Is GPTZero Accurate? Benchmarks, Studies and False Positives

On long English prose from mainstream chatbots, recent independent tests put GPTZero among the more accurate AI detectors: a 2025 University of Chicago study found it wrongly flagged about 0.7% of human-written passages and missed at most 3% of AI text. It is not "99% accurate" everywhere, though. The same study found it missed about half of AI text after a "humanizer" rewrite, it is weaker on very short text and outside English, and its 2023 version produced many false positives. GPTZero itself says "No AI detector is 100% accurate" and that "results should not be used to punish or as the final verdict" (GPTZero FAQ).

Below: GPTZero's own accuracy claims, labelled as vendor claims; what each independent evaluation measured; how it handles languages; what triggers false positives; and how to read a GPTZero result. Every figure links to its source, and everything was checked in September 2026.

GPTZero's own accuracy claims

GPTZero's homepage advertises "99% Accuracy." Its FAQ says GPTZero is "the most accurate AI detector, with a 99% accuracy rate when spotting AI-generated text vs. human writing," and that it detects "mixed documents" with "a 96.5% accuracy rate" (GPTZero).

The detailed numbers are on its benchmarking page, published 5 February 2026 and described as updated quarterly. Each domain benchmark uses 1,000 human texts and 1,000 AI texts, generated with GPT-5.2, Gemini 3 Pro, Claude Sonnet 4.5 and Grok 4 Fast. For detector version 4.3b, GPTZero reports:

Source: GPTZero benchmarking page.

Credit where it's due: GPTZero names its model versions, publishes some raw predictions, and explains how it builds each dataset, which is more than most detector vendors do. But these remain vendor benchmarks. GPTZero chose the texts, generated the AI samples, and set the thresholds, and some datasets, including the essays, aren't public "due to data usage policies and copyright constraints."

GPTZero also points to RAID, an academic benchmark. In an October 2025 post, it says it detected "95.7% of AI texts while only incorrectly predicting 1% of human texts as AI," rising to "over 99% when filtering out discontinued LLMs like GPT3.5." That figure checks out against RAID's public leaderboard files: GPTZero's submission, released 7 October 2025, scores 95.7% at a 1% false-positive rate on RAID's test set without adversarial attacks, and 92.9% with them (RAID leaderboard results). RAID keeps its test labels hidden and scores the predictions that detector makers submit, so this is a vendor-run submission, independently scored.

What independent evaluations found

The story those rows tell is one of change. The early-2023 GPTZero was the outlier in Weber-Wulff's study: the authors wrote that "half of the positive classifications would be false accusations, which makes this tool unsuitable for the academic environment," and noted that, "with the notable exception of GPT Zero," every tool erred toward calling text human. By 2024 and 2025, the independent tests show a much more cautious detector.

The most useful recent study is the one by Brian Jabarian and Alex Imas at the University of Chicago (NBER working paper 34223, not peer-reviewed). They matched human news, blogs, reviews, novels, restaurant reviews and résumés with AI versions from GPT-4.1, Claude Opus 4, Claude Sonnet 4 and Gemini 2.0 Flash. Their findings on GPTZero:

• False positives stayed low. GPTZero's false-positive rate held at about 0.7% across every threshold tested.
• It caught most unedited AI text. It missed between 0.2% and 3.0%, depending on the model and threshold.
• Very short text was harder. On passages under 50 words, GPTZero had "a substantially higher miss rate."
• Humanizers hurt it badly. After passages went through the StealthGPT humanizer, GPTZero "largely loses its capacity to detect AI-generated text," missing around half or more.
• It wasn't the best in the study. Pangram was the only tool that met the authors' strictest policy cap. They put GPTZero and Originality in "a secondary tier," noting that "minimizing FPR favors GPTZero."

GPTZero's February 2026 benchmark claims 91.8% recall on bypassed text for a newer version, so that humanizer result may not describe today's model. It is, however, the most recent independent test we found.

Why the numbers disagree so much

Four things explain most of the spread between "54% accurate" and "99% accurate."

1. Version. GPTZero says it shipped 15 model releases in 2025 alone (GPTZero). A 2023 result describes a different product.
2. Threshold. RAID and the Chicago study set thresholds themselves to compare tools fairly. In the app, GPTZero chooses where "AI" starts, and the Chicago authors note detector thresholds "are not disclosed to users through the available APIs."
3. Text type. RAID's domains are news, books, abstracts, reviews, Reddit, recipes, Wikipedia and poetry, and its authors saw "a clear bias towards domains and models that the detectors have trained on" (RAID). A benchmark result on news articles says little about a lab report.
4. Who ran it. GPTZero's own benchmarks are the most flattering, independent academic tests the least, and a vendor-submitted run scored by RAID sits in between.

Languages: "strongest on English prose"

GPTZero is candid about this. It "fully supports English, German, Portuguese, French and Spanish," and says accuracy "is strongest on English prose" (GPTZero FAQ). Its benchmarking page adds that more than 85% of the texts users scan are English. "Multilingual AI detection" is listed as a Premium plan feature.

