in recent published tests it rarely flags human writing. Its weak spots are AI text that has been run through a humanizer, text from newer reasoning models, and long documents that mix human and AI writing, where independent studies found it missed a large share. GPTZero advertises 99% accuracy; independent figures range from close to that down to very low, depending on the test.

What does GPTZero claim about its accuracy?

At the time of writing, GPTZero's homepage shows a 99% accuracy figure, and its FAQ says it classifies AI writing correctly "99 out of 100 times", citing its own testing and benchmarks by partners at Penn State's AI Research Lab. It also reports accuracy above 99% when results are filtered to modern models such as GPT-4, and 96.5% on documents mixing human and AI writing. It says it is "de-biased" for people writing in English as a second language.

The same page is candid about limits. It says no AI detector is 100% accurate, that accuracy is strongest on longer English prose, and that results "should not be used to punish or as the final verdict". Without an account, you can check up to 10,000 characters per scan.

These are vendor claims. The rest of this article looks at numbers from outside GPTZero. How each study was set up is explained in are AI detectors accurate?; here we stick to GPTZero's results.

How does GPTZero score on the RAID benchmark?

RAID is an open benchmark from ACL 2024 with a public leaderboard that anyone can submit a detector to. GPTZero highlights its RAID result, and we checked it. On the leaderboard at the time of writing, GPTZero catches 95.7% of AI texts while flagging 1% of human texts, with no attacks applied. With the benchmark's adversarial attacks included, that falls to 92.9%. In both views, more than 20 other entries score higher, including Winston AI, Grammarly and QuillBot.

Two caveats apply. RAID's AI texts come from older models such as ChatGPT, GPT-4, GPT-2, Mistral and Llama chat models, and each text is either fully human or fully AI. It says little about mixed documents or this year's models.

What did the University of Chicago study find?

The 2025 Becker Friedman Institute working paper by Brian Jabarian and Alex Imas ran GPTZero against Pangram, Originality.ai and an open-source model on six everyday genres. For GPTZero, three findings stand out:

• GPTZero and Originality.ai formed a second tier behind Pangram. If keeping false positives low matters most, the authors found GPTZero the better of the two.
• Neither was judged suitable for very short passages of under 50 words.
• After the AI passages were rewritten by the StealthGPT humanizer, GPTZero "largely" lost its ability to detect them, with miss rates around 50% and above across most genres and models.

GPTZero disputes part of this. In January 2026 it published a re-run of the same benchmark, saying the researchers had used the wrong field of its API. With its December 2025 model and its own mapping of labels, GPTZero reported catching 99.3% of AI passages with a 0.05% false positive rate, slightly ahead of Pangram on its scoring. That re-run is GPTZero's own and has not been independently checked, and it reports no results on humanized text or short passages.

How did GPTZero do against expert human readers?

Russell, Karpinska and Iyyer's ACL 2025 study pitted detectors against people who use AI for their own writing. In its set of 300 non-fiction articles, GPTZero caught 85.3% of the AI-written ones while wrongly flagging 0.7% of human ones, a very low false positive rate. The average hides a clear pattern by source:

• On articles from GPT-4o, it caught 100%.
• On articles from OpenAI's o1-Pro reasoning model, it caught 76.7%.
• On o1-Pro articles written to dodge known AI tells, the study's humanized set, it caught 46.7%.

The authors' verdict was that GPTZero "struggles significantly" on o1-Pro text, with and without humanization. Each condition had 30 AI articles, so single figures can move by several points. Pangram and the expert readers did better overall on the same articles, as we set out in how accurate is Pangram.

How does GPTZero handle long papers and mixed writing?

A June 2026 study from Vrije Universiteit Brussel ran four detectors on 160 long academic papers. GPTZero's results on AI content were weak:

• Of the 40 papers that ChatGPT's Deep Research mode wrote from start to finish, 70% got a GPTZero score under 20% and the rest scored between 20% and 40%.
• On 40 hybrid papers, where GPT-4o had written the literature review and conclusion, none of GPTZero's scores fell in the right range. On the humanized versions, one did.
• On 40 human papers by graduate students writing in a second language, GPTZero gave some small non-zero scores, but all stayed under 20%.

A 2026 University of Notre Dame study of published academic abstracts adds two findings. GPTZero flagged 38% to 49% of abstracts that an AI model had only lightly refined, which matters if your institution allows that kind of help. And after the Undetectable AI humanizer rewrote the AI versions, GPTZero missed about 96% of them. Our article on whether AI humanizers work compares that result with other detectors.

Does GPTZero flag non-native English writers?

It did in 2023, when a Stanford study found seven detectors, GPTZero among them, flagged most essays by English learners as AI; we go through that study in AI detectors and non-native English. GPTZero says it has since trained its model to reduce false positives on such writing. The Brussels study found no GPTZero score above 20% on 40 papers by non-native graduate students, which is encouraging, though 40 papers is a small sample.

So, how accurate is GPTZero overall?

It is a capable detector for unedited text from widely used models, with a false positive rate that recent studies put very low. It is much less dependable on humanized text, on output from newer reasoning models, and on long documents where AI wrote only some sections. Treat a GPTZero score as one piece of evidence, as GPTZero itself advises.

If you want a free check with no account, our own detector has no character limit, scans every word rather than a sample and runs in your browser without uploading anything. Its report adds a sentence map and an AI-likeness score, and how it is tested is set out in our open benchmark. Our list of GPTZero alternatives sets out where each tool fits.

Sources

1. GPTZero homepage and FAQ, checked October 2026
2. GPTZero: RAID benchmark announcement (15 October 2025), checked October 2026
3. RAID Benchmark leaderboard, checked October 2026
4. Dugan et al., RAID: A Shared Benchmark for Robust Evaluation of Machine-Generated Text Detectors (ACL 2024), checked October 2026
5. Jabarian and Imas, Artificial Writing and Automated Detection, BFI Working Paper 2025-116, checked October 2026
6. GPTZero: re-evaluation on the Chicago Booth benchmark (12 January 2026), checked October 2026
7. Russell, Karpinska and Iyyer, People who frequently use ChatGPT for writing tasks are accurate and robust detectors of AI-generated text (ACL 2025), checked October 2026
8. Van Vlasselaer, Van Droogenbroeck and Spruyt, Who wrote this? Evaluating the reliability of AI detection tools in higher education (International Journal for Educational Integrity, 2026), checked October 2026
9. Karr, Khvatskii, Hua and Chawla, Why AI Detection Fails for Academic Integrity (ACM AI Summit 2026), checked October 2026
10. Liang et al., GPT detectors are biased against non-native English writers (Patterns, 2023), checked October 2026