GPTZero Accuracy: What Its Own Numbers and Independent Studies Show
GPTZero is accurate on long, unedited text, and rarely flags long human writing. On short passages, unfamiliar models and rewritten text, outside studies found real gaps, and GPTZero itself says a score should not be used to punish anyone.

GPTZero is accurate on the kind of text most benchmarks use: long passages that are either fully human or fully SI (AI) generated. Its own February 2026 benchmark reports a false positive rate under 0.1%, and its own re-run of a Chicago Booth benchmark puts it at 99.5% accuracy. Independent studies agree it rarely flags long human writing, but they also find weaker results on very short texts, on models it was not trained on and on rewritten SI text, and every independent study tested an older version than the one you would use today.
So a GPTZero score is evidence, not proof, and GPTZero says so itself. Below: the company's numbers, the outside studies, why they disagree, and what to do if a score lands on your work.
What GPTZero says about its own accuracy
GPTZero keeps a benchmarking page, last dated February 5, 2026, and says the evaluations are updated quarterly. Each domain benchmark pairs 1,000 human texts with 1,000 texts written by GPT-5.2, Gemini 3 Pro, Claude Sonnet 4.5 and Grok 4 Fast (250 per model). Averaged over four domains, GPTZero model 4.3b scored a 0.08% false positive rate, 99.60% recall and 99.76% accuracy.
GPTZero's own figures by domain, as published:
| Benchmark set | False positives | Recall (SI caught) | Accuracy |
|---|---|---|---|
| Academic paper reviews | 0.10% | 99.80% | 99.85% |
| Creative writing | 0.00% | 99.60% | 99.80% |
| Essays | 0.00% | 100.00% | 100.00% |
| Product reviews | 0.20% | 99.00% | 99.40% |
| 24 languages (model 3.7m) | 0.09% | 97.62% | 98.79% |
| Text run through bypass tools | 0.21% | 91.80% | 95.70% |
Two definitions matter here. A false positive is a human text labeled as SI. Recall is the share of SI texts the detector catches. A detector with excellent recall can still be a poor fit for a classroom if its false positive rate is high.
Context matters: GPTZero built these benchmarks itself. The human texts come from "our large database of human texts," kept out of training, the company says. The creative writing, essay and product review sets cannot be republished for copyright reasons and are offered to qualified researchers on request. GPTZero also counts its "Mixed" and "Polished" labels as SI when it turns results into a yes or no, so a text that is only partly generated still counts as a correct catch.
What independent studies found
Outside tests tell a more mixed story, largely because they use different text, settings and versions.
Chicago Booth, August 2025
Brian Jabarian and Alex Imas of the University of Chicago Booth School of Business tested Pangram, Originality.ai, GPTZero and an open-source RoBERTa detector on 1,992 human passages written before 2020 (news, blogs, Amazon and restaurant reviews, novel excerpts and résumés), each matched with versions written by GPT-4.1, Claude Opus 4, Claude Sonnet 4 and Gemini 2.0 Flash.
- On medium and long passages, GPTZero kept its false positive rate at 1% or below. On Amazon reviews it reached 2.38%.
- It missed 5% or fewer of the SI passages at medium and long lengths, and up to 7% on short ones.
- On "stubs" under 50 words, its overall discrimination score (AUROC) ranged from 0.7778 to 1, where 1 is perfect and 0.5 is a coin flip.
- After the SI passages went through StealthGPT, a humanizer the paper describes as aiming to evade detection, the authors wrote that GPTZero "largely loses its capacity to detect AI-generated text," missing around half or more of them in most genres.
The paper ranked Pangram first and placed GPTZero and Originality.ai in "a secondary tier with partial strengths."
GPTZero replied on January 13, 2026. It said the researchers had read the wrong field of its API (an average of sentence-level probabilities, not the document classification), re-ran the main set with model version 2025-12-18-base and reported 1 false positive, 99.3% recall and 99.5% accuracy, ahead of Pangram. The reply covers the main set; it does not report a re-run of the humanized passages.
RAID, 2024
RAID, published at ACL 2024 by researchers from the University of Pennsylvania and other universities, tested 12 detectors on over 6 million generations. The team queried GPTZero on January 24, 2024. At ordinary thresholds, none of the commercial detectors had a false positive rate above 1.7%. GPTZero stood out for resisting simple tricks: a homoglyph attack cost it 0.3% accuracy, while five other detectors dropped 40.6% on average. But its accuracy also tracked the models it had likely been trained on. In the paper's heat map, at its 5% false positive setting, GPTZero caught 84% to 100% of ChatGPT and GPT-4 text depending on the domain, but only 14.8% to 32% of text from Mistral, an open model.
