Is ZeroGPT Accurate? What Published Research Shows

ZeroGPT is one of the most popular free AI detectors around. If you're reading this, there's a decent chance it just told you something you wrote yourself is AI generated. So is ZeroGPT accurate? Sometimes it's very good. Sometimes it's badly wrong. And the published research shows its results swing a lot depending on the kind of writing you give it.

What ZeroGPT is and how it says it works

ZeroGPT is a web-based detector. You paste in text and it returns a percentage score, with the sentences it thinks are AI-written highlighted. On its homepage, the company advertises 98.4% detection accuracy, a false positive rate under 1%, and coverage of more than 35 LLM families. It credits a proprietary system it calls DeepAnalyse and says the model was trained on internet text, educational datasets and its own synthetic AI datasets.

The homepage doesn't explain how those accuracy figures were measured. Its FAQ says the tool analyzes linguistic and statistical signals to estimate how likely it is that text came from an LLM. To its credit, the FAQ also admits that highly polished, formulaic or "low-entropy" human writing can resemble AI, and it tells educators to use results as evidence within a fair review process.

What independent research found

ZeroGPT has been included in several published evaluations. They don't agree with each other, and that disagreement is the most useful finding of all.

The RAID benchmark (ACL 2024)

RAID, presented at ACL 2024 by Liam Dugan and colleagues, is one of the largest detector tests published so far: over 6 million generations from 11 language models across 8 domains, run against 12 detectors including ZeroGPT. The authors' overall conclusion was that current detectors are easily fooled by adversarial attacks, changes in sampling and models they haven't seen before.

The ZeroGPT results that matter most here are about human text. To compare detectors fairly, the researchers tried to tune each one so it flagged no more than 5% of human-written text. According to the full paper, ZeroGPT couldn't reach that target in several domains. Even at its strictest setting in that calibration step, it still flagged 29% of human-written news articles, 48% of Wikipedia text and 52% of recipes. Across domains it plateaued at a 16.9% false positive rate. On Reddit posts, scientific abstracts, reviews and poetry, it stayed at or under 5%.

The same paper also reports that at ZeroGPT's default cutoffs, its false positive rate across the whole dataset was under 2%. So the average looks fine. The problem is the spread: some kinds of perfectly human writing get flagged far more than others.

Weber-Wulff and colleagues (2023)

A team of academic integrity researchers tested 14 detection tools, including ZeroGPT, Turnitin and GPTZero, on 54 test documents, and published the results in the International Journal for Educational Integrity. Their verdict was blunt: the tools are "neither accurate nor reliable." ZeroGPT's overall accuracy landed somewhere between 59% and 74% depending on how partially correct answers were scored.

On human writing, though, ZeroGPT did well in this study. It produced no false accusations on the 18 human-written documents, which included texts written in other languages and machine-translated into English. Its weakness here was missing AI text, not flagging people.

Physics essays (Yeadon and colleagues, 2024)

In a study of 300 short physics essays, half written by students and half by GPT-4, five commercial detectors were compared. ZeroGPT came out on top with 98% accuracy and a precision of 1.0, which means none of the essays it flagged as AI were actually human.

Medical abstracts (Cooperman and Brandão, 2024)

A small study in Foot & Ankle Surgery: Techniques, Reports & Cases found the opposite. Six detectors each checked 12 abstracts, six written by humans and six by AI. ZeroGPT had the highest false positive rate of the group, labelling 5 of the 6 human-written abstracts as AI. That's a tiny sample, so don't lean on the exact percentage. But it's a clear example of formal scientific prose tripping the tool.

The Declaration of Independence

The most widely shared example came in October 2024, when Decrypt ran the Declaration of Independence through four detectors. ZeroGPT said it was 97.93% AI-generated. QuillBot's detector called it 100% human, and GPTZero gave it an 89% probability of being human. A document from 1776 obviously wasn't written by a language model.

The results side by side

So is ZeroGPT accurate? It depends on the writing. On some kinds of text, independent tests come close to the company's own numbers. On formal, encyclopedic, scientific or tightly structured writing, the evidence shows it can flag human work at high rates. No single figure, the company's or anyone else's, tells you how it will treat your essay.

