Guide

GPTZero in 2026: How It Works, How Accurate It Is, and What to Do If It Flags You

A plain-English 2026 guide to GPTZero: how its classifier scores a document, what changed in GPTZero 4o, what its 99% accuracy claim is measured on, what peer-reviewed testing found, what Writing Replay records, what it costs, and the pending Superhuman acquisition.

September 27, 2026 · The AI Rankings

Quick answer: GPTZero is a consumer AI detector that classifies text sentence by sentence and returns a document-level verdict of human, mixed or AI, and it is the most widely used detector of its kind — its own pricing page claims over 10 million teachers and students in more than 100 countries as of 26 September 2026. GPTZero’s published claim is 99% accuracy with a false-positive rate at or below 1%, and its newest model, GPTZero 4o, announced on 24 September 2026, claims one false positive in 10,402 student essays on the PERSUADE corpus (GPTZero). Independent testing on long academic papers found something different: a peer-reviewed study published in the International Journal for Educational Integrity in June 2026 found GPTZero classified all 40 human papers correctly but scored 0% strict accuracy on 40 fully AI-generated academic papers and on 40 hybrid papers, and 2.5% on humanised text (Van Vlasselaer et al.). Three things follow. GPTZero is free to point at your own writing and needs no account, which makes it the most common way a student first meets a detector. Its own FAQ states that “these results should not be used to punish students”. And the company is subject to a pending acquisition by Superhuman, the company formerly named Grammarly, announced on 23 June 2026 with no completion announced as of 26 September 2026.

This guide covers GPTZero specifically: the product, the score, the evidence and the ownership. For the underlying detection mechanism across all tools, see how AI detectors work. For ranked alternatives with independent accuracy scores, see best AI detectors. For the institutional detector most students are actually assessed by, see Turnitin AI detection.

What GPTZero actually is

GPTZero is an AI detection product sold direct to individuals and institutions, rather than a feature inside a learning management system.

It launched as a student project and stayed consumer-first. Edward Tian, then a Princeton undergraduate, put GPTZero online on 3 January 2023, and it drew 30,000 uses in its first week — enough to crash the hosting it was built on. The company raised $3.5 million in seed funding in May 2023 and $10 million in a Series A in mid-2024.

Anyone can self-check, which is the opposite of Turnitin. A student can paste their own essay into GPTZero before submitting it. They cannot do that with Turnitin, whose AI writing indicator is visible only to instructors and administrators.

It is now a suite, not a single checker. GPTZero’s own navigation as of 26 September 2026 lists an AI Detector, a Hallucination Detector, a Plagiarism Check, Authorship and Writing Replay, Expert Feedback, a Grammar Checker, a Chrome extension and AI Vision for images.

It has institutional distribution as well as consumer reach. GPTZero’s pricing page describes the company as the “Official AI detector partners of the American Federation of Teachers”, and its newsroom lists a post dated 8 July 2026 describing AP Verify integrating GPTZero for generative AI text detection in its verification workflow. Those are partnership claims published by GPTZero, not independently audited deployments.

GPTZero’s user figures differ between pages. GPTZero’s pricing page says “over 10 million teachers and students”, while the site metadata on its Chrome extension page says “over 17 Million Users” (GPTZero Chrome). The two figures may be counting different things — registered users against total users, or people against extension installs — but GPTZero does not reconcile them.

How GPTZero’s detector works

GPTZero runs a sentence-by-sentence classification model that assigns each sentence a probability and confidence that it was created by AI, then aggregates those into a document-level verdict (GPTZero technology).

GPTZero’s technology page describes “an end-to-end deep learning approach, trained on text datasets from the web, education, and AI-generated from a range of LLMs”. It does not describe the product as running perplexity and burstiness calculations, which is how GPTZero was commonly explained in 2023. Perplexity and burstiness are covered in how AI detectors work; a 2026 GPTZero score is the output of a trained classifier.

