How ZeroGPT works, why it flags human writing, and what to do about it

What ZeroGPT is, and what it is not

ZeroGPT is run by Olive Works LLC, a company registered in Wyoming in the United States, and the site has been online since January 2023. The detector is free to use without paying, and the same site sells a paraphraser, a summariser, a chatbot, image and video detectors, WhatsApp and Telegram bots, an API, and its own AI humanizer, which its homepage advertises as "Rewrite to pass review".

Three things are easy to confuse with it. GPTZero (gptzero.me) is a different detector made by a different company; we cover it in how GPTZero works. The famous April 2023 screenshot of the US Constitution being flagged as AI came from GPTZero, not ZeroGPT, although ZeroGPT flagged it too when Ars Technica tested it. And zerogpt.net, zerogpt.cc and the browser extensions that carry the ZeroGPT name are not made by zerogpt.com, which links to no extension of its own.

How ZeroGPT says it works

ZeroGPT calls its method DeepAnalyse and describes four stages: macro analysis of the whole text, sentence analysis, micro analysis, and a final assessment. Its FAQ lists the signals as "token patterns, burstiness, entropy, and ensemble classifier features trained on mixed datasets", and says the model was trained on text from the internet, educational datasets and its own synthetic AI text. The model itself is not published: "Proprietary internal model weights and implementation details are not shown."

Burstiness and entropy are versions of the ideas most detectors share: AI text tends to be evenly predictable, while people vary their rhythm and word choice. How AI detectors work explains them without the jargon.

ZeroGPT states its accuracy three different ways on its own site: "98.4% Detection accuracy" and a "<1% False positive rate" on its DeepAnalyse page, "up to 98%" with an error rate "lower than 2%" on its homepage, and "pushing toward >98% on internal evaluations" in its FAQ. We found no published method or independent audit behind any of these figures.

How to read a ZeroGPT result

A result is a sentence, a percentage labelled "AI GPT" and highlighted sentences that it suspects were generated. The sentence is one of nine, in this order from most human to most AI:

• "Your text is Human written"
• "Your text is Most Likely Human written"
• "Your text is Most Likely Human written, may include parts generated by AI/GPT"
• "Your text is Likely Human written, may include parts generated by AI/GPT"
• "Your text contains mixed signals, with some parts generated by AI/GPT"
• "Your text is Likely generated by AI/GPT"
• "Your text is Most Likely AI/GPT generated"
• "Most of Your text is AI/GPT Generated"
• "Your text is AI/GPT Generated"

Read the verdict and the number together

The two do not always agree. A 2023 study of 16 detectors noted that ZeroGPT's "text descriptions do not always correspond to specific numeric values", and a 2025 study of research abstracts found ZeroGPT flagging some human abstracts as AI even though their scores were below 30%. If you are shown only one of the two, ask for both, and for the highlighted text.

Length matters too. ZeroGPT sets no hard minimum, but it advises at least 150 to 200 words and says that 500 to 1,000 words or more "improves stability". A score on one paragraph tells you very little. The free version checks up to 15,000 characters at a time, about 2,500 words.

Why ZeroGPT flags human writing

ZeroGPT's best-known mistakes are famous documents. In July 2023 Ars Technica ran part of the US Constitution through it and got "Your text is AI/GPT Generated", and a passage from the Book of Genesis came back as 88.2% AI-generated. In October 2024 Decrypt reported that the preamble of the Declaration of Independence scored 97.75% AI for one tester and 97.93% in its own re-run. In May 2025 the Dallas Express reported that the 1836 Texas Declaration of Independence scored 86.54%.

The likely reason is the one Ars Technica gave: text that appears everywhere online is extremely familiar to a model trained on the internet, so it reads as predictable, and predictable is what these detectors call AI. ZeroGPT's own FAQ admits the wider version of the problem: "Highly polished, formulaic, or low-entropy writing can resemble AI." That describes a lot of careful student and professional writing.

Independent tests point the same way. The 2024 RAID benchmark could not get ZeroGPT's false-positive rate below 16.9%, and at the thresholds it tested ZeroGPT flagged 48% of human Wikipedia text and 52% of human-written recipes. A 2023 study in the International Journal for Educational Integrity found ZeroGPT correct on all nine human-written English texts but on only five of nine human texts that had been machine-translated into English. A 2023 Stanford study that included ZeroGPT found seven detectors labelling, on average, 61% of TOEFL essays by non-native English speakers as AI. Most of these studies tested older versions; ZeroGPT says it retrains regularly.

Our own weekly benchmark uses a different detector and shows the same failure on the same kind of text: well-known passages written long before AI, scored as likely AI.

If someone flags your work with ZeroGPT

Start with ZeroGPT's own words, because they are clear. Its terms of use say that "a ZeroGPT detection score does not constitute definitive proof" and that "you must not treat a ZeroGPT Output as the sole basis for a decision that could materially affect another person's rights, education, employment". The instructions for educators in its results panel say: "These outcomes should not be utilized to penalize students."

• Ask for the exact text that was checked, the verdict sentence, the percentage and the highlighting. Check that it was ZeroGPT (zerogpt.com) and not a site with a similar name.
• Point to the published record: the same tool has labelled the US Constitution and the Declaration of Independence as AI, which shows what a flag on its own is worth.
• Show how you wrote it. Version history in Google Docs or Word, outlines, notes, sources and earlier drafts are far stronger evidence than any detector score.
• If you wrote in another language and translated, say so. Machine-translated human text is one of the cases where ZeroGPT has been shown to struggle, and our guide for non-native English writers covers the wider problem.
• Offer to talk through the content: what you argued, why you chose your sources, what you would change. Then follow your institution's or your client's process; if you are accused of using AI sets out the steps.

Where AntiGPT fits, honestly

We do not run ZeroGPT, and we publish no figures for it; its terms also forbid using its results to benchmark a competing product without permission. AntiGPT's detector is a different, open model, so its score will not match ZeroGPT's. Anyone can check text on ZeroGPT for free, and if a reader of yours uses it, it is worth knowing how your text reads to it.

What AntiGPT is for is the writing. If you draft with AI help where that is allowed, it rewrites the draft so it reads like you, shows an estimated score before and after, and lets you redo single sentences that still sound like a template. It cannot make any detector certain to agree, and no tool can. The free readability checker and sentence variety checker show two of the properties detectors react to, in your own text, without an account.