Is GPTZero trustworthy?

For long, unedited output from a single AI model, its record is strong: a 2025 study of 78 essays caught all 28 AI-written ones with confidence never below 98%. For judging whether a human wrote something, it is far weaker, and its own site says no AI detector is 100% accurate.

Can GPTZero be wrong?

Yes, in both directions. The same 2025 study found it wrongly flagged 8 of 50 human essays, a 16% false positive rate. A June 2026 four-detector study found the opposite failure: it returned nothing on 40 papers written entirely by GPT-4o.

Is GPTZero as accurate as Turnitin?

They fail differently. In the December 2023 Weber-Wulff evaluation of 14 tools, Turnitin had the lowest false positive rate (0%) and GPTZero the highest (50%), while GPTZero had the lowest false negative rate (8%). In the June 2026 study both scored 0% strict accuracy on fully AI-written papers.

Why does GPTZero say I used AI when I didn't?

Its classifier scores how predictable your wording is, and careful, plain, tightly structured prose is predictable. The effect is strongest for non-native English writers: across seven detectors, Liang et al. measured a 61.22% average false positive rate on TOEFL essays.

Is GPTZero accurate on short text?

Less so. Every classifier of this kind needs length to accumulate evidence, and the authors of the 2025 study noted that even medium-length texts could not be predicted accurately every time.

Is GPTZero Accurate? What Its Numbers Mean

Two groups of researchers have attempted to determine how often GPTZero identifies human-written essays as AI-generated. One found that the rate was 0.05% while the other found that the rate was 50%. These measurements were made with the same tool two years apart, and both figures are correct.

That is the real answer to "is GPTZero accurate" and it is more useful than a single percentage. GPTZero is very good at one specific job and unreliable at a different one and the score on your screen does not tell you which job it just did.

Here is what the tool claims, what other people measured and which of those numbers applies to your document.

What GPTZero publishes about itself

From GPTZero's own pages, read on 15 September 2026:

Claim	Figure	Where
False positive rate	"no more than 1% when evaluating AI versus human text"	Technology page
Mixed human and AI documents	96.5% accuracy	Technology page
False positives on TOEFL essays	reduced to 1.1%	Technology page
RAID benchmark	95.7% of AI texts detected at a 1% false positive rate	Accuracy page
RAID, excluding retired models	"over 99% when filtering out discontinued LLMs like GPT3.5"	Accuracy page
Chicago Booth benchmark, Aug 2025	99.5% accuracy, 0.05% false positive rate, 99.3% recall	GPTZero's write-up

The company deserves credit for the sentence that follows those figures: "No AI detector is 100% accurate and AI itself is changing constantly." This statement is more honest than those of most of the company's competitors.

The Chicago Booth figures deserve a note. The study was run by researchers at the University of Chicago Booth School of Business, which makes it independent but the write-up I read is hosted by GPTZero and GPTZero comes first in it. That is not a reason to discount the numbers. It is a reason to read them next to the ones nobody in the industry chose to publicise.

What GPTZero stopped doing in 2023

Most explanations of GPTZero describe it as measuring perplexity and burstiness. However, the GPTZero site itself states otherwise:

As of autumn 2023, GPTZero no longer uses perplexity and burstiness for its AI detection because we migrated to a deep-learning based architecture.

Perplexity remains as one of seven indicators within the current model, along with deep learning and internet text search. The current implementation of the model uses a sentence-by-sentence classifier to determine the likelihood of a passage being machine-written, as well as a component called Paraphraser Shield by GPTZero.

This matters when interpreting a result. The perplexity score is something that can be reasoned about. For example, GPTZero's own article states that "generally, a perplexity above 85 is more likely than not from a human source". Deep learning classifiers return percentages, without any indication of how they arrived at those percentages. The advice given in older blog posts on how to increase the perplexity score to beat GPTZero refers to a system that was retired three years ago. This article examines what has replaced it.

What independent tests found

Four studies have measured GPTZero directly.