The API exposes the verdict as one of three classifications — HUMAN_ONLY, MIXED or AI_ONLY — alongside sentence-level and paragraph-level detail. The consumer report shows the same thing as a percentage with highlighted sentences.

Confidence bands sit underneath the percentage. GPTZero categorises results as “uncertain”, “moderately confident” or “highly confident”, and states an error rate under 1% for the highly confident band (GPTZero technology). A result the tool itself labels uncertain is not a weaker finding of AI use; it is the tool declining to make one.

Length changes the answer. GPTZero’s FAQ states that “the accuracy of our model increases as more text is submitted to the model”, and that document-level classification is more accurate than paragraph-level (GPTZero FAQ). By GPTZero’s own account, a single paragraph gets a less reliable verdict than a full document.

Paraphraser Shield is GPTZero’s named defence against evasion. The company describes it as defending against bypass attempts including paraphrasing and homoglyph attacks — the trick of swapping visually identical characters from other alphabets into a document. Independent results on paraphrased and humanised text are covered below.

GPTZero 4o: what changed in September 2026

GPTZero announced its current detection model, GPTZero 4o, in a post dated 24 September 2026. Anything you read about GPTZero’s accuracy that predates that announcement is measuring a different model.

GPTZero describes four changes (GPTZero):

Source: Introducing GPTZero 4o, 24 September 2026. GPTZero also describes the model as punctuation-insensitive, meaning punctuation changes alone should not move the result.

AI Patterns names the features behind a score. Instead of a percentage alone, it points to specific patterns — ideas organised in threes, “Not X, but Y” constructions — that a writer can examine and dispute. Those patterns also appear in human writing. The feature launched separately on 19 August 2026 and was folded into 4o.

What GPTZero will and will not scan

Source: GPTZero FAQ and GPTZero technology, read 26 September 2026.

GPTZero excludes heavily modified text. The FAQ states: “Our classifier is not trained to identify AI-generated text after it has been heavily modified after generation.” A generated draft that has been substantially reworked is therefore outside what the tool is built to catch.

GPTZero also warns about procedural text. The FAQ states the classifier “can sometimes flag other machine-generated or highly procedural text as AI-generated”.

How to read a GPTZero score

The score has three layers.

A percentage without its confidence band is an incomplete result. GPTZero publishes an error rate for high-confidence classifications only.

Mixed is a separate classification. GPTZero’s API returns MIXED as a first-class classification, and its own benchmarking page gives mixed documents a separate and lower accuracy figure — 96.5%, against 99% for the clean human-versus-AI case (GPTZero benchmarking).

Which AI models GPTZero says it can detect

GPTZero’s technology page lists coverage of “ChatGPT (GPT-3, GPT-4, GPT-5, GPT-6)”, plus Google Gemini, Meta’s Llama and Anthropic’s Claude. Its September 2026 model announcement adds per-model recall figures for four current frontier models:

Source: Introducing GPTZero 4o, 24 September 2026. These are GPTZero’s own measurements on its own test sets, and the model names are GPTZero’s.

Two caveats apply to that table. Recall is the easier half of the problem. Catching AI text is straightforward if you are willing to flag human text along with it; the number that constrains a detector is the false-positive rate, and it is reported separately. And coverage of a model family is not coverage of what people actually submit. Every figure above is measured on unedited model output. GPTZero’s own FAQ excludes heavily modified AI text from its trained scope, so a 99.89% recall figure on raw GPT-5.6 prose says little about a paragraph a student generated and then rewrote.

What GPTZero claims about accuracy

GPTZero’s headline claim is a 99% accuracy rate on AI versus human text, with a false-positive rate at or below 1% and a false-negative rate of 0% (GPTZero benchmarking, published 30 January 2025 and last updated 27 July 2026). In GPTZero’s own wording: “GPTZero has an accuracy rate of 99% when detecting AI-generated text versus human writing, meaning we correctly classify AI writing 99 out of 100 times.”

