CiteOwl

Blog/Using AI honestly

How accurate is GPTZero?

The CiteOwl Team· Updated 29 August 2026· 13 min read

On the largest public benchmark for AI text detection, GPTZero's own October 2025 submission caught 95.7% of machine-written documents while wrongly flagging 1% of human ones. The scored file is public, so the figure is checkable rather than advertised. It still does not answer the question you arrived with, which is whether the essay that just got flagged was written by a machine. That depends on something no benchmark measures: how much AI writing is in the pile your essay was scanned alongside.

Four of our pages cover Turnitin. This one covers the free detector students paste into at midnight.

Three different questions, one word

"Accuracy" is doing three separate jobs here, and none of them follows from the others.

The first is a vendor's self-reported accuracy: a number measured on a corpus the company assembled, at a threshold it chose. The second is an independent benchmark result: the same detector measured by somebody else, on a public dataset, at a stated false positive rate. The third is the one you want, the chance that this flagged essay was written by AI.

The third cannot be read off the first two, because it depends on a quantity outside the detector: how common AI writing is in the batch being scanned. A detector that is right 99% of the time where half the documents are machine written behaves very differently in a seminar where one paper in fifty is. Section six does that arithmetic; everything before it is the evidence going in.

This page will not tell you how to make writing pass a detector. That is a different problem with a worse answer, covered in do AI humanizers actually work.

What GPTZero says about GPTZero

The headline on gptzero.me is "99% Accuracy". The FAQ underneath expands it: "Independent benchmarks, like our partners at Penn State's AI Research Lab, and our own large-scale testing show that GPTZero is the most accurate AI detector, with a 99% accuracy rate when spotting AI-generated text vs. human writing." Three more precise figures sit behind it, all read on 29 August 2026.

The API documentation states a condition: "Confidence categories are tuned such that when the confidence_category field is high 99.1% of human articles are classified as human, and 98.4% of AI articles are classified as AI." Those rates describe the scans the model itself calls high confidence, not every scan.

A January 2024 engineering post adds the caveat most coverage drops. The bands are tuned so that "the average error rate is less than 1% for the 'high' confidence predictions". Average is load bearing, because "the number of possible documents is vast" and some kinds sit above that rate.

In February 2026 GPTZero published a methodology paper, which is more than most detectors offer. GPTZero: Robust Detection of LLM-Generated Texts reports, at a threshold fixed to 1% false positives, 99.8% accuracy on essays and 99.4% on scientific abstracts, and claims "GPTZero is the only detector which consistently has a sub 1% false positive rate across all domains while achieving a recall >97%."

It is still the vendor evaluating itself, on a closed model with proprietary training data, running the competitors itself. We tried to reproduce the essay figure and could not. Section 5.1 describes the benchmarks as "1000 human and 1000 corresponding AI texts, with the AI texts being generated by the following 4 LLMs (250 texts per LLM): GPT-5.2, Gemini 3 Pro, Claude Sonnet 4.5, and Grok 4 Fast". Appendix C.2 describes those sets differently, calling the essay set "a balanced dataset of 200 texts" and the abstract set a "balanced subset of 500 scientific abstracts generated by Mistral". The appendix links a public repository, which we downloaded on 29 August 2026 and counted.

The README describes the repository as data for a different, earlier conference submission, and its AI text comes from GPT-4, Mistral and Davinci, not the four 2025 models the paper names. The dull explanation is likely right: an appendix carried over from an older draft. The consequence is less dull. Essay accuracy, the figure a student cares about most, cannot be recomputed from the files the paper points at, and the essay file is not balanced: three human documents for every AI one.

What the public benchmark actually recorded

RAID is the largest shared benchmark for this task: over six million generations across eleven generators, eight domains, eleven adversarial edits and four decoding strategies, published at ACL 2024 by Dugan and colleagues. It matters structurally, because it fixes the false positive rate before measuring anything, so every detector is scored at the same tolerance for wrongly accusing somebody.

Its finding about the field is one sentence, and this page is built on it: "We found that our detectors were capable of achieving the high accuracies cited in many viral reports, but only at similarly high FPR."

GPTZero appears in RAID twice, twenty-one months apart, and the pair is the clearest demonstration available of why a detector number is easy to abuse.

The paper's authors queried GPTZero's API on 24 January 2024. At a threshold tuned to 5% false positives, that version scored 66.5% overall, and underneath the average it was wildly uneven: 99.4% on ChatGPT output and 97.1% on GPT-4 output, against 28.9% on base Mistral and 24.8% on base MPT. An excellent ChatGPT detector, a poor detector of anything else.

In October 2025 GPTZero submitted its current model's predictions to the same benchmark, where the maintainers score them and publish the file. We downloaded it and computed the breakdowns.

Every row is a true statement about GPTZero's accuracy on RAID. They differ by date, by threshold and by which slice is counted, which is why a figure quoted without those three facts tells you nothing. The leaderboard is public, and so is the per-domain detail: at a 1% false positive rate it caught 96.2% of AI-written abstracts, so academic prose sits near the top of its range.

The adversarial breakdown is the other thing worth extracting. Ten of the benchmark's eleven edits cost GPTZero at most one point. The eleventh, machine paraphrasing, took it from 95.7% to 71.6%. That is a fact about how a headline conceals a twenty-four point spread, not a route anywhere: it still leaves seven in ten caught.