The independent data on other languages is thin and older. On RAID's extra Czech and German news sets, the 2024 version of GPTZero caught 33.6% and 49.5% of AI text at a 5% false-positive rate (RAID, Table 7). GPTZero's own 2026 multilingual benchmark reports far better numbers for a newer version. Until someone else tests that, treat scores outside those five languages, and especially outside English, with extra caution.

False positives: when GPTZero flags human writing

• Famous or heavily copied text. A viral April 2023 screenshot showed GPTZero calling part of the US Constitution "likely to be written entirely by AI." GPTZero's creator Edward Tian told Ars Technica the Constitution "is a text fed repeatedly into the training data of many large language models."
• Short text. GPTZero says "accuracy improves with longer inputs (document-level results are stronger than paragraph or sentence-level)" (GPTZero FAQ). In 2023 its minimum input was 250 characters, about 50 words (Weber-Wulff et al.).
• Machine translation. In the 2023 Weber-Wulff test, six of GPTZero's nine false accusations were on machine-translated human writing (Weber-Wulff et al.).
• Non-native English. GPTZero was one of the seven detectors in Liang et al.. GPTZero now says it is "de-biased for ESL" and trains "to reduce false positive rate for ESL writing to 1%," a vendor claim we couldn't find independently tested on the current version. The wider debate is in AI detection and ESL students.
• Your writing, polished by AI. GPTZero separates "human text that was polished by an LLM" into its own "Mixed" category, and in its benchmarks it counts that as AI (GPTZero). If you ran your own draft through an AI editor, expect a Mixed result rather than Human.

How to read a GPTZero result

GPTZero's guide to understanding scan results (December 2024) explains more than most users realize:

• The label comes first. A Quick Scan returns "AI," "human," or "mixed," and GPTZero's taxonomy also distinguishes paraphrased AI and AI-polished human text.
• The percentage is a probability, not a share of your text. In GPTZero's words, it "is NOT saying that '6% of your document contains AI'" but is "our model's probability that your document was written by either AI or human." That's different from Turnitin, whose percentage is a share of the text; see GPTZero vs Turnitin.
• Read the confidence label. By GPTZero's own figures, "highly confident" means an error rate under 2%, "moderately confident" around 10%, and low confidence 14% or higher. It says that when unsure it leans toward "human" with low confidence.
• Highlights show influence, not guilt. Advanced Scan highlights the sentences that most affect the score, and GPTZero notes that its detector "holistically analyzes the document," so unhighlighted sentences still count.
• It isn't proof. GPTZero says it doesn't "recommend using an AI detection result as the only proof for academic punishment or discipline."

What to do if GPTZero flags your writing

1. Note the label, percentage and confidence. A low-confidence Mixed result is weak evidence even by GPTZero's own standards.
2. Check length and language. Short passages and languages outside GPTZero's five fully supported ones are where errors concentrate.
3. Gather process evidence. Version history, dated drafts, notes and sources. GPTZero's own Writing Report can rebuild a Google Doc's history "edit by edit," and GPTZero calls it "evidence of process, not a verdict" (GPTZero).
4. Test the tool on your older writing. Our ZeroGPT accuracy guide has a step-by-step method for sanity-checking any detector on texts where you know the answer. A second opinion from a different model, such as SupWriter's AI detector, can show whether the flag is specific to GPTZero, though no single score settles anything (how we test).
5. Contact GPTZero if it's a disciplinary matter. Its guide says: "If you are facing disciplinary issues as a result of a false positive (including from other tools), please reach out to our team."
6. Know your rights in the process. Our guide on how to prove your writing is human covers what to bring to a meeting and how to appeal.

FAQ

Is GPTZero really 99% accurate?

That's GPTZero's own claim, based on its own benchmarks (GPTZero). Independent results depend on the version and the text: 54% of documents correct in a March 2023 study (Weber-Wulff et al.), and about 0.7% false positives with at most 3% of unedited AI text missed in a 2025 study (Jabarian and Imas).

Is GPTZero more accurate than ZeroGPT?

On the one benchmark that tested both on identical text, mostly yes. In RAID, on ChatGPT, GPT-4 and Cohere chat output generated with random sampling, GPTZero caught 88.5% of AI text and ZeroGPT 65.8% at a 5% false-positive rate, and at a plain 0.5 cutoff GPTZero flagged 0.00% of human text to ZeroGPT's 1.42% (RAID). Those are 2024 versions. More in is ZeroGPT accurate?

Can GPTZero detect Claude and GPT-5?

GPTZero says it works across models including "GPT-5, GPT-6, Gemini, Claude, Llama, Deepseek" (GPTZero). Independently, the Chicago study found it missed 2.9% of Claude Opus 4 text and 0.85% of Claude Sonnet 4 text at a 0.5 threshold (Jabarian and Imas). We found no independent test on GPT-5 yet.

Does GPTZero work in other languages?

It "fully supports" English, German, Portuguese, French and Spanish, and says it is "strongest on English prose" (GPTZero FAQ). Treat results in other languages with caution.

---

Want a second opinion on a draft? Check it with SupWriter's AI detector. 300 words free, no credit card required.