The same benchmarking page gives separate and lower figures for mixed documents: 96.5% accuracy, a 0.9% false-positive rate and a 4.4% false-negative rate.

Sources: GPTZero benchmarking and Introducing GPTZero 4o.

GPTZero publishes more about its validation than most detector vendors. It names a third-party validator — Penn State’s AI/ML Research Lab — publishes the datasets it benchmarks on, and names its competitor by version when it compares.

GPTZero also publishes head-to-head comparisons against Pangram, its closest competitor on false-positive control. For GPTZero 4o it reports:

Source: Introducing GPTZero 4o, 24 September 2026. These are vendor-run comparisons of a competitor’s product, published by the vendor on 24 September 2026 and not yet independently replicated.

One vendor claim conflicts with its cited source. GPTZero published a post on 13 January 2026 reporting that it topped the University of Chicago Booth School of Business benchmark, with 99.5% accuracy, 99.3% recall and a 0.05% false-positive rate against Pangram’s 99.1% and 98.9% and Originality.ai’s 85.0% and 81.3% (GPTZero). The Chicago Booth Review’s own write-up of that research — by Brian Jabarian and Alex Imas, whose working paper is NBER 34223 — published on 2 December 2025, reaches a different conclusion, stating that “Pangram is the only AI detector maintaining policy-grade levels on our main metrics when evaluated on all four generative AI models” (Chicago Booth Review). The study has been revised more than once since its September 2025 working-paper release, and GPTZero has shipped model updates across the same period, so the two may be describing different runs against the same dataset. We could not reconcile them from the published versions, so both are reported.

What independent testing actually found

Three independent results are summarised below. They measure earlier GPTZero models than 4o.

The 2026 peer-reviewed study. Marijke Van Vlasselaer, Filip Van Droogenbroeck and Bram Spruyt ran a preregistered test of GPTZero, Copyleaks, Turnitin and Pangram against 160 English-language academic papers of at least 4,000 words. GPTZero classified 70% of the fully AI-generated papers as outright false negatives and the remaining 30% as partial false negatives, for a mean error of −85.70 percentage points against the true AI proportion. On hybrid documents it classified all 40 as false negatives. On human papers, the authors report that all four tools “correctly identified fully human texts” (VUB).

Two qualifications apply to every detector in the study. The AI text was generated in May 2025 — the fully AI papers with ChatGPT Deep Research, the inserted passages in the hybrid papers with GPT-4o — and GPTZero has shipped several detection models since, including 4o on 24 September 2026. And the study used long documents, the format GPTZero’s own documentation says it handles best.

RAID gives a per-generator breakdown. In the RAID paper, at a fixed 5% false-positive rate, GPTZero scored 99.4% on ChatGPT output and 28.9% on output from the Mistral base model, averaging 66.5% across 11 generators. It held up unusually well against adversarial attacks — 64.0% under paraphrase, 66.2% under homoglyph substitution, barely below its own baseline — where most detectors in the benchmark collapsed. Originality.ai averaged 85.0% and ZeroGPT 65.5% on the same test. RAID dates from May 2024 and measures a much older GPTZero.

The Carnegie Mellon paper looks at why detectors vary by generator. Yixuan Even Xu and colleagues found that commercial detectors, GPTZero included, judge text from unaligned base models far more human than text from the instruction-tuned versions of the identical model: 96.7% human for Llama3-8B base against 30.3% for the instruct variant. The authors conclude that detectors are identifying “artifacts of instruction tuning” — the assistant voice — rather than machine generation as such.

The false-positive problem, and who it falls on

The best-known evidence on who false positives fall on is from 2023. Weixin Liang, Mert Yuksekgonul, Yining Mao, Eric Wu and James Zou of Stanford tested seven GPT detectors against 91 TOEFL essays written by non-native English speakers and 88 essays by US eighth-graders. The detectors “misclassified over half of the TOEFL essays as ‘AI-generated’”, at an average false-positive rate of 61.22%, while achieving near-perfect accuracy on the US eighth-grade essays (Liang et al., Patterns, 2023). The published version does not break the figures down per detector, so no specific rate can be attributed to GPTZero.

