Guide
GPTZero: How It Works, How Accurate It Is, and What to Do If It Flags You
A plain-English guide to GPTZero: how its classifier scores a document, what changed in GPTZero 4o, what its accuracy claims are measured on, what independent testing found, what Writing Replay records, and who owns it.
Quick answer: GPTZero is an AI detector that classifies text sentence by sentence and returns a document-level verdict of human, mixed or AI. Its pricing page says it is used by over 10 million teachers and students in more than 100 countries. GPTZero’s published claim is 99% accuracy with a false-positive rate at or below 1%, and its current model, GPTZero 4o, made available to all users on 20 September 2026, reports one false positive in 10,402 student essays on the PERSUADE corpus (GPTZero). In a peer-reviewed study published in the International Journal for Educational Integrity in June 2026, the authors report GPTZero correctly identified the fully human papers, and gave papers written entirely by AI an average AI score of 14.3% (Van Vlasselaer et al.). That study tested an earlier GPTZero model. GPTZero is free to run on your own writing, and its FAQ states that “these results should not be used to punish students”. Superhuman, the company formerly named Grammarly, completed its acquisition of GPTZero on 30 September 2026.
This guide covers GPTZero specifically. For the underlying detection mechanism across all tools, see how AI detectors work. For ranked alternatives, see best AI detectors. For the institutional detector most students are assessed by, see Turnitin AI detection.
What GPTZero is
GPTZero is an AI detection product sold directly to individuals and institutions.
It launched as a student project. Edward Tian, then a Princeton undergraduate, put GPTZero online in January 2023. TechCrunch reported in June 2024 that the company had raised a $10 million Series A.
Anyone can self-check. A student can paste their own essay into GPTZero before submitting it. That is not possible with Turnitin, whose AI writing indicator is visible only to instructors and administrators.
It is a suite. GPTZero’s navigation 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 partnerships. GPTZero’s pricing page describes the company as the “Official AI detector partners of the American Federation of Teachers”, and GPTZero has published a case study describing AP Verify using GPTZero for generative AI text detection.
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”. Perplexity and burstiness, which were often used to explain GPTZero in 2023, are covered in how AI detectors work.
The API returns 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).
Length matters. 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).
Paraphraser Shield is GPTZero’s named defence against evasion, which it says covers paraphrasing and homoglyph attacks (swapping in visually identical characters from other alphabets).
GPTZero 4o: what changed
GPTZero made its current detection model, GPTZero 4o, available to all users on 20 September 2026 (GPTZero). Results published before then measure earlier models.
| Change | What GPTZero says |
|---|---|
| Training data | ”scaled up our training data across academic writing and social media writing” |
| Architecture | ”revamped the model architecture to be robust to paraphrasing” |
| AI Patterns | Identifies and explains ten recurring stylistic and rhetorical patterns, such as organising ideas in threes or “Not X, but Y” framing |
| Masked Detection | Excludes headers from analysis so that changing a title does not change the score |
GPTZero also describes the model as punctuation-insensitive. AI Patterns points to specific patterns behind a score that a writer can examine; 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
| Requirement | What GPTZero states |
|---|---|
| Text length per scan | Up to 150,000 characters, or a batch of up to 250 files |
| Optimal length | Longer is better; document-level beats paragraph-level |
| Languages | English, Spanish, French, German “and other languages” |
| Text type | Performs best on longer texts and English prose |
| Heavily edited AI text | Outside what the classifier is trained to identify |
Source: GPTZero FAQ and GPTZero technology.
Heavily modified text. The FAQ states: “Our classifier is not trained to identify AI-generated text after it has been heavily modified after generation.”
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
| Layer | What it is |
|---|---|
| Document verdict | Human, mixed or AI, with a percentage |
| Confidence band | Uncertain, moderately confident or highly confident; GPTZero states an error rate under 1% for the highly confident band |
| Sentence highlights | Per-sentence probabilities, plus AI Patterns |
Read the percentage with its confidence band. GPTZero publishes an error rate for high-confidence classifications only.
