The short answer
GPTZero is an AI detector with a free plan that scores whole documents and highlights sentences. In Plagino's March 2026 benchmark of 1,000 documents, it caught 88.9% of AI-written documents and wrongly flagged 5.7% of human-written ones. Plagino caught 97.4% with 0.8% false flags; Turnitin caught 98.2% with 1.4%.
What is the GPTZero AI detector and how does it work?
GPTZero is an AI content detector that estimates whether text was written by a large language model such as ChatGPT, Claude or Gemini. A Princeton undergraduate, Edward Tian, launched it in January 2023, according to Wikipedia's GPTZero entry. Its home page now says GPTZero is part of Superhuman, and Wikipedia reports that Superhuman acquired it in June 2026.
A scan with this AI detection tool returns a score for the whole document and highlights the sentences that drive it. Through its API, GPTZero sorts a document into three classes: human only, mixed or AI only, with a confidence category on top, as its FAQ explains. Early versions leaned on perplexity (how predictable the wording is) and burstiness (how much sentence style varies). Wikipedia reports a move to a deep learning classifier in autumn 2023, and GPTZero's how it works page describes a model that turns text into numbers and draws a boundary between human and AI writing. It also says GPTZero favors avoiding false accusations over catching every AI text.
What changed in GPTZero in 2026
As of October 2026, GPTZero's default model is GPTZero 4o, which its release post says became available to all users on 20 September 2026. An earlier release, model 4.8b on 1 August 2026, began masking headers so they no longer affect the score. Masked text shows in gray.
GPTZero also says punctuation, such as adding or removing an em dash, should not decide the result. These are vendor statements. Plagino tested the March 2026 version of the tool, so they matter for how far our numbers still apply, a point we return to below.
Is GPTZero accurate? Results from Plagino's benchmark
We tested GPTZero in Plagino's benchmark alongside six other detectors. The corpus had 1,000 documents: 500 AI-written (GPT-4o, Claude and Gemini, some lightly human-edited) and 500 human-written. Every tool saw the same files through its own web app on default settings, in March 2026. A document counted as flagged if the tool flagged it at all.
We report two numbers because they describe different mistakes. The detection rate is the share of AI-written documents a tool caught. The false-positive rate is the share of human-written documents it wrongly called AI. For someone deciding whether to act on a flag, the second number matters more.
| Detector | AI documents caught | 95% interval | Human documents wrongly flagged | 95% interval |
|---|---|---|---|---|
| Turnitin (reference) | 98.2% | 96.6–99.1% | 1.4% | 0.7–2.9% |
| Plagino | 97.4% | 95.6–98.5% | 0.8% | 0.3–2.0% |
| GPTZero | 88.9% | 85.8–91.4% | 5.7% | 4.0–8.1% |
GPTZero caught 88.9% of the AI-written documents, fifth of seven detectors. Its interval does not overlap with Plagino's, so on this sample Plagino caught more AI writing. Roughly 1 in 9 AI documents slipped past GPTZero.
The false-positive rate is the larger gap. GPTZero wrongly flagged 5.7% of human-written documents, about 1 in 18 and the second-highest of the seven. That is about seven times Plagino's 0.8% and four times Turnitin's 1.4%, and the intervals do not overlap. In a class of 100 honest essays, a rate like that would put about six students under suspicion.
GPTZero is also not close to Turnitin in this test. It trailed by 9.3 points on detection and flagged four times as many human documents. Turnitin and GPTZero are separate products, though, so a GPTZero result does not predict what your institution's Turnitin report will show.
Why do GPTZero's own accuracy numbers differ from independent tests?
GPTZero advertises 99% accuracy on its home page, and its figures for false positives are far below the 5.7% we measured. Both can be true, because each number describes a different test. The table below puts the main figures side by side, with who produced each one and the catch that comes with it. Dates and figures are as of October 2026.