The 2026 evidence points the other way on false positives. The human papers in the Van Vlasselaer study were written by non-native English-speaking graduate students, and the authors report that all four tools, GPTZero included, classified them correctly, with false positives “rare across all tools, suggesting improvement compared to earlier studies”. GPTZero’s 4o claim of one false positive in 10,402 PERSUADE essays is a vendor figure on a corpus of US student argumentative writing, not a multilingual corpus.

There is GPTZero-specific litigation. In February 2025 a student at Yale’s Executive MBA programme sued Yale University and individual administrators in the District of Connecticut, alleging he was flagged by GPTZero on a final exam, pressured to confess, and disciplined through irregular proceedings — with claims including discrimination and retaliation under Title VI of the Civil Rights Act on the basis that he is a non-native English speaker (Crowell & Moring client alert, 6 March 2025). The case is Doe v. Yale University et al, 3:25-cv-00159-SFR. We found no reported ruling on the merits as of 26 September 2026, and a filed allegation is not a finding.

GPTZero’s own guidance. Its FAQ states that “these results should not be used to punish students”, that “there always exist edge cases with both instances where AI is classified as human, and human is classified as AI”, and recommends that educators ask students to demonstrate understanding in a controlled setting and examine edit history or brainstorming notes instead.

Writing Replay and Origin: the pivot from detection to process

Since 2024 GPTZero has also built tools that record how a document was written.

Writing Replay records the construction of a document and plays it back. GPTZero introduced it with GPTZero Docs on 27 June 2024, describing a “video replay [that] captures the entire typing process” and showing which words were present in the original draft and which appeared in the final one (GPTZero). The company’s framing at launch was that with replay attached, “the origin of a document cannot be called into question”.

The Chrome extension extends the same capture into Google Docs. It provides live AI scan results and typing-pattern analysis as you write, a Writing Report covering writing-activity timeline, largest copy-and-pastes and average revision duration, AI-generated feedback delivered into Google Docs comments against a rubric or prompt, and AI detection on any web page (GPTZero Chrome extension). The extension is free to install.

Process capture records behaviour; a classifier score infers from style. A replay showing hours of drafting, restructures and a paste from a cited source is a record of what happened. Turnitin Clarity and Grammarly Authorship work on the same principle — see how AI detectors work.

It only works if it is switched on before the writing starts, and if students know about it. Retroactive process evidence does not exist. An institution that wants this has to put it in the assignment brief, which makes it a policy decision before it is a software purchase. For classroom tooling that does not rest on detection at all, see best AI for teachers.

What it costs

Source: GPTZero pricing, read 26 September 2026, and GPTZero support.

Three pricing facts are worth stating precisely, and one is unavailable.

Overages are billed at $0.00046 per word, and GPTZero’s support documentation states a buffer of up to 1 million words over plan before an upgrade is required (GPTZero support).

The plan line-up has consolidated. GPTZero’s pricing page on 26 September 2026 shows two consumer tiers, Premium and Professional. The Essential tier that third-party reviews were still listing in 2025 does not appear on it.

GPTZero’s US dollar list prices are data not available from this research. The pricing page displays prices in the currency of the visitor’s region rather than a fixed USD figure, and the widely republished USD prices in third-party reviews date from 2025 and disagree with each other. This page records the word allowances and points to the live page for the price. A promotional code advertised on the page at the time of reading discounts annual plans by a further 30%, so any quoted figure has a shelf life measured in weeks.

The free checker’s monthly word allowance is also data not available. GPTZero’s pricing page no longer shows a free-tier card, and the 10,000-words-a-month figure that circulates widely traces to 2025 documentation we could not confirm against the live product on 26 September 2026.

How GPTZero differs from the tools it gets compared with

This is a positioning comparison, not a ranking. For ranked accuracy across the whole market, with independent scores, see best AI detectors.