Study	What was tested	GPTZero result
Weber-Wulff et al., Dec 2023	14 tools, 54 test cases	50% false positive rate, the worst in the field; 8% false negative rate, the best in the field
Liang et al., Apr 2023	7 detectors, 91 TOEFL essays by non-native English writers	field average false positive rate 61.22%; 18 of 91 essays flagged by all seven detectors at once
Dik, Erdem & Dik, Jun 2025	78 essays: 28 AI-generated, 50 human-written	100% of AI essays caught, never below 98% confidence; 84% accuracy on human essays, a 16% false positive rate
Van Vlasselaer et al., Jun 2026	160 documents, four detectors	~97.5% on fully human text; 0% strict and 30% inclusive on fully AI papers; 0% on hybrid text; 2.5% on text re-prompted to sound human

Read the Weber-Wulff row again. The same tool had the highest false positive rate and the lowest false negative rate of fourteen. These are not two flaws in the tool. This is one behaviour in 2023: GPTZero leans hard towards answering "AI". This catches nearly all machine text and sweeps up human text as well.

Read the 2026 row, where the lean has gone the other way. On 40 papers written end to end by GPT-4o, GPTZero returned nothing under strict scoring. In contrast, on the same study's human documents, it was near perfect. So, three years apart, the tool has moved from over-calling AI to under-calling it against current models.

What that means for your document

The Dik study provides the clearest picture of the shape of GPTZero's abilities because the two tasks were separated. GPTZero correctly identified every instance of a machine-written essay that had not been touched by humans, with confidence levels that never dropped below 98%. However, it incorrectly identified eight instances of human-written essays out of fifty.

So the honest reading of this is: a high score on long, unedited text that is output by a single model is usually right. A high score on a human essay is something that happens roughly once in every six attempts according to the only study that measured this kind of score, and significantly more often than that if the writer is not a native speaker of English: Liang and colleagues found the field average to be 61.22% for essays written by non-native speakers for the TOEFL test. GPTZero claims to have lowered the false positive rate on these essays to 1.1%, a real improvement, measured on its own test set rather than an independent one. Our page on detector bias against ESL writers goes into this gap in more detail.

The direction of the error is what nobody tells students. A detector that over-calls AI hurts innocent writers. A detector that under-calls it hurts nobody except the institution paying for it. Between 2023 and 2026 GPTZero appears to have traded one for the other, which is defensible as product design and makes the word "accurate" almost meaningless without a date attached.

When a GPTZero result is worth something

Long, unedited output from one model. This is the case it was built for and the case it wins: 100% in the Dik study, 95.7% on RAID.
Short passages. Weak. Every classifier in this category needs length to accumulate evidence, and the Dik authors noted that medium-length texts "cannot be accurately predicted every time".
Writing by non-native English speakers. Treat with real suspicion, whichever figure you believe.
Mixed or edited text. GPTZero claims 96.5% here. The 2026 study measured 0% on hybrid documents. That is the widest gap between claim and independent result anywhere in this post.
As the only evidence in a misconduct case. No. Both peer-reviewed studies above say so in their own conclusions, and GPTZero's own site says no detector is 100% accurate.

If you want to know how your writing would score before anyone else runs it through, use a checker with no stake in the answer: Sapling's free detector needs no account, and on our own fixed samples it has separated 1918 public-domain prose from raw AI cleanly on every date we have tested it. If you are working from an AI-assisted draft, the description of what a rewriting pass will do to your GPTZero score, and what it will not do, is on our GPTZero page.

The short version

GPTZero has reliably caught unedited AI text since 2023. The problem is its record with human text. A study done in 2025 found that GPTZero produced false positives in 16% of student essays. A study done in 2023 found false positives in 50% of student essays. Also, the field to which GPTZero belongs has averaged false positives of 61.22% in essays written by non-native English speakers. GPTZero, however, claims a false positive rate of 1% or lower on its own benchmarks. Yet a study done in June 2026 found that GPTZero returned nothing on papers written completely by GPT-4o. Any number you have seen quoted for GPTZero is true only of one test set at one date, and the score on your screen gives no indication of which one.

If you want the longer treatment

Are AI detectors accurate? Every published number, in one table — the same treatment applied to the whole field.
How AI detectors work: perplexity, classifiers and watermarks — what replaced perplexity and burstiness.
AI detector comparison: Turnitin vs GPTZero vs Grammarly — the three tools students meet most.
Why an AI detector says your essay is AI when it isn't — what to do about a false positive.
What AI detector do colleges actually use? — whether your institution even has this switched on.

Is GPTZero Accurate? What Its 2026 Numbers Actually Mean