| Source | Who ran it | False-positive result | Catch |
|---|---|---|---|
| GPTZero benchmarking page (first posted February 2026) | GPTZero | 0.08% average across four text types (model 4.3b) | GPTZero built the test sets: 1,000 human and 1,000 AI texts per type, from GPT-5.2, Gemini 3 Pro, Claude Sonnet 4.5 and Grok 4 Fast |
| GPTZero 4o release post (September 2026) | GPTZero | 1 false positive in 10,402 student essays (PERSUADE dataset) | GPTZero ran the check on a public essay dataset; no outside team repeated it |
| Graphite study (May 2026), as reported by GPTZero | Outside researchers, reported by the vendor | 1.36% on 15,000 pre-ChatGPT human articles | Plagino has not read the study itself, only GPTZero's summary of it |
| RAID benchmark, as reported by GPTZero | Outside dataset, reported by the vendor | Set at 1% by design; GPTZero reports 95.7% of AI texts caught at that setting | This is a fixed threshold, not a measured false-positive rate |
| Dik, Erdem and Dik preprint (June 2025) | Academic preprint, not peer reviewed | 8 of 50 human essays flagged (16%); all 28 AI essays caught | Small sample; essays of 40 to 800 words; GPTZero version not stated |
| ZDNET test (2025) | Journalist | Scored 80%: 4 of 5 test texts right, missing one ChatGPT-written text | Five texts only, so it shows variation more than a rate |
| Plagino benchmark (March 2026) | Plagino | 5.7% of 500 human documents (4.0–8.1%) | Web app on default settings; an older GPTZero model than today's default |
The sources are GPTZero's benchmarking page, its 4o release post and its RAID results post, the Dik, Erdem and Dik preprint on arXiv, and ZDNET's 2025 detector round-up.
Three things explain most of the gap. First, the test texts differ. GPTZero tests on text it selects after checking that it was not used in training, while an outside test uses documents the vendor has never seen. Second, the model keeps changing. GPTZero says it shipped 15 model releases in 2025 alone, and our March 2026 run predates both the August 2026 header masking and GPTZero 4o. Third, counting rules differ. GPTZero's own mapping counts any mixed result as AI, and we counted a document as flagged if any part of it was flagged.
Small independent tests add noise. In the 50-essay preprint, a single extra flagged essay moves the rate by 2 points. A vendor figure is a claim about its own test set, not a promise about your essay. Our guide to AI detector false positives explains how to read a rate.
Why does GPTZero say I used AI when I didn't?
A flag means GPTZero's model found patterns it links with AI writing, and human writing can share those patterns. It is a probability, not proof. GPTZero's own FAQ says results should not be used to punish students.
The same FAQ lists when the tool is weaker. Accuracy rises as more text is submitted, so a whole document is more reliable than a paragraph. It is strongest on English prose by adults, and it can flag highly procedural or machine-generated text, such as lab steps or boilerplate. In the Dik preprint, human essays under 100 words and over 350 words were the ones most likely to be flagged.
Writers who learned English as a second language face extra risk. A 2023 Stanford study found that seven detectors, GPTZero among them, misclassified over half of TOEFL essays as AI on average. GPTZero says it has since trained its model to cut false positives on English-learner writing to about 1%. That is a vendor claim, and Plagino has not tested it.
If a flag lands on work you wrote yourself, do not rewrite it to chase a lower score. Gather evidence of your process and use your institution's procedure. Our guide for students falsely accused of using AI walks through that process.
What to do if GPTZero flags your own writing
- 1
Keep your drafts and history. Save your notes, outlines and earlier drafts. If you wrote in Google Docs or Word, version history shows how the text grew over time.
- 2
Scan the whole document. A single paragraph gives a weaker result than the full text, because GPTZero says accuracy improves with length. Scan the full piece, not just the part that worries you.
- 3
Read the highlighted sentences. Check whether the highlights fall on definitions, method steps, lists or standard phrases. Those are the places where human writing most often looks formulaic.
- 4
Note the result class and date. Write down whether the result said AI, mixed or human, the percentage, and the date. GPTZero updates its model often, so a result from last month may differ today.
- 5
Use the official process. Share your drafts with your instructor and ask how the school reviews AI flags. A detector score alone should not decide an academic case.
Is GPTZero free, and what does it cost?
GPTZero has a free plan. Its student page says you can scan up to 10,000 words a month and use the Chrome extension at no cost. Paid plans raise the limits and add features. The details below come from the GPTZero pricing page, checked in October 2026.