Sources: GPTZero FAQ, Turnitin FAQs, Van Vlasselaer et al.. Where a row says “not published here”, the vendor does not state it on the pages cited — we have not substituted an estimate.

The main difference is who runs it. GPTZero is a self-check a writer can run before anyone else sees the text. Turnitin is run by an instructor, the student does not see the report unless it is shared, and its result can feed a disciplinary process. A clean GPTZero result is not a defence against a Turnitin flag.

What to do, by situation

Best for a writer who wants a free sanity check before submitting: GPTZero’s free web checker, read with its confidence band. It costs nothing and takes seconds. It does not predict what Turnitin will say, and given GPTZero’s stated exclusion of heavily modified text and the 2026 false-negative results, a clean result is weak evidence that no AI was used.

Best for a student who wants to be able to prove they wrote something: process capture, switched on before you start. GPTZero’s Writing Replay, your platform’s native version history, or both. None of it can be enabled retroactively, so it has to be on before you start. See best AI for students for the assistants on the other side of the same desk.

Best for a student who has already been flagged: your writing process, not a counter-test. Do not run the essay through four detectors and email the screenshots. Preserve document version history, drafts, outlines, notes and timestamps. Ask which tool produced the score, what the exact figure was, and what confidence band it carried — GPTZero publishes an error rate only for high-confidence results. Ask whether your institution’s policy permits detector output as sole evidence, — GPTZero’s own FAQ says its results “should not be used to punish students”.

Best for an instructor deciding whether to act on a score: read the vendor’s own guidance first, then design the conversation. GPTZero recommends asking students to demonstrate understanding in a controlled setting and examining edit history or brainstorming notes.

For an institution writing policy: the 2026 study’s authors say detector output should not be used as sole evidence. They conclude that detection tools “can provide useful initial flags” but “should not be used as sole evidence in high-stakes decision-making”. GPTZero’s FAQ says the same about its own results.

If you write in English as a second language, or you have a diagnosed learning difference, put that in writing early. It is at the centre of the Yale litigation, and an institution that fails to consider it has made a procedural error you can name in an appeal.

Who owns GPTZero now

On 23 June 2026 Superhuman announced its intent to acquire GPTZero (BusinessWire). Superhuman is the company formerly named Grammarly, which rebranded on 29 October 2025 while keeping Grammarly as the name of the writing product. GPTZero’s own announcement the same day said it “plans to join Superhuman” and that “nothing changes for our daily users”, promising continued AI detection, hallucination detection and writing replay plus updated integrations with learning management systems, email clients, Chrome extensions and mobile and desktop apps (GPTZero). GPTZero was advised by Cooley (Cooley).

No completion of that acquisition has been announced as of 26 September 2026, and the deal value was not disclosed. GPTZero’s website footer linked to both Superhuman and Grammarly on that date.

If the deal closes, one company will own a leading AI writing assistant, a free AI detector, a paid AI detector agent, GPTZero’s consumer detector and two writing-process trackers.

Summary

• The product: a sentence-by-sentence classifier with a document verdict of human, mixed or AI, a confidence band, and since 24 September 2026 the GPTZero 4o model with AI Patterns.
• GPTZero’s claims: 99% accuracy and a false-positive rate at or below 1% on clean text; 96.5% on mixed documents; one false positive in 10,402 PERSUADE essays for 4o.
• Independent results: all human papers classified correctly but 0% strict accuracy on fully AI and hybrid papers in the June 2026 peer-reviewed study; 66.5% average in RAID (2024), with strong results on ChatGPT and weak results on base models. Both measured earlier models than 4o.
• GPTZero’s own limits: not trained on heavily modified AI text, can flag procedural text, and “should not be used to punish students”.
• Process capture: Writing Replay and the Chrome extension record how a document was written.
• Ownership: Superhuman announced a deal to acquire GPTZero on 23 June 2026; no completion had been announced by 26 September 2026.