What an independent audit found

The most useful third-party evidence is not a leaderboard, it is an audit. Artificial Writing and Automated Detection, NBER working paper 34223, by Brian Jabarian and Alex Imas of Chicago Booth, September 2025. They matched 1,992 human-written passages across six everyday genres with AI versions from GPT-4.1, Claude Opus 4, Claude Sonnet 4 and Gemini 2.0 Flash, then scored four detectors.

Its GPTZero numbers are the closest thing to a usable false positive rate for a normal document. Across the whole threshold range the authors imposed it held flat at 0.0071, roughly seven human passages in a thousand, and ran from 0.0000 on résumés to 0.0238 on Amazon reviews at per-genre thresholds. The miss rate on ordinary AI output ran from 0.002 to 0.030 by model. The paper ranks Pangram first and puts GPTZero and Originality.AI in a second tier where "minimizing FPR favors GPTZero, while maximizing AUROC favors OriginalityAI".

Then there is rewritten text, where the evidence conflicts.

We cannot resolve that. The three sit months apart, use different rewriting tools, and one is run by the party with an interest in the answer. What can be said is what has not been done: nobody independent has re-run a humanizer test on the model GPTZero shipped in 2026. Either of the first two rows, quoted as the current state of the world, is an unreplicated result.

The false positives, and who gets them

The number that does damage is the false positive rate, and its most cited measurement is three years old.

Liang, Yuksekgonul, Mao, Wu and Zou ran 91 TOEFL essays by non-native English speakers and 88 by US eighth-graders through seven detectors, GPTZero among them, on 15 March 2023. The version of record, in Patterns, reports that while the detectors classified the US student essays correctly, "they incorrectly labeled more than half of the TOEFL essays as 'AI-generated' (average false-positive rate: 61.3%). All detectors unanimously identified 19.8% of the human-written TOEFL essays as AI authored, and at least one detector flagged 97.8% of TOEFL essays as AI generated."

The mechanism carries forward, because it has not changed. Perplexity-based detectors read a narrow vocabulary as machine writing, and a narrow vocabulary is what second-language writing looks like on the page. The authors' verdict on vendor claims: "Claims of GPT detectors' '99% accuracy' are often taken at face value by a broader audience, which is misleading at best, given the lack of access to a publicly available test dataset, information on model specifics, and details on training data." Their recommendation: "we strongly caution against the use of GPT detectors in evaluative or educational settings".

Now the qualification most pages citing that study leave out: it measured a generation of detectors that no longer exists. GPTZero today says it is "the only AI detector de-biased for ESL (English Second Language) learners" and retrains "to reduce false positive rate for ESL writing to 1%". That is a vendor claim with no published test behind it, and nobody has re-run the TOEFL experiment on current models. So 61.3% describes March 2023 detectors, and no replication exists in either direction. Our page on whether AI detectors will flag your writing covers the statistical features involved.

Why a 99% detector still flags honest essays

Take the best documented operating point here: 95.7% of AI documents detected, 1% of human ones wrongly flagged. Picture a thousand essays. If fifty were written with AI, the detector flags about forty-eight of them and about ten of the nine hundred and fifty honest ones, so forty-eight of fifty-eight flags are right. If instead ten were AI, it flags about ten of those and still about ten honest essays, and half of every flag is wrong. Same detector, same threshold. Only the pile changed.

The middle column uses GPTZero's own submission: 95.7% detected at 1% false positives. The right-hand column pairs that submission's 96.2% on abstracts with the 10% domain-level false positive rate RAID recorded on abstracts. Both pairs come from the sources named above; the arithmetic is ours and is nothing more than counting.

The RAID authors put it best in explaining their own choice of metric. They report accuracy at a fixed false positive rate because the question is "What percentage of generations are detected given that we tolerate an x% chance of wrongly accusing someone".

And here is the part nobody knows. There is no measurement of how much AI writing sits in your course, and there cannot easily be one: the ground truth would mean knowing who used AI, which is what the detector was for. The number that decides what a flag means is the one nobody has. Anyone telling you a 99% accurate detector makes a flagged essay 99% likely to be AI written has skipped this section.

Two caveats, since the table is arithmetic rather than measurement. It assumes the false positive rate on your writing matches the benchmark rate, which GPTZero itself warns is an average. And it treats documents as independent, which they are not when one marker scans a whole class.

What a score is good for

GPTZero's own guidance is more careful than the way its output gets used. The FAQ says results "should not be used to punish students" and calls the classifier a tool "to flag situations in which a conversation can be started (for example, between educators and students)". The engineering post adds that "we do not encourage punitive actions based solely on the results of a single scan". The company selling the detector calls a scan the start of a conversation, which is what the arithmetic supports.

If you have been flagged, the score is evidence of a statistical resemblance and nothing more, and what a Turnitin AI score means transfers to GPTZero unchanged. The route through an accusation is in accused of using AI on an essay, and what a detector does inside a plagiarism platform is in can Turnitin detect ChatGPT.

If you have not been flagged, the productive question is not what a detector will score. It is whether the work is yours, which is about how you used the tool rather than which words came out. Is using AI to write essays cheating is the question a marker is actually asking.

And if you run the scans: the table above is why a single result cannot carry a case. At low prevalence most flags are wrong, even from a detector performing exactly as advertised. That does not make the tool useless. It makes it a shortlist.

None of which settles the problem underneath, that a finished draft carries no record of where anything came from. CiteOwl works the other way round. It searches OpenAlex, CrossRef, Unpaywall and the web, reads the papers it retrieves, writes each claim from a passage it keeps beside the sentence, and hands you every edit as a diff to accept or reject. The result is not a score. It is a document whose sources and decisions were each yours on the way through.