The short answer
GPTZero is reasonably good at the easy cases: text that is entirely written by an AI model, or entirely by a person. It is less dependable on short passages, AI text that has been heavily edited, and documents that mix the two, and GPTZero says so itself. Its headline “99% accuracy” is the company’s own figure, and the independent research is thinner and more mixed. Treat a GPTZero result as a reason to look closer, never as proof.
What GPTZero claims about its accuracy
On its homepage FAQ, GPTZero describes itself as having “a 99% accuracy rate when spotting AI-generated text vs. human writing”. The more specific numbers come from a post on the RAID benchmark, published by GPTZero on October 15, 2025:
| GPTZero’s claim | What it refers to |
|---|---|
| 95.7% of AI texts detected | On RAID, while about 1% of human texts were wrongly flagged as AI |
| Over 99% | The same benchmark, filtered to newer models (discontinued models such as GPT-3.5 removed) |
| 96.5% | Accuracy on “mixed documents” of human and AI writing, per the homepage FAQ |
RAID is a third-party dataset that GPTZero describes as covering 672,000 texts across 11 domains, 12 language models and 12 adversarial attacks. Two caveats sit in GPTZero’s own post. First, it notes that RAID, like many benchmarks, “considers only fully human or fully AI texts”, which is not what most real student work looks like. Second, the comparison it publishes is GPTZero running and reporting its own results. The same FAQ adds that GPTZero does not believe “any AI detector is perfect” and that accuracy goes up as more text is submitted.
What independent tests found
Published, independent tests of GPTZero specifically are few, and most are small. One that ranks for this question is Dik, Erdem and Dik (2025), a preprint that ran 28 AI-generated and 50 human-written essays of different lengths through GPTZero. In the authors’ summary, a vast majority of the AI papers were detected accurately, scoring 91% to 100% AI, while the human essays gave mixed results, including a handful of false positives. Their conclusion is that GPTZero is effective at flagging purely AI-generated text but its reliability in confirming human-written text is limited, and that educators should be cautious about relying on detection tools alone.
That pattern, strong on AI text, weaker at clearing human text, is the one that matters for students, because a wrong “AI” label on your own work is the costly error. A 78-essay study cannot give a reliable error rate, so read it as a warning sign rather than a measurement. Reviews that quote a single independent percentage for GPTZero usually do not name the study behind it, and we have left those numbers out. For the wider evidence on detectors in general, see do AI detectors work?
How GPTZero decides
Many reviews still say GPTZero measures “perplexity” (how predictable the wording is) and “burstiness” (how much that varies across the text). GPTZero’s support centre says it no longer uses either for detection: it moved to a deep-learning model in autumn 2023. The general idea is unchanged, though. A model trained on large amounts of human and AI writing learns which patterns are typical of each and places a new text on one side of a boundary. Because the two kinds of writing overlap, some texts land on the wrong side. Our guide on how AI detectors work explains this in more detail.
What you see is not a single yes or no. According to GPTZero’s guide to reading its results, the model sorts a text into three classes: written by a human, written by AI, or a mix of the two. Each result carries a confidence level. High confidence means the text resembles what the model was trained on; low confidence means it does not, and GPTZero says those results should be read with caution.
A worked example
Say a student pastes a 600-word essay and sees roughly 70% human, 20% mixed and 10% AI, with low confidence. That does not mean 10% of the essay was written by AI. It is a probability across three classes, and the low confidence is the model saying the text is unlike its training data. The fair reading is “probably human, uncertain”, not “partly AI”. A high-confidence 95% AI result on a full essay is a stronger signal, but still a prediction that a person has to check against the writing process.
Where GPTZero struggles, in its own words
GPTZero’s support article on the limitations of its classifier is more candid than its homepage. It lists:
- Short text. Document-level results are more reliable than paragraph-level ones, and paragraph-level results more reliable than sentence-level ones. A highlighted sentence is the weakest signal on the page.
- Text unlike its training data. The model works best on text similar to what it was trained on, which GPTZero describes as mostly English prose written by adults. Writing by younger students or in other languages is further from that.