Mixed is a separate classification. GPTZero’s benchmarking page gives mixed documents a separate accuracy figure of 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 4o announcement reports recall figures for four models:
| Model GPTZero tested | AI recall GPTZero reports |
|---|---|
| GPT-5.6 | 99.89% |
| Gemini 3.6 | 99.87% |
| Grok 4.5 | 99.37% |
| Claude 5 | 99.04% |
Source: Introducing GPTZero 4o. These are GPTZero’s own measurements on its own test sets. Recall measures how much AI text is caught; the false-positive rate, reported separately, measures how much human text is wrongly flagged.
What GPTZero says about accuracy
GPTZero’s benchmarking page states: “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” (GPTZero benchmarking). It names Penn State’s AI/ML Research Lab as a third-party validator.
| GPTZero’s published claim | Figure | Case it applies to |
|---|---|---|
| Accuracy | 99% | Clean AI versus clean human text |
| False-positive rate | At or below 1% | Clean AI versus clean human text |
| Accuracy | 96.5% | Mixed human and AI documents |
| False-positive rate | 0.9% | Mixed human and AI documents |
| False-negative rate | 4.4% | Mixed human and AI documents |
| False positives | 1 in 10,402 essays | PERSUADE student writing corpus, GPTZero 4o |
Sources: GPTZero benchmarking and Introducing GPTZero 4o.
GPTZero also publishes comparisons with Pangram. For GPTZero 4o it reports:
| Benchmark | GPTZero 4o | Pangram |
|---|---|---|
| EpochAI | 0% false positives, 2.69% false negatives | 2.86% false negatives (Pangram 4) |
| Graphite | 1.36% false positives, 0.07% false negatives | 1.84% false positives, 1.40% false negatives |
| DetectRL | 96.53% average F1 | 95.30% average F1 |
Source: Introducing GPTZero 4o. These are comparisons run and published by GPTZero.
GPTZero’s Chicago Booth figures come from its own re-run. On 13 January 2026 GPTZero reported 99.5% accuracy, 99.3% recall and a 0.05% false-positive rate on the University of Chicago Booth School of Business benchmark, saying it re-ran the benchmark after the researchers used the wrong API field for its score (GPTZero). The Chicago Booth Review’s write-up of the research by Brian Jabarian and Alex Imas (NBER 34223), published on 2 December 2025, states 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).
What independent testing found
The results below measure earlier GPTZero models than 4o.
| Study | Date | What it tested | GPTZero result |
|---|---|---|---|
| Van Vlasselaer, Van Droogenbroeck & Spruyt, International Journal for Educational Integrity | June 2026 | 160 academic papers: 40 each fully human, fully AI, hybrid and humanised | Human papers correctly identified (average AI score 1.9%); fully AI papers 14.3% AI on average; hybrid 21.9% against an actual 46.3%; humanised 21.2% against an actual 44.8% |
| Dugan et al., RAID, ACL | May 2024 | 11 generators across 8 domains with 11 adversarial attacks, at a fixed 5% false-positive rate | 66.5% average accuracy; 99.4% on ChatGPT; 28.9% on the Mistral base model; 64.0% under paraphrase attack |
| Xu, Zhong, Raghunathan, Fang & Kolter, Carnegie Mellon | May 2026 | Base models versus instruction-tuned models from the same family | Judged Llama3-8B base continuations 96.7% human, against 30.3% for the instruction-tuned version |
The 2026 Vrije Universiteit Brussel study tested GPTZero, Copyleaks, Turnitin and Pangram (VUB). The authors report that all four tools “correctly identified fully human texts”; GPTZero’s average AI score on those papers was 1.9%, against 0.0% for the other three. On the 40 papers written entirely with GPT-4o Deep Research, GPTZero’s average AI score was 14.3%, a mean error of −85.70 percentage points; Copyleaks’ result was almost identical. The authors suggest the Deep Research model may explain why most tools did less well on fully AI papers than on hybrid ones.
RAID measured GPTZero at 99.4% on ChatGPT output and 28.9% on the Mistral base model, averaging 66.5% across 11 generators at a fixed 5% false-positive rate. It scored 64.0% under paraphrase and 66.2% under homoglyph substitution, close to its baseline; five other detectors dropped an average of 40.6% under homoglyphs. Originality.ai averaged 85.0% and ZeroGPT 65.5% on the same test.
The Carnegie Mellon paper found that commercial detectors, GPTZero included, judge text from base models far more human than text from instruction-tuned versions of the same model: 96.7% human for Llama3-8B base against 30.3% for the instruct variant. The authors conclude detectors are identifying “artifacts of instruction tuning”.