Frequently asked questions

Is GPTZero accurate?

It depends entirely on what you give it. On clean, unedited AI output GPTZero reports 99% accuracy with a false-positive rate at or below 1%, and its September 2026 model claims one false positive in 10,402 essays on the PERSUADE student writing corpus. On long academic documents the independent results are much lower: a peer-reviewed study in the International Journal for Educational Integrity in June 2026 found GPTZero scored 0% strict accuracy on 40 fully AI-generated academic papers, 0% on 40 hybrid papers and 2.5% on humanised text, while correctly classifying all 40 human papers. In that study GPTZero produced no false accusations on human papers and missed the AI-generated ones; those papers were generated in May 2025, before GPTZero’s current model.

Does GPTZero detect ChatGPT?

GPTZero says it does: its technology page lists coverage across GPT-3, GPT-4, GPT-5 and GPT-6, and its 4o announcement reports 99.89% recall on GPT-5.6. Independent results vary. The 2024 RAID benchmark measured 99.4% on ChatGPT output at a fixed 5% false-positive rate, GPTZero’s best result of any generator in that test. The June 2026 peer-reviewed study, whose fully AI papers were written with ChatGPT Deep Research, found GPTZero detected none of them on its strict measure. GPTZero’s own FAQ states its classifier “is not trained to identify AI-generated text after it has been heavily modified after generation”, so a draft that has been rewritten is a different and much harder case.

Is GPTZero free?

There is a free checker at gptzero.me that requires no account, and the Chrome extension is free to install. GPTZero’s pricing page as displayed on 26 September 2026 lists two paid consumer tiers — Premium at 300,000 words a month and Professional at 500,000 words a month with LMS integration and batch scanning of up to 250 files — and directs teams to sales. The free tier’s monthly word allowance is not stated on the current pricing page, and the 10,000-words figure widely quoted online traces to 2025 documentation, so we are recording it as data not available rather than repeating it.

Why does GPTZero say my own writing is AI?

GPTZero’s own FAQ acknowledges that its classifier “can sometimes flag other machine-generated or highly procedural text as AI-generated” and that edge cases exist in both directions. Stanford researchers found in 2023 that seven detectors misclassified over half of TOEFL essays by non-native English speakers as AI-generated; a 2026 peer-reviewed study found all four tools it tested, GPTZero included, classified non-native graduate students’ papers correctly. Check the confidence band on your result: GPTZero publishes an error rate under 1% only for high-confidence classifications, and a result labelled uncertain is the tool declining to make a finding.

Is GPTZero the same as ZeroGPT?

No. They are different products from different companies with confusingly similar names, and the mix-up is common enough that people cite one while using the other. GPTZero was launched by Edward Tian on 3 January 2023 and is the product covered on this page. ZeroGPT is a separate free detector, and in the RAID benchmark the two scored 66.5% and 65.5% average accuracy respectively at a fixed 5% false-positive rate. For how both sit against the rest of the market, see our best AI detectors comparison.

Does GPTZero detect paraphrased or humanised text?

Partially. GPTZero’s named defence is Paraphraser Shield, which it says handles paraphrasing and homoglyph attacks, and RAID measured it as unusually robust to those specific attacks — 64.0% under paraphrase against a 66.5% baseline, where most detectors collapsed. But the June 2026 peer-reviewed study found GPTZero achieved 2.5% strict accuracy on humanised academic papers and classified all 40 hybrid human-and-AI papers as false negatives. GPTZero’s FAQ says heavily modified AI text is outside its trained scope. Our guide to AI humanisers covers that side of the arms race.

How does GPTZero compare to Turnitin?

They solve different problems and are not substitutes. GPTZero is a consumer tool anyone can run on their own text before submitting it; Turnitin’s AI writing indicator is visible only to instructors and administrators at institutions that license Turnitin Originality, and students cannot pre-check. On the one study that tested both against the same documents — the June 2026 International Journal for Educational Integrity paper — GPTZero and Turnitin both scored 0% strict accuracy on fully AI-generated papers and both produced no false positives on human papers. A clean GPTZero result is therefore not a defence against a Turnitin flag. See our full Turnitin AI detection guide for how the institutional side works.

