Short answer: GPTZero is one of the stronger AI detectors in published tests, and it does well on clearly AI-written text. It is not error-free: independent studies have caught it flagging human writing, and one 2025 study found it far more likely to wrongly label AI-polished writing by non-native English authors as fully AI-written (PeerJ Computer Science). Treat its score as a signal, compare it with a second tool such as our free AI detector, and never use any score as proof on its own.

Below: GPTZero's own claims, its benchmark entry, independent studies, and how to use its result fairly.

What GPTZero claims about its accuracy

GPTZero is an AI detector built around classrooms. You paste text or upload a file, and it estimates how much of it is AI-generated, with sentence-level highlights. Its free plan covers up to 10,000 words a month (GPTZero pricing). In June 2026, Superhuman, the company formerly known as Grammarly, announced it was acquiring GPTZero (BetaKit).

GPTZero's homepage advertises 99.9% accuracy. Its FAQ says that on the RAID benchmark it detected 95.7% of AI texts while flagging 1% of human texts as AI, and that it reaches 96.5% accuracy on "mixed" documents that combine human and AI writing (GPTZero). It also says it trains its model to keep the false positive rate for ESL (English as a second language) writing at 1% (GPTZero).

Its technology page adds that its results come in confidence categories ("uncertain", "moderately confident" and "highly confident"), tuned so the average error rate is under 1% for highly confident predictions on its internal evaluation set (GPTZero technology).

GPTZero is also clear about limits. Its FAQ says no AI detector is 100% accurate, that results "should not be used to punish or as the final verdict", and that accuracy is better on longer documents and on English prose (GPTZero). Its technology page adds that short texts and heavily edited passages make detection harder (GPTZero technology).

What the RAID benchmark actually shows

RAID is an academic benchmark from the University of Pennsylvania, built from more than 6 million AI generations across 11 models, 8 domains and 11 adversarial attacks (Dugan et al., RAID, arXiv). Detector makers run the test set themselves and submit their predictions, which RAID then scores.

GPTZero's submission, dated October 2025, is public. On the texts without attacks, it caught about 95.7% of AI texts at a 1% false-alarm threshold, which matches its FAQ. Across the full benchmark, including attacks such as paraphrasing and character tricks, the figure is about 92.9% at the same threshold and about 97.2% when 5% of human texts may be flagged (RAID results file for GPTZero).

Those are solid results, but a benchmark measures how much AI text is caught at a fixed false-alarm rate on RAID's own texts. It does not tell you how often GPTZero will be right about your student's essay or cover letter.

What independent tests found

PeerJ Computer Science study (2025)

The most useful recent study compared GPTZero, ZeroGPT and DetectGPT on 72 journal abstracts written before ChatGPT and AI-written versions from ChatGPT o1 and Gemini 2.0 Pro. On this clear-cut task, GPTZero reached 97.22% accuracy with no human abstracts flagged (Pratama, PeerJ Computer Science, 2025).

The second experiment is the one teachers and editors should read. When the human abstracts were only polished by an AI model for readability, GPTZero labelled 25% of the non-native authors' abstracts as fully AI-generated, against 11% for native authors. The author concluded that tools with higher overall accuracy, such as GPTZero, showed stronger bias against non-native authors in this setting (Pratama, 2025). For students and researchers in India who write in English as a second language and use grammar tools, that matters.

University of Chicago working paper (2025)

Economists at the Becker Friedman Institute tested Pangram, Originality.ai, GPTZero and an open-source RoBERTa model on a large corpus of human and AI text. They found that most commercial detectors "perform remarkably well" on their set, but that on very short passages and rewritten text the performance of detectors other than Pangram "becomes case-dependent" (Jabarian and Imas, BFI).

Older studies (2023)

  • A Stanford study ran essays by non-native English writers through seven detectors, GPTZero among them. On average the detectors flagged 61.22% of these human-written essays as AI (Liang et al., 2023). That average covers all seven tools as they were in 2023, not GPTZero today.
  • A review of 14 detection tools, GPTZero included, found none of them reliable enough on its own, with all below 80% accuracy on the test set (Weber-Wulff et al., 2023).
  • In April 2023 a viral screenshot showed GPTZero calling part of the US Constitution "likely to be written entirely by AI" (Ars Technica). Famous text is highly predictable to language models, which is what many detectors read as AI.

GPTZero has changed a lot since 2023, so these show the kind of mistakes detectors make, not how often GPTZero makes them now.

A journalist's hands-on test (2025)

ZDNet's David Gewirtz runs the same five texts through detectors every few months. In October 2025 GPTZero got four of five right, but which one it missed changed between rounds: in April it got his own writing wrong, later it got a ChatGPT text wrong (ZDNet). Five texts is far too few to measure accuracy, but results clearly shift between updates.