GPTZero shows prices in local currency. The page showed Pakistani rupees when we viewed it, so Plagino could not confirm US dollar prices from GPTZero itself. A third-party listing reports Premium at $155.88 a year and Professional at $299.88 a year. Treat those figures as unverified and confirm your own price before you subscribe.
| GPTZero | Plagino | |
|---|---|---|
| Free use | Free plan: 10,000 words a month and the Chrome extension | 3 scans per month, no card |
| Paid plans | Premium: 300,000 words a month, 50,000 characters per scan. Professional: 500,000 words a month, 150,000 characters per scan. Annual billing saves 45% | Pro: $19/month for 20 scans per day |
| Teams and schools | Team plans with shared credits and unified billing, Enterprise through sales, and an API | Team: custom pricing for departments and writing centres |
| How you submit | Paste text or upload a file, Chrome extension, Google Docs writing replay, Canvas and Google Classroom integrations | .docx upload in the web app |
| Report | Document score, sentence highlights and a downloadable AI report on paid plans | PDF with sentence-level AI highlights and matched sources |
| Plagiarism check | Separate plagiarism checker product | Open-access academic literature via CORE only |
| Writing-process proof | Writing replay shows a document being written in Google Docs | None |
GPTZero's paid plans are measured in words a month, while Plagino's are measured in scans a day. Plagino's pricing page has the current plan details. If you need a plagiarism check as well, Plagino's plagiarism checker matches against open-access research through CORE, and does not search the public web.
What does GPTZero do with the text you scan?
GPTZero's FAQ says it stores text scanned in its dashboard and uses it only in aggregate to improve the service. It says it does not store documents sent through its API, and its privacy policy, last updated 2 April 2026, repeats that API text is not stored by default. A support article adds that GPTZero keeps a copy of a submission but does not reproduce it or disclose it to third parties.
GPTZero also says it is SOC 2 Type II audited and complies with FERPA, GDPR and CCPA, and that personal data is deleted on request. These are the company's own statements. If you are scanning an unpublished thesis or manuscript, read the policy first. Plagino's own terms are on its privacy page.
Where GPTZero is the better AI checker
GPTZero does more than Plagino in several areas, and for some users those matter more than a lower false-positive rate.
Proof of how a document was written
GPTZero's Google Docs tool replays a document as it was written, so a reader can see copy and pastes and typing patterns. That gives a student a way to show process, which a score alone cannot do. Plagino has no equivalent.
Classrooms, teams and developers
GPTZero lists Canvas and Google Classroom integrations, batch scanning of up to 250 files, shared team credits and an API. Its pricing page names the American Federation of Teachers as an official partner. A teacher grading a whole class, or a team building detection into its own app, will find more here. Plagino has no LMS integration. Our page on Plagino for teachers covers what it does offer.
Free allowance, languages and open benchmarking
The free plan covers 10,000 words a month, against 3 scans a month on Plagino's free plan. GPTZero says it fully supports English, German, Portuguese, French and Spanish, and its mixed class gives a more nuanced answer than a single score. It also publishes benchmark datasets and model versions, which is more transparent than most vendors.
Where Plagino is the better choice
Plagino is our tool, so read this with that in mind. The clearest difference in our benchmark was false positives. Plagino wrongly flagged 0.8% of human-written documents against GPTZero's 5.7%, and the gap is outside the margin of error. If you are a student checking your own essay, or a teacher who wants to avoid accusing an honest writer, a lower false-positive rate is the number to prioritize. Plagino also caught more AI writing (97.4% against 88.9%) on this sample.
Plagino is built for academic drafts. You upload a .docx file and get a PDF report with sentence-level AI-writing highlights and matched academic sources, which you can keep and discuss with a tutor. The free plan gives 3 scans a month with no card, and Pro costs $19 a month for 20 scans a day. If you are looking for a GPTZero alternative with a lower measured false-flag rate, start there.
Whichever tool you use, a score is a signal, not proof. For other head-to-head results, see Plagino vs Originality.ai, Plagino vs Copyleaks, Plagino vs Winston AI and Plagino vs Scribbr. Our guide to the best alternatives to Turnitin compares the wider field.
Test conditions and limitations
These limits apply to Plagino's results as much as to GPTZero's.
What the benchmark does not cover
The corpus was English only, and the AI documents came from GPT-4o, Claude and Gemini, so results on other languages or newer models may differ. We scored at document level, so one flagged sentence in a human essay counted as a false positive. We used each tool's default settings in March 2026.
GPTZero has changed since then, with header masking in August 2026 and GPTZero 4o in September 2026. Our numbers describe the tool as we tested it, not necessarily the one you will use today. We plan to re-test GPTZero in the next quarterly benchmark and update this page.
