GPTZero is accurate in some conditions and unreliable in others, and the numbers you see depend on who measured them. GPTZero reports that it catches 95.7% of AI texts while mislabeling 1% of human texts on a public benchmark. Independent research is thinner, smaller and more cautious: it finds GPTZero good at spotting plain AI output and less dependable at confirming that a person wrote something.
Key takeaways
- The best-known GPTZero figures (95.7% detected, 1% false positives) come from GPTZero’s own write-up of a benchmark run. They describe fully human or fully AI text, not edited or mixed writing.
- The independent GPTZero study we found was small: 28 AI essays and 50 human essays. It found AI text detected well and a handful of human essays wrongly flagged.
- GPTZero’s percentage is a probability about the whole document. It is not the share of the text that is AI.
- GPTZero itself says a result should not be the only proof for academic discipline.
What does GPTZero say about its own accuracy?
In a post dated 15 October 2025, GPTZero says its detector identified 95.7% of AI texts while classifying 1% of human texts as AI, and that accuracy rises above 99% when discontinued models such as GPT-3.5 are removed. It calls this result “independent, 3rd party” because the test set is RAID, a benchmark built by university researchers and presented at ACL 2024.
Several conditions sit behind those numbers:
- The vendor ran and reported the comparison. The RAID authors released the data and a leaderboard, but the 95.7% and 1% figures, and the ranking against other tools, are presented in GPTZero’s post. The post presents one table without adversarial attacks and one with them, but the sentence giving 95.7% and 1% does not say which setting it comes from, and the tables are images we could not read as text. The no-attack table is the likelier source, which is our inference and not GPTZero’s statement.
- RAID only has fully human or fully AI texts. GPTZero says it also handles mixed writing and AI text that has been rewritten, but the post gives no benchmark figures for those cases.
- The false positive rate is tied to a chosen threshold. GPTZero says the metric it cares about most is the detection rate when false positives are held to 1%. That is a sensible way to test, but it is a setting on a benchmark, not a promise about every document you will ever check.
The RAID paper itself is a reminder of why conditions matter. Its authors tested 12 detectors and concluded that current detectors are easily fooled by adversarial attacks, and that performance also falls with different sampling settings and with generators the detector has not seen. A high score on one run says little about a different kind of text or a newer model.
GPTZero’s news page also lists newer posts we did not read, including a GPTZero vs Pangram comparison dated 10 November 2026 and an “Introducing GPTZero 4o” post dated 18 October 2026. The 95.7% and 1% figures may therefore have been superseded by newer vendor numbers, which we have not verified.
What have independent studies found?
Independent evidence on GPTZero specifically is limited, and we found no large peer-reviewed evaluation of its false positive rate.
The study we found is a 2025 arXiv paper by Dik, Erdem and Dik, Assessing GPTZero’s Accuracy in Identifying AI vs. Human-Written Essays. It tested short (40 to 100 words), medium (100 to 350) and long (350 to 800) essays: 28 AI-generated and 50 human-written. Most AI essays were detected, with 91 to 100% of their content attributed to AI. Human essays gave more varied results, including a few false positives. The authors conclude that GPTZero is effective for purely AI-generated content but less reliable for confirming human authorship, and advise educators not to rely on it alone. With 78 essays, treat this as a pointer, not a measurement.
Two broader findings frame everything. A 2023 study by Weber-Wulff and colleagues, Testing of Detection Tools for AI-Generated Text, tested 12 public tools plus Turnitin and PlagiarismCheck and called them “neither accurate nor reliable”, with a lean toward labeling text as human. We could not confirm from the abstract whether GPTZero was among them, so we cite it as field-wide evidence only. And GPT detectors are biased against non-native English writers (Liang and colleagues) found that the detectors it evaluated consistently misclassified non-native English writing as AI-generated. The abstract does not name the detectors, so we cannot say how GPTZero would fare. Both are reasons to be careful with any tool, and our guide to how accurate AI detectors are covers the wider evidence.
Be wary of the many review sites that publish a tidy “GPTZero is 87% accurate” figure. Most we saw were small tests or repeated GPTZero’s own claims, and they disagreed with each other.
How do I read GPTZero’s probabilities and highlights?
GPTZero’s help pages explain the output, and the details are easy to misread.
The percentage is a probability about the document, not a share of the text. GPTZero’s quick scan sorts text into AI, mixed or human, and each result carries a probability and a confidence statement. Its documentation says the number is “NOT saying that ‘6% of your document contains AI’”. Its own example reads a “4% AI” score as roughly 4 in 100 similar texts being fully AI, about 57 mixed and about 39 fully human. If that sounds unfamiliar, it is a different idea from the “share flagged” figure on many tools; what an AI detection percentage means explains the two readings.