False positives
The best-known study of who false positives affect is from 2023. Weixin Liang and colleagues at Stanford tested seven detectors on 91 TOEFL essays 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% (Liang et al., Patterns, 2023).
A lawsuit. In February 2025 a student in Yale’s Executive MBA programme sued Yale University and individual administrators in the District of Connecticut, alleging that he was flagged by GPTZero on a final exam and disciplined through irregular proceedings, with claims including discrimination under Title VI of the Civil Rights Act (Crowell & Moring, 6 March 2025). Yale disputes the claims; Poets&Quants reported Yale’s position that its Honor Committee’s finding rested on similarities between the student’s answers and ChatGPT responses rather than on GPTZero. A judge denied the student’s request for an injunction in May 2025, and the case was ongoing as of August 2026.
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.
Writing Replay and Origin
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 28 June 2024, describing a “video replay [that] captures the entire typing process” (GPTZero).
The Chrome extension offers a Writing Report that replays a document’s edit history, typing-pattern analysis, and AI detection on any webpage (GPTZero Chrome extension). It is free with a GPTZero account, and the extension page says “20M+ GPTZero users”.
Process capture records behaviour; a classifier score infers from style. Turnitin Clarity and Grammarly Authorship work on the same principle. Process capture has to be switched on before writing starts. For classroom approaches that do not rest on detection, see best AI for teachers.
What it costs
| Plan | Words per month | What it adds |
|---|---|---|
| Free web checker | Not stated on the pricing page | Basic scan at gptzero.me |
| Premium | 300,000 | Advanced AI Scan, multilingual AI detection, downloadable AI reports, AI highlights in the Chrome extension |
| Professional | 500,000 | Everything in Premium, plus scanning up to 250 files at once, page-by-page scanning, enterprise-grade security and LMS integration |
| Team and Enterprise | Shared team credits | Professional features across an organisation; contact sales |
| API | Metered | 30,000 requests per hour on all subscription plans |
Source: GPTZero pricing and GPTZero support. GPTZero shows prices in the visitor’s local currency. Overage rates are shown at purchase; on the 300,000-word plan GPTZero’s support page caps overage at 700,000 extra words, 1 million in total (GPTZero support).
How GPTZero differs from Turnitin and Pangram
This is a comparison of how the tools are used, not a ranking. For ranked accuracy, see best AI detectors.
| GPTZero | Turnitin AI writing detection | Pangram | |
|---|---|---|---|
| Who can run it | Anyone; free basic check | Instructors and administrators at institutions licensing Turnitin Originality | Anyone; free daily checks |
| Can a student self-check | Yes | No | Yes |
| Where the result appears | gptzero.me, Chrome extension, Word add-in, API | Inside the Similarity Report | Consumer report and API |
| Low-score handling | Percentage shown with a confidence band | No score or highlights between 1% and 19% | — |
| Process-capture product | Writing Replay and the Chrome extension | Turnitin Clarity | — |
| Owner | Superhuman (acquisition completed 30 September 2026) | Turnitin | Pangram Labs |
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, and the student does not see the report unless it is shared. A GPTZero result does not predict a Turnitin result.
What to do, by situation
A writer checking work before submitting. GPTZero’s free web checker gives a quick read; look at the confidence band as well as the percentage. GPTZero says it is not trained on heavily modified AI text.
A student who wants to be able to show they wrote something. Use process capture from the start: GPTZero’s Writing Replay, your platform’s version history, or both. See best AI for students for the assistants students use.
A student who has been flagged. 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. 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”.
An instructor deciding whether to act on a score. GPTZero recommends asking students to demonstrate understanding in a controlled setting and examining edit history or brainstorming notes.
An institution writing policy. The 2026 study’s authors 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.
Who owns GPTZero
Superhuman, the company formerly named Grammarly, announced on 23 June 2026 that it would acquire GPTZero (BusinessWire), and said on 30 September 2026 that GPTZero “is now part of Superhuman” (Superhuman). Terms were not disclosed. Grammarly rebranded as Superhuman on 29 October 2025, keeping Grammarly as the name of the writing product. At the announcement, GPTZero said that “nothing changes for our daily users”, promising continued AI detection, hallucination detection and writing replay (GPTZero). GPTZero was advised by Cooley (Cooley).