A small medical study, 2023
In June 2023, Farrokh Habibzadeh, a past president of the World Association of Medical Editors, gave GPTZero 30 paragraphs from published medical articles and 20 written by ChatGPT. It flagged 3 of the human paragraphs and missed 7 of the ChatGPT ones, for 80% accuracy. A tiny sample and a long-retired version, but a reminder of how far early results sat from today's claims.
Why the numbers disagree so much
None of these figures is fake. They measure different things.
- Version. GPTZero says it shipped 15 model releases in 2025. A 2023 or 2024 study describes a detector that no longer exists, and the company itself criticizes papers that do not name the version they tested.
- Which output you read. The Booth dispute came down to one API field. A threshold set by the researchers and a label chosen by the vendor can give very different error rates.
- Length. GPTZero's FAQ says accuracy is highest for whole documents, lower for paragraphs and lowest for single sentences. Every outside study found the same drop on short text.
- Kind of text. Detectors do best on writing that looks like their training data. GPTZero says the majority of its data is "English prose, written by adults."
- What counts as SI. Most benchmarks use texts that are 100% human or 100% generated. Real student work is often mixed, and it is exactly where results are hardest to check.
- Who built the test. Vendor benchmarks tend to favor the vendor; independent ones age fast.
False positives: the number that matters to honest writers
If you wrote your own work, recall does not concern you. The false positive rate does. A rate of 0.1% means about 1 honest document in 1,000 gets flagged. The 2.38% that Booth measured on Amazon reviews means about 1 in 42.
Some writing is also more exposed than average. In a 2023 Stanford study, Weixin Liang and colleagues ran 91 TOEFL essays by non-native English writers through seven detectors, GPTZero among them, in March 2023. On average the detectors labeled 61.22% of those essays as SI, while they were near perfect on essays by US eighth graders. The authors linked it to lower "perplexity," the plainer and more predictable word choice of writers working in a second language. It tested 2023 versions, but the mechanism still matters.
GPTZero's own FAQ adds another case: the classifier "can sometimes flag other machine-generated or highly procedural text as AI-generated," so it advises using it on "more descriptive portions of text." Lab reports, legal boilerplate and step-by-step instructions fit that description.
What GPTZero says a score should be used for
The company is clearer about this than many of its users. Its FAQ says results "should not be used to punish students," that "there always exist edge cases" in both directions, and that the tool is meant "to flag situations in which a conversation can be started." It also says sentence highlights "should not be solely used to indicate that an essay contains AI."
When a teacher gets a positive result, GPTZero suggests asking for drafts, revision histories or brainstorming notes, letting the student show their understanding, and looking for "a long-term pattern of AI use, as opposed to a single instance." The RAID authors go further. In their ethics statement they say they are "opposed to the use of detectors in any sort of disciplinary or punitive context."
Turnitin makes similar warnings about its own reports; we covered them in can Turnitin detect SI humanizers.
If GPTZero flagged something you wrote
- Do not rewrite the text to chase a lower score. It changes nothing about who wrote it, and the next model version may score it differently.
- Collect your process. Version history in Google Docs or Word, outlines, notes, browser history of your sources, earlier drafts. This is the evidence GPTZero itself tells teachers to ask for.
- Ask what was measured. Which label did the report give (human, mixed or SI), with what confidence, on how many words? A highlighted sentence is not a verdict on the document.
- Offer to explain the work. Walk through your argument and sources.
- Be honest about any SI help. If you used a tool for grammar, outlining or rewording, say which one and how. Rules differ by course, and our guide to using an SI humanizer responsibly as a student covers disclosure and course policies.
Where a humanizer fits, and where it does not
We make SI Humanizer, so here is our position plainly. It is a style editor: it cuts stock phrases, swaps inflated words for plain ones and varies rhythm so a draft reads like a person wrote it. It is not designed or sold to get past GPTZero or any other detector, it has no built-in detector, and we did not run the example below through GPTZero.