Each rate rests on 500 documents. That separates large gaps, such as the false-positive difference between GPTZero and Plagino, but not small ones. Plagino has not benchmarked GPTZero's plagiarism checker, writing replay or non-English detection.
For a wider look at how detectors err and why, read our guide to AI detector false positives, or try the Plagino AI detector on your own draft.
Frequently asked questions
Is GPTZero accurate?
It catches most AI writing but flagged more human writing than most rivals in Plagino's test. In Plagino's March 2026 benchmark of 1,000 documents, GPTZero caught 88.9% of AI-written documents and wrongly flagged 5.7% of human-written ones. GPTZero's own published tests report far lower false-positive rates, so treat any single figure as a claim about one test set.
Is GPTZero free?
Yes, GPTZero has a free plan with 10,000 words a month and a Chrome extension. Paid Premium and Professional plans raise the limit to 300,000 and 500,000 words a month, as of October 2026. Prices show in local currency on its pricing page, so check your own price before you subscribe.
How does GPTZero work?
GPTZero runs your text through a trained model that decides whether it looks human, mixed or AI-written. Early versions used perplexity and burstiness, which measure how predictable the wording is and how much sentence style varies. Its own explainer now describes turning text into numbers and drawing a boundary between human and AI writing, with sentence highlights showing what drove the result.
Is GPTZero as good as Turnitin?
Not in Plagino's March 2026 benchmark. Turnitin caught 98.2% of AI-written documents and wrongly flagged 1.4% of human ones, against 88.9% and 5.7% for GPTZero. The two are separate products, and Turnitin is sold only to institutions, so most students cannot run their own Turnitin AI check.
Does Turnitin use GPTZero?
No. GPTZero and Turnitin are competing products from separate companies. GPTZero's benchmarking page names Turnitin as a competitor and says Turnitin limits access to educators, so GPTZero cannot test it freely. A GPTZero score therefore does not reliably predict what your school's Turnitin report will show.
Why does GPTZero say my writing is AI when I wrote it myself?
GPTZero found patterns it links with AI writing, and human writing can share them. Short passages, formulaic text such as method steps, and writing by non-native English speakers are more at risk. GPTZero's own FAQ says its results should not be used to punish students, so keep your drafts and ask for a review.
Can GPTZero detect ChatGPT?
Yes, GPTZero is built to detect text from ChatGPT, Claude, Gemini and other models. In Plagino's benchmark, which used documents from GPT-4o, Claude and Gemini, it caught 88.9% of AI-written documents, so roughly 1 in 9 slipped through. GPTZero says its newer models catch more, but Plagino has not re-tested them.
What does GPTZero do with my text?
GPTZero stores text scanned in its dashboard and says it uses it only in aggregate to improve the service. Text sent through its API is not stored, according to its FAQ. Its support center says a copy of your submission is kept but never shown to third parties, so read the privacy policy before scanning unpublished work.
Is GPTZero fair to non-native English writers?
GPTZero says it has trained its model to cut false positives on English-learner writing to about 1%, but that is a vendor claim. A 2023 Stanford study of seven detectors, GPTZero among them, found they misclassified over half of TOEFL essays as AI on average. Plagino has not run its own test on non-native writing, so treat a flag on such writing with extra care.
Does GPTZero check for plagiarism?
Yes, GPTZero sells a separate plagiarism checker alongside its AI detector. The plagiarism checker looks for copied sources, while the AI detector scores writing patterns. Plagino's plagiarism checker matches against open-access research through CORE and does not search the public web.
What does a mixed result mean in GPTZero?
Mixed means GPTZero thinks a document contains both human and AI writing. GPTZero says it offers this third class and can separate human text that was polished by AI. In Plagino's benchmark, a document counted as flagged if a tool flagged any part of it.
What is a good GPTZero alternative for students?
It depends on the job. Plagino suits students who want a PDF report with sentence-level highlights and matched sources, and its benchmark false-flag rate was lower than GPTZero's. Copyleaks, Originality.ai and Winston AI also have comparison pages on this site, and none of these tools can promise to match your institution's checker.
Who owns GPTZero?
GPTZero is now part of Superhuman, which Wikipedia reports acquired it in June 2026. Edward Tian launched GPTZero in January 2023. The company says it has served more than 17 million users.