Confidence is about error rate. GPTZero describes “highly confident” as an error rate under 2%, “moderately confident” as around 10% and “low confidence” as 14% or higher. It says that when it is unsure it leans toward “human” with low confidence, to avoid accusing someone falsely. A low-confidence “human” is therefore not strong evidence of anything.
Highlights show influence, not guilt. In the advanced scan, highlighted sentences are the ones with the greatest effect on the overall score. GPTZero says unhighlighted sentences can still affect the score, and because the detector reads the whole document, the result may depend on a pattern across most of the text rather than a few lines. A highlight is a place to start reading, not a finding about that sentence.
| What you see | What it means | What it does not mean |
|---|---|---|
| “X% AI” | Estimated likelihood about the whole document | X% of the sentences were written by AI |
| Confidence level | The error rate for documents like this one | A guarantee that this result is right |
| Highlighted passage | Strongest influence on the score | Proof that this passage was machine-made |
| “Human” with low confidence | The model leaned human because it was unsure | Strong evidence of human authorship |
Does GPTZero get false positives?
Any detector does, and GPTZero’s own materials acknowledge it. Its FAQ says results “should not be used to punish students”, that sentence-level classification should not be used alone to indicate that an essay contains AI, and that accuracy improves as more text is submitted. Its help pages say it does not recommend using a result as the only proof for academic discipline. The FAQ also says the classifier is not trained to identify AI text that has been heavily modified, which matches the point above about edited writing.
The situations where care matters most are the same as for the whole field: short passages, formal or formulaic writing, translated text, and writers working in a second language. Why your writing may show as AI and AI detection false positives cover those cases. If a GPTZero result is being used against someone, what to do when a detector flags student writing sets out a fairer process.
What about price and free limits?
GPTZero offers a free tier and paid plans, but we could not verify the free plan’s monthly limit. Its pricing page loads most plan details client-side. On 15 November 2026 the text we could read listed only a team plan at 500,000 words per member per month with a two-seat minimum, and its price fields showed placeholder values. The FAQ says a free account can scan texts up to 150,000 characters, or a batch of up to 250 files. Third-party sites quote different numbers, so check the GPTZero pricing page directly before relying on any figure, including ours.
Why do GPTZero vs ZeroGPT or Turnitin comparisons mislead?
Head-to-head posts are tempting because they end in a winner, but the winner often depends on the test.
- Different thresholds. A tool tuned to flag cautiously looks better on false positives and worse on missed AI. Comparing detection rates without the matching false positive rate is not a comparison.
- Different samples. A handful of essays, one genre or one model can put any tool on top. A 40-sample test and a benchmark of hundreds of thousands of texts are not the same kind of evidence.
- Vendor authorship. Many comparisons are written by a company that sells one of the tools being compared, including GPTZero’s own benchmark posts. That does not make them false, but it is a reason to look for the underlying data.
- Moving targets. Detectors are updated and new language models appear, so a result can age quickly.
A fair comparison states the dataset, the text lengths, the models used, the false positive rate and the date. If a page does not give those, treat its ranking as an opinion. Our comparison of detectors teachers use sticks to what the vendors document themselves.
How should I use a GPTZero result?
Use it as one signal. Check the length first, because very short text gives thin evidence. Read the highlighted passages and ask whether they are generic, formal or simply well-polished. Consider who wrote the text and whether they use editing tools openly. Then look for stronger evidence, such as drafts, version history, sources and a conversation with the writer. How to tell if an essay was written by AI lists checks that carry more weight than any score.
If you already have the text, run it through our AI-generated text detector and review the flagged passages rather than relying on the overall score. We label passages and do not publish an accuracy number, for the reasons on the accuracy page, so a second tool is a second signal, not a tiebreaker.
Frequently asked questions
Is GPTZero accurate enough to use on its own?
No. GPTZero’s own documentation says a result should not be the sole basis for academic discipline. It is most useful as a prompt to look more closely, alongside drafts, sources and a conversation with the writer.
What is GPTZero’s false positive rate?
GPTZero reports 1% on the RAID benchmark, which contains fully human and fully AI texts. That figure comes from GPTZero’s own post and does not cover edited, translated or non-native writing. Real-world rates depend on the text and threshold, and we found no large independent measurement.
Does GPTZero’s percentage show how much of the text is AI?
No. GPTZero says the percentage is the model’s estimated likelihood about the whole document, and that “6% AI” does not mean 6% of the document contains AI.
Is GPTZero free?
It has a free tier, but we could not verify the current limits from GPTZero’s pages. Check its pricing page for the present terms.