GPTZero's own claims next to what the RAID benchmark and independent studies found

Claims and tests side by side

Source What was tested Main finding
GPTZero homepage and FAQ GPTZero's own testing Very high accuracy and a low false positive rate, including for ESL writing (GPTZero)
RAID, Oct 2025 submission Public benchmark, run by GPTZero Strong detection at a fixed false-alarm rate; lower with attacks (RAID)
Pratama, PeerJ 2025 72 abstracts plus AI and AI-polished versions Near-perfect on clear cases; AI-polished non-native writing flagged far more often (PeerJ)
Jabarian and Imas, 2025 4 detectors, large corpus Commercial tools strong overall; short and rewritten text harder (BFI)
ZDNet, Oct 2025 5 texts Four of five right; the miss changed between rounds (ZDNet)

The pattern: GPTZero is strong on fully human vs fully AI text, and weaker, like every detector, on short, edited or AI-polished human writing.

Why GPTZero flags human writing

A false positive is human writing labelled as AI. For GPTZero, as for other detectors, the usual causes are:

  1. AI-polished drafts. A human draft run through a grammar or rewriting tool takes on the smooth style detectors associate with AI. This is where the PeerJ study saw the biggest gap between native and non-native authors (PeerJ).
  2. Short text. A paragraph gives little evidence. GPTZero itself says document-level results are stronger than paragraph or sentence results (GPTZero).
  3. Formulaic writing. Definitions, standard introductions, legal text and famous passages are predictable by nature.
  4. Mixed documents. A text that is partly human and partly AI is hard to label as a whole, which is why GPTZero reports it as a separate category.

How to use a GPTZero score fairly

If a GPTZero result will affect a grade, a job or a publication:

  1. Check the whole document, not one paragraph, and note GPTZero's confidence label, not only the percentage.
  2. Read the highlighted sentences. Are they the writer's own argument, or references and standard phrases?
  3. Ask whether AI polishing was allowed. If your rules permit grammar tools, an "AI" result on a polished draft may be within the rules.
  4. Run a second detector. Strong disagreement between two tools is itself a reason to slow down. Our reviews of ZeroGPT and QuillBot's detector show how differently tools behave.
  5. Look at drafts and version history, and compare with the writer's earlier work.
  6. Talk to the writer before deciding anything. GPTZero itself calls its tool a conversation starter (GPTZero technology).

Checking with ProofAI

ProofAI is an AI detector made in India, priced in rupees. Paste text into the AI text detector or upload a PDF, DOCX, DOC or TXT file up to 25 MB in the document checker, and you get an AI probability with the passages that drove it. It is free for 10 checks a day without logging in and 25 with a free account, counted per check rather than per word. It also checks images; video detection is coming soon.

ProofAI has the same limits as every detector: it can be wrong, especially on short, formulaic or AI-polished text, so use it as a second opinion, not a verdict. We have not run GPTZero and ProofAI on a shared test set, so we publish no head-to-head figure. For plans, limits and when GPTZero is the better pick, see GPTZero alternative; for how we test our own detector, see the accuracy page.

FAQ

Is GPTZero accurate?
It is among the stronger detectors in published tests. A 2025 peer-reviewed study found it near-perfect at separating original human abstracts from AI-written ones, but much more likely to flag AI-polished writing by non-native authors as fully AI (PeerJ). It is accurate enough to be a useful signal, not to be proof.

Does GPTZero give false positives?
Yes. GPTZero says no detector is 100% accurate (GPTZero), and independent tests have recorded human writing labelled as AI, from the US Constitution in 2023 (Ars Technica) to AI-polished academic abstracts in 2025.

Is GPTZero's 99.9% accuracy claim true?
It is GPTZero's own figure from its own testing (GPTZero). Its public RAID submission shows strong but lower figures once adversarial attacks are included (RAID). Neither tells you how it will do on your kind of writing.

Is GPTZero biased against non-native English writers?
GPTZero says it trains its model to keep ESL false positives low (GPTZero). One 2025 study still found a large gap when non-native authors polished their own writing with AI (PeerJ). If you write in English as a second language, keep your drafts.

Can a teacher fail a student because of a GPTZero score?
A score alone should not decide that, and GPTZero itself says results should not be used to punish or as the final verdict (GPTZero). Use it to start a conversation, and look at drafts and earlier work.

Conclusion

GPTZero is a capable detector with a strong benchmark record and an honest limits section. But published tests show where it, and every detector, struggles: short text, edited text and AI-polished writing by second-language authors. Use its score to decide where to look, check with a second tool when it matters, and let drafts and a conversation with the writer decide.

Originally published on ProofAI.