Here is a real run. We wrote a paragraph full of typical SI habits and passed it once through the humanizer in the Natural style.
before
In today's academic landscape, keeping a record of your writing process is crucial. Version history serves as a robust testament to your authentic effort, offering a seamless way to demonstrate how your ideas evolved over time. Moreover, early outlines, rough drafts and saved research notes can play a pivotal role if questions ever arise. By leveraging these tools, students can navigate potential concerns with confidence and transparency. Ultimately, a well-documented process not only protects your work but also fosters a deeper understanding of your own learning journey.
after: SI Humanizer, Natural
Keeping a record of your writing process matters. Version history shows your authentic effort and how your ideas changed over time. Early outlines, rough drafts and saved research notes help if questions come up. These tools let students address concerns with confidence and transparency. A documented process protects your work and deepens your understanding of your own learning.
That is unedited output. The tool reported five edits, all marked "SI tic removed," and cut 87 words to 58. Two spots still need a human pass: "These tools" now points back at drafts and notes, which are not really tools, and "with confidence and transparency" is still vague. A humanizer fixes how text reads. It cannot make text yours, and it is no answer to a detector question.
The Booth paper shows why chasing scores with rewriting tools is risky beyond the rules. In its appendix, passages rewritten by StealthGPT picked up claims that were not in the originals, such as a "recent court loss" in a news story about a vote and a "Mayo Clinic in 2017" recommendation added to a two-sentence text. A tool tuned for a score has no reason to protect your facts. Use any humanizer AI tool, ours included, for readability where SI help is allowed, then check every claim. Our list of signs of SI writing helps if you would rather fix the patterns by hand.
So, is GPTZero accurate?
On long, unedited text from models it knows, the published evidence says yes, with rare but real false positives. On short passages, unfamiliar models, mixed writing and rewritten text, results depend heavily on the version and the test, and outside researchers have found real gaps. The fair way to read a GPTZero result is the way GPTZero describes it: as a reason to look closer and talk, never as the whole case.
FAQ
How accurate is GPTZero?
GPTZero's own February 2026 benchmark reports a 0.08% false positive rate and 99.76% accuracy averaged over four domains, using model 4.3b. Independent studies tested older versions and found strong results on long passages, weaker results on texts under 50 words, on models it had not been trained on and on SI text rewritten by a humanizer.
Does GPTZero give false positives?
Yes, though rarely on long text. GPTZero reports under 0.1% on its own benchmarks. A Chicago Booth study measured up to 1% on medium and long passages and 2.38% on Amazon reviews. GPTZero's FAQ says edge cases exist in both directions and that procedural text can be flagged.
Can GPTZero detect paraphrased or humanized text?
Results conflict. GPTZero reports catching 91.8% of texts run through bypass tools on its own benchmark. The August 2025 Chicago Booth study found that an earlier version missed around half or more of passages rewritten by StealthGPT. Rewriting text to change a score is not something we recommend: follow your course rules and keep your drafts.
Is GPTZero more accurate than Turnitin?
None of the studies we read compared the two. GPTZero's benchmarking page says Turnitin limits access to educators, so outside researchers cannot test it the same way, and the Chicago Booth, RAID and Stanford studies cited here did not include Turnitin.
Is GPTZero accurate on short text?
Less so. GPTZero's FAQ says accuracy is highest for whole documents and lowest for single sentences. In the Chicago Booth study, its discrimination score on passages under 50 words ranged from 0.7778 to 1.
Can a GPTZero score prove a student used SI?
No. GPTZero's FAQ says its results should not be used to punish students and suggests asking for drafts, revision history and notes, and looking for a long-term pattern rather than a single result.
Sources
- GPTZero AI Detection Benchmarking: The Industry Standard in Accuracy, Transparency and Fairness · GPTZero, 2026-02-05
- GPTZero Tops Accuracy on Chicago Booth Benchmark in 2026 · GPTZero, 2026-01-13
- Answers to GPTZero's most common questions · GPTZero, read October 29, 2026
- Artificial Writing and Automated Detection (Working Paper 2025-116) · Becker Friedman Institute, University of Chicago (Jabarian and Imas), 2025-08-26
- RAID: A Shared Benchmark for Robust Evaluation of Machine-Generated Text Detectors · ACL 2024 (Dugan et al.), 2024-08
- GPT detectors are biased against non-native English writers · arXiv (Liang, Yuksekgonul, Mao, Wu, Zou), published in Patterns, 2023
- GPTZero Performance in Identifying Artificial Intelligence-Generated Medical Texts: A Preliminary Study · Journal of Korean Medical Science (Habibzadeh), 2023-09