Can teachers see if I used ChatGPT with GPTZero?

Not reliably from a detection score alone, and increasingly that is not what they are looking at. GPTZero’s classifier returns a probability, not a record, and its accuracy on edited or hybrid text is poor. What does produce a record is GPTZero’s process-capture side: the Chrome extension logs typing patterns, writing-activity timelines, the largest copy-and-pastes and average revision duration in Google Docs, and Writing Replay plays back a document’s construction. That evidence only exists if the tooling was switched on before you started writing, and it cannot be generated retroactively.

What should I do if GPTZero flags my essay?

Preserve your writing process evidence first — version history, drafts, outlines, notes and timestamps carry far more weight than a second detector’s opinion. Ask what the exact score was and which confidence band it carried, since GPTZero publishes an error rate only for high-confidence results. Point to GPTZero’s own FAQ, which states that “these results should not be used to punish students” and recommends that educators examine edit history or brainstorming notes instead. Check whether your institution’s policy permits detector output as sole evidence, because most now state that it does not. If English is not your first language or you have a diagnosed learning difference, say so in writing early: that circumstance is at the centre of a pending Yale lawsuit over a GPTZero flag, filed in February 2025 in the District of Connecticut.

Who owns GPTZero?

GPTZero is an independent company founded by Edward Tian in January 2023, subject to a pending acquisition. On 23 June 2026 Superhuman — the company formerly named Grammarly, which rebranded on 29 October 2025 — announced its intent to acquire GPTZero, with the deal value undisclosed. No completion has been announced as of 26 September 2026. GPTZero said at the time that it “plans to join Superhuman” and that “nothing changes for our daily users”. If the deal closes, one company will own both a leading AI writing assistant and the most-used consumer AI detector.

Does GPTZero work in languages other than English?

GPTZero’s FAQ states support for English, Spanish, French, German and other languages, and multilingual AI detection is listed as a Premium plan feature. However, GPTZero’s technology page says the model “performs best of longer texts and English prose”, and the company publishes no per-language accuracy or false-positive figures. Every independent study cited on this page tested English text only. A non-English score rests on less published evidence than an English one.

What is the GPTZero 4o model?

GPTZero 4o is GPTZero’s current detection model, announced on 24 September 2026. GPTZero describes it as adding expanded training data across academic and social media writing, a revamped architecture designed to resist paraphrasing, punctuation-insensitive scoring, a Masked Detection feature that excludes headers so a title change does not move the score, and AI Patterns, which identifies and explains ten recurring stylistic patterns such as organising ideas in threes or “Not X, but Y” framing. GPTZero reports AI recall of 99.89% on GPT-5.6, 99.87% on Gemini 3.6, 99.37% on Grok 4.5 and 99.04% on Claude 5, and one false positive in 10,402 PERSUADE essays. These are vendor figures, published on 24 September 2026, and have not been independently replicated.

Written 26 September 2026 from GPTZero’s own technology page, FAQ, pricing page, support documentation and newsroom as published on that date, plus the peer-reviewed International Journal for Educational Integrity study, the RAID benchmark paper, the Carnegie Mellon base-model paper, the Stanford non-native-speaker study, the Chicago Booth Review write-up and NBER working paper, and the BusinessWire, Cooley and GPTZero acquisition announcements. GPTZero shipped a new detection model on 24 September 2026, so every independent result cited here measures an earlier model and is dated accordingly. GPTZero’s own Chicago Booth figures conflict with the Chicago Booth Review’s published conclusion, and both are reported rather than reconciled. Where GPTZero publishes no figure — US dollar list prices, the free tier’s word allowance, per-language accuracy and per-model false-positive rates — this page records “data not available” rather than estimating.

← All guides