- Heavily edited AI text. The classifier is not trained to catch AI text that was substantially modified after it was generated. GPTZero has a separate paraphrase detection feature, which it describes as in beta and “not infallible”.
- Procedural writing. Highly procedural or formulaic text, such as instructions, can be flagged as AI, so GPTZero suggests using it on more descriptive passages.
On languages, GPTZero’s language support page says it detects AI text in English and more than 20 other languages, including Spanish, French, German, Portuguese, Chinese and Japanese.
False positives: the error that matters
A false positive is human writing labelled as AI. GPTZero’s explainer on how it works says it prioritizes minimizing them, and its homepage FAQ claims to train its model to keep the false positive rate for writing by English learners at 1%. Those are, again, its own statements.
The concern about English learners comes from research such as Liang and colleagues (2023), which tested seven widely used GPT detectors on 91 TOEFL essays written by non-native English speakers and found an average false positive rate of 61.3%, while essays by US eighth-graders were classified accurately. The paper does not name the seven detectors in its text, and detectors have been retrained many times since, so it does not tell you how GPTZero performs today. What it shows is why plain, predictable writing is at higher risk with any detector.
Our guide to AI detector false positives explains the causes, and how to prove you didn’t use AI is a checklist if it has happened to you. If you came here looking for ZeroGPT, a different product with a similar name, see ZeroGPT vs Turnitin.
GPTZero vs Turnitin
The two are built for different situations. GPTZero is a tool anyone can open in a browser. Turnitin’s AI detection is sold to institutions and shown to instructors. The table compares what each company states, from GPTZero’s site and Turnitin’s AI detection FAQ and file requirements.
| GPTZero | Turnitin | |
|---|---|---|
| Who can run it | Anyone; no account needed for scans up to 10,000 characters | Institutions that license Turnitin Originality; students cannot see the AI report |
| Minimum length | No minimum stated; accuracy improves with more text | At least 300 words of prose |
| Result | Human, mixed or AI, with a confidence level and highlighted sentences | Percentage of qualifying prose likely AI; 1% to 19% shown only as an asterisk |
| Languages | English and more than 20 others | The FAQ lists English, Spanish, Japanese, Arabic, Dutch, German and Turkish; the file requirements page lists the first four |
| Paraphrased AI text | Separate feature, in beta | Checked automatically, English only |
| Stated false positive rate | About 1% of human texts, on the RAID benchmark | Under 1% for documents with more than 20% AI writing |
| Use of the result | “Should not be used to punish or as the final verdict” | “Should not be used as the sole basis for action” |
The two false positive figures look alike but come from different tests on different texts, so they do not show that one tool is more accurate than the other. GPTZero’s help article on Turnitin claims it is better at avoiding false positives, without publishing data for a direct comparison. The practical point is simpler: a GPTZero result tells you nothing reliable about what Turnitin will show, and the reverse is also true. If your school uses Turnitin, our guide to whether Turnitin detects AI covers how its report works.
Using a GPTZero result fairly
GPTZero’s own advice to educators is a good standard for anyone reading a result:
- Look at the process, not just the score. Drafts, revision history and notes show how a text came together. GPTZero suggests asking for them.
- Check understanding directly. A conversation or an in-class task about the work says more than a percentage.
- Look for a pattern. GPTZero recommends looking for a long-term pattern of AI use rather than acting on a single instance.
- Follow the policy. What counts as acceptable AI use, and what happens next, is set by the course and the institution, not by any detector.
If you are a student checking your own draft, the honest use of a detector is to find passages worth rereading, not to tune a score. If your course allows AI, disclose and cite it; see how to cite ChatGPT. The Plagiarism Checker Plus AI detector is another option for a second opinion: free without an account up to 1,000 words, a free account includes 3,000 words a month, the text needs at least 50 words, and it is never used to train AI models. It is a separate detector and can disagree with GPTZero or Turnitin, which is itself a reminder that none of them is a verdict.