AI detectors

GPTZero: how accurate it is, what it flags, and what we found testing it on student writing

By ESL Humanizer editorial team · Updated October 2026

Short answer

GPTZero is one of the most widely used AI detectors and is free for scans of up to 10,000 characters. It claims 99% accuracy and says it is de-biased for non-native writers. In our October 2026 test it classified all 299 scans of pre-ChatGPT student writing as human, English learners included, and caught all 30 AI-written controls. But when learner texts were fully polished by AI, all 19 were labelled mixed, while grammar-only corrections stayed human. A GPTZero score is still an estimate, not proof.

What GPTZero is

An AI detector that estimates how much of a text is likely AI-written and highlights the sentences it attributes to AI. It also flags documents that mix human and AI writing.

Checks for AI writing	Yes
Where it’s used	Teachers and schools, often through browser and LMS integrations, and students checking their own work
Free access	Free scans of up to 10,000 characters; longer scans need a free account
Paid	Monthly or annual plans with higher word limits and extra features — see GPTZero's pricing page

How to use GPTZero

Open gptzero.me and paste your text, up to 10,000 characters without an account.
Run the scan. GPTZero estimates how much of the text is likely AI-written and highlights the sentences that drove the score.
Read a “mixed” result carefully: it means GPTZero thinks part of the text was AI-written or AI-edited, which is what fully AI-polished writing triggered in our test.
Don't confuse GPTZero with ZeroGPT — they are different companies and often disagree.

Want to see which of your own sentences read as AI-like, and why? Our free AI writing checker highlights the phrasing detectors react to, in your browser, without storing your text.

What GPTZero says about itself

GPTZero claims a 99% accuracy rate, reports detecting 95.7% of AI texts while flagging 1% of human texts on the RAID benchmark, and calls itself “the only AI detector de-biased for ESL (English Second Language) learners,” saying it trains its model to keep the false positive rate on ESL writing to about 1%. See GPTZero’s website for current details.

What published tests of GPTZero found

We tested GPTZero ourselves in October 2026 on 386 scans: essays by English learners and native speakers from the PERSUADE 2.0 corpus, short learner passages from English Language Learners Stack Exchange written before ChatGPT, AI-written controls, and learner texts edited in different ways. It flagged none of the 299 human scans, caught all 30 AI controls, kept grammar-only corrections human, and labelled all 19 fully AI-polished learner texts as mixed. In the independent 2025 Chicago Booth study, GPTZero also kept false positives low. Caveat: public essays may be familiar to detectors, so new writing could score differently.

What independent research says

Vendor figures come from the vendor’s own tests. Independent research paints a more cautious picture of AI detection in general:

In a peer-reviewed Stanford study, seven detectors misclassified an average of 61% of human-written TOEFL essays as AI-generated, while essays by US eighth-graders were classified almost perfectly (Liang et al., 2023).
An independent test of 14 detection tools found none above 80% accuracy (Weber-Wulff et al., 2023).
In our own October 2026 test, GPTZero flagged none of 299 scans of pre-ChatGPT student writing, learners included — but labelled all 19 learner texts that had been fully polished by AI as AI-assisted (our GPTZero study). Only GPTZero was tested, so other detectors may behave differently.
OpenAI withdrew its own AI classifier in 2023 for low accuracy (OpenAI).

What non-native writers should know

Our test supports GPTZero's claim that learners' own writing is rarely flagged: none of the learner texts we scanned were.
What changed the result was AI polishing. Letting a chatbot rewrite every sentence turned human learner texts into “mixed”; fixing only grammar didn't.
Short texts give any detector less to go on, so check a whole essay rather than one paragraph.

If GPTZero flagged your work

Check whether you used an AI tool to rewrite whole sentences — in our test, that is what produced “mixed” labels.
Ask which sentences were highlighted, and share your drafts and version history.
Point out that GPTZero itself says no detector is 100% accurate and that a score shouldn't be the final verdict.

Need to reply in writing? Use our appeal letter template.

Wrote it yourself? Make it sound natural.

Paste a paragraph. ESL Humanizer fixes non-native phrasing, keeps your meaning and highlights every change — free for up to ~250 words.

Try it free

Quick answers

Is GPTZero accurate?

GPTZero claims 99% accuracy. In our own test of 386 scans it classified every human text correctly, including English learners' writing, and caught every AI-written control. Its weak spot for students is AI-edited writing: fully AI-polished versions of learners' texts were labelled mixed.

Is GPTZero free?

Yes, for scans of up to 10,000 characters. Longer scans need a free account, and paid plans raise the word limits.

Does GPTZero flag non-native English speakers?

GPTZero says it is de-biased for ESL writers. In our test it flagged none of the learner texts we scanned, but the 2023 Stanford study found detectors in general flagged 61% of TOEFL essays, so keep evidence of your writing process anyway.

What does “mixed” mean in GPTZero?

It means GPTZero thinks the document combines human and AI writing. In our test, learner texts that a chatbot had fully polished were labelled mixed, even though a person wrote the original.

Is GPTZero the same as ZeroGPT?

No. They are separate products from different companies, and they often give different scores on the same text.