Why detectors like ZeroGPT misfire on human writing

Most AI detectors lean on some measure of predictability. Language models write by picking likely next words, so AI text tends to have low perplexity: each word is easy to guess from the ones before it. Detectors treat low perplexity as a sign of AI.

The trouble is that plenty of human writing is predictable too, for good reasons:

• Formal and academic writing uses standard phrasing, set structures and field-specific vocabulary. That's what students are taught to do.
• Reference and news writing follows house styles that flatten individual quirks. News and Wikipedia were two of RAID's worst domains for ZeroGPT.
• Formulaic genres like recipes repeat the same patterns over and over. Recipes were RAID's single worst domain for ZeroGPT.
• Non-native English writing often uses a smaller, safer vocabulary and simpler sentences.

That last point has the strongest evidence behind it. In a Stanford study published in Patterns in July 2023, Weixin Liang and colleagues ran seven popular detectors, ZeroGPT among them, on 91 TOEFL essays by non-native English speakers and 88 essays by US eighth graders. The detectors were near-perfect on the US essays. On the TOEFL essays, the average false positive rate across detectors was about 61%. Of the 91 essays, 89 were flagged as AI by at least one detector, and 18 were flagged by all seven.

The researchers then tested why. When they used ChatGPT to make the TOEFL essays' word choices sound more like a native speaker's, the average false positive rate dropped to about 12%. When they simplified the US essays' vocabulary to mimic non-native writing, it jumped from about 5% to about 57%. Same ideas, different word choices, very different verdicts. The essays that every detector flagged had significantly lower perplexity.

Those are averages across seven tools, not ZeroGPT's own numbers. But ZeroGPT's FAQ concedes the same weakness when it says low-entropy writing can look like AI. GPTZero has a similar problem, which we looked at in how accurate is GPTZero.

What to do if ZeroGPT flagged your writing

1. Don't panic, and don't rewrite in a hurry

A score is not a finding. ZeroGPT itself frames its output as a likelihood and tells educators to weigh it within a fair review. Rewriting your essay in a panic won't help your case.

2. Gather evidence of your process

Your strongest defence is showing how the work was made. Pull together:

• Version history from your editor, if it keeps one
• Earlier drafts, outlines and handwritten notes
• The sources you read, with the dates you accessed them if you have them
• Older writing of yours that shows the same voice and habits

Going forward, write somewhere that saves version history. It's free insurance.

3. Talk to your instructor early

Ask calmly which tool was used, what the score was and which passages were highlighted. Offer to walk through your drafts or explain your argument in person. Being able to talk fluently about your own work is persuasive in a way a percentage isn't. If English is your second language, it's fair to point to the Stanford study and explain why careful, formal second-language writing gets misread.

4. Use several signals, not one score

Detectors often disagree, as the Declaration test showed, so one tool's verdict shouldn't settle anything. You can check the text with another tool, such as our free AI detector, to see whether the results line up. Treat every detector, ours included, as one signal with known limits, not proof either way.

5. Know your school's policy

Read your institution's academic integrity and AI policy so you know how flags are handled and what appeal rights you have. Our essay flagged as AI recovery guide walks through the conversation and the appeal step by step.

And if you do use AI tools where your school or employer allows it, WriteHumanly's humanizer is built to make an AI-assisted draft read naturally, in your own voice, while keeping every fact intact. It's a writing aid, not a way around a policy, and no tool can promise what a detector will say.

Frequently asked questions

Is ZeroGPT accurate?

It depends on the text. ZeroGPT claims 98.4% accuracy and under 1% false positives, and one study of physics essays found no false positives at all. But the RAID benchmark found it flagged 29% to 52% of human-written news, Wikipedia and recipe text at its strictest setting, and a small medical study saw it flag 5 of 6 human-written abstracts.

Why does ZeroGPT say my writing is AI when I wrote it?

Detectors look for predictable wording. Formal, academic, formulaic and non-native English writing is often predictable for perfectly human reasons, