GPTZero's Published False-Positive Figures Run From 0.08 to 1 Percent, Each on Different Writing

Four figures for wrongly flagged human writing, three on the company's own pages and one reported by Nature. GPTZero's 2025 comparison page reports 99.6 percent accuracy, a 0.13 percent false-positive rate and 99.4 percent recall ². Its FAQ reports 99.1 percent of human articles classified as human, but only within its high-confidence band ¹. On the home page the headline rounds to 99 percent accuracy ³.

Those numbers do not disagree about arithmetic. They disagree about scope. A confidence band is a filter, so the FAQ's 99.1 percent describes only the documents the model was already sure about, and says nothing about the ones it marked uncertain ¹. The comparison page's 0.13 percent is a single figure over a whole benchmark set.

That page exists to score a rival. It puts Pangram at 97.5 percent accuracy and a 0.20 percent false-positive rate against its own 99.6 and 0.13 ². A figure one vendor publishes about another is worth reading, and worth reading as marketing. The fuller version is pangram-vs-gptzero.

A third figure sits on the home page and is easy to walk past. The page says that on RAID, a third-party benchmark, the detector caught 95.7 percent of AI texts while wrongly predicting 1 percent of human texts as AI ³. That sentence is GPTZero's account of its own RAID result, not RAID's finding about GPTZero. The RAID paper fixes every detector's threshold at a 5 percent false-positive rate, and it publishes neither number ⁴.

Nature supplies a fourth, off the company's pages altogether: in August 2026 it reported that GPTZero had marketed a 0.08 percent false-positive rate from internal tests in February 2026, while its technical paper states the same result more cautiously as sub 1 percent ⁶.

The vendor's own figures for wrongly flagged human writing run from 0.08 percent to 1 percent. The highest of them sits on the home page.

Each was measured on different writing, and only two of the six by anyone other than GPTZero. The years differ too. RAID tests how well detectors hold up against adversarial attacks and against generators they have not seen ⁴. Bloomberg Businessweek tested 500 real college application essays ⁵. A benchmark corpus and a marking pile are not the same population.

The Bloomberg test answers the obvious objection, that the essays might themselves have been machine-written.

The essays went to Texas A and M University in the summer of 2022, shortly before ChatGPT was released, which as Bloomberg puts it all but guarantees they were not machine-written ⁵. They arrived by public records request, which puts them outside the data detectors train on ⁵. GPTZero and Copyleaks flagged 1 to 2 percent of them as likely AI, in some cases claiming near-total certainty ⁵.

RAID found a different weakness. Across eight open and four closed detectors, it reported current systems easily fooled by adversarial attacks and by generators they had never seen ⁴. Variations in sampling strategy and repetition penalty fooled them too.

A figure measured on plain model output says little about a paper that went through a paraphraser once.

Nature named the structural problem in August 2026: the only accuracy rates current enough to describe today's writing models are the ones firms announce from internal testing, and those are not externally verified ⁶. The same report is where GPTZero's February 2026 figure of 0.08 percent comes from, next to the technical paper's more cautious sub 1 percent ⁶. That holds for every vendor here.

Read the date off every one of these figures.

Those who write plainly, in the one study that isolated the mechanism. A 2023 Stanford study published in Patterns ran seven detectors over 91 TOEFL essays written by people and recorded an average false-positive rate of 61.22 percent, against near-perfect accuracy on essays by US eighth graders ⁷.

That study tested detectors as a class rather than GPTZero alone.

Plain writing is not borrowed writing.

The follow-up experiment settles what the flag tracks. Rewriting those same human TOEFL essays with richer vocabulary cut the average false-positive rate from 61.22 percent to 11.77 percent ⁷. Simplifying the vocabulary of the US eighth-grade essays pushed misclassification the other way, from 5.19 percent to 56.65 percent ⁷.

Authorship never changed in either direction; word choice did.

GPTZero names the exposure on its home page, saying the model is repeatedly trained to bring the false-positive rate for ESL writing down to 1 percent ³. A target on a marketing page is an intention rather than a measurement.

A rate stated for ESL writing is not a rate for a course. Even met, one percent of a large course's human-written papers is several papers a term that somebody has to read again.

Bloomberg Businessweek reported the same exposure from the other end: the students most likely to be wrongly accused are those who write in a more generic manner, among them neurodivergent students and students who speak English as a second language ⁵. What that looks like from the student's side is non-native-speaker-flagged-ai-detector.

More than the sales page suggests, because the denominator is submissions rather than students. Vanderbilt University ran this arithmetic in August 2023. It took Turnitin's stated 1 percent rate against the 75,000 papers submitted in 2022, got roughly 750 that could have been wrongly labeled, and gave that arithmetic as one of its reasons for switching the detector off ⁸.

Take 800 submissions, one large course collecting four assignments from 200 students. Assume, and it is only an assumption, that a rate measured on a vendor's benchmark holds on a course's own papers, and that all 800 submissions were written without AI. GPTZero's own 0.13 percent then puts about one wrongly flagged paper in that term ²; the 1 percent its home page reports for its own RAID result puts eight ³. Both are counts of papers, not of students.

That spread is the whole procurement decision.

At the lower figure a department argues over a handful of cases a year. The higher one means a small disciplinary process nobody planned for.

Neither figure predicts which eight papers. A false-positive rate is a property of a corpus, not a ranking of the papers in one class, and it will not fall evenly across the students who wrote them. Working the multiplication out for a specific course is ai-detector-false-positive-rate-arithmetic.

GPTZero says so itself. Its FAQ states that the results should not be used to punish students, and recommends a holistic assessment of student work instead ¹. The same page notes that document-level accuracy runs higher than paragraph-level, which runs higher than sentence-level ¹.

That makes the highlighted sentences the part of the report with the least behind it.

One case shows how far a score gets on its own. A UC Davis senior was referred for academic dishonesty in 2023 after a professor ran his essay through an AI detector that the reporting does not name. He cleared himself with the edit history of the Google document, which showed the paper being built over two or three hours ⁹. The same reporting records the detector flagging Martin Luther King's I Have A Dream speech and passages from Genesis ⁹.

The score opened that case; the record closed it.

Asking for the draft history costs one email and settles more than a second detector run will. The practical version of that request is google-docs-version-history-ai.

A detector that publishes a number owes the corpus and the date beside it, and that test applies as much to Human, at human.olive.is, as to the four GPTZero figures above. Built for college-level academic writing, it reports an in-house measurement on held-out sets rather than an independent audit, measured September 2026: 0 wrongly flagged by Human's default model of 600 human-written papers it had never been trained on; the ceiling on that is 0.498%. This is an estimate from our detector. A flag should be reason to look closer and not verified proof of cheating. Whoever published the figure, the paper in the marking pile still has to be read.