Summary
- The product: a sentence-by-sentence classifier with a document verdict of human, mixed or AI and a confidence band, running the GPTZero 4o model since 20 September 2026.
- 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: in the June 2026 peer-reviewed study, human papers were correctly identified and fully AI papers averaged a 14.3% AI score; RAID (2024) measured 66.5% on average. Both tested earlier models than 4o.
- GPTZero’s own guidance: not trained on heavily modified AI text, can flag procedural text, and results “should not be used to punish students”.
- Ownership: part of Superhuman since 30 September 2026.
Frequently asked questions
Is GPTZero accurate?
GPTZero reports 99% accuracy with a false-positive rate at or below 1% on clean, unedited text, and its 4o model reports one false positive in 10,402 essays on the PERSUADE student writing corpus. In a peer-reviewed study in the International Journal for Educational Integrity in June 2026, the authors report GPTZero correctly identified the human papers, and its average AI score on papers written entirely by AI was 14.3%. That study tested a model earlier than 4o.
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. The 2024 RAID benchmark measured 99.4% on ChatGPT output at a fixed 5% false-positive rate. GPTZero’s FAQ states its classifier “is not trained to identify AI-generated text after it has been heavily modified after generation”.
Is GPTZero free?
There is a free checker at gptzero.me, and the Chrome extension is free with a GPTZero account. Paid tiers are 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; teams are directed to sales. GPTZero’s pricing page does not state the free tier’s monthly allowance.
Why does GPTZero say my own writing is AI?
GPTZero’s FAQ says its classifier “can sometimes flag other machine-generated or highly procedural text as AI-generated” and that edge cases exist in both directions. A 2023 Stanford study found seven detectors misclassified over half of TOEFL essays by non-native English speakers as AI-generated. Check the confidence band on your result: GPTZero publishes an error rate under 1% only for high-confidence classifications.
Is GPTZero the same as ZeroGPT?
No. They are different products from different companies with similar names. GPTZero was launched by Edward Tian in January 2023 and is the product covered on this page. ZeroGPT is a separate free detector. In the RAID benchmark they scored 66.5% and 65.5% average accuracy respectively at a fixed 5% false-positive rate.
Does GPTZero detect paraphrased or humanised text?
GPTZero says its Paraphraser Shield handles paraphrasing and homoglyph attacks, and RAID measured 64.0% under paraphrase against a 66.5% baseline. In the June 2026 study, GPTZero’s average AI score on humanised papers was 21.2% against an actual 44.8%. GPTZero’s FAQ says heavily modified AI text is outside its trained scope. For that side of the subject, see AI humanisers.
How does GPTZero compare to Turnitin?
They are used differently. GPTZero is a 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. A GPTZero result does not predict a Turnitin result. See Turnitin AI detection.
Can teachers see if I used ChatGPT with GPTZero?
GPTZero’s classifier returns a probability, not a record. What produces a record is GPTZero’s process-capture side: Writing Replay and the Chrome extension’s Writing Report replay how a document was built. That evidence only exists if the tooling was switched on before writing started.
What should I do if GPTZero flags my essay?
Preserve your writing process evidence first: version history, drafts, outlines, notes and timestamps. Ask what the exact score was and which confidence band it carried. GPTZero’s FAQ states that “these results should not be used to punish students” and recommends that educators examine edit history or brainstorming notes. If English is not your first language or you have a diagnosed learning difference, say so in writing early.
Who owns GPTZero?
Superhuman, the company formerly named Grammarly. Superhuman announced the acquisition on 23 June 2026 and said on 30 September 2026 that GPTZero “is now part of Superhuman”. Terms were not disclosed. GPTZero was founded by Edward Tian in January 2023.
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 a Premium feature. GPTZero’s technology page says the model “performs best of longer texts and English prose”, and it publishes no per-language accuracy figures. The independent studies cited on this page tested English text.
What is the GPTZero 4o model?
GPTZero 4o is GPTZero’s current detection model, available to all users from 20 September 2026. GPTZero says it scaled up training data across academic and social media writing, revamped the architecture to be robust to paraphrasing, made scoring punctuation-insensitive, added Masked Detection so a title change does not move the score, and added AI Patterns, which explains ten recurring stylistic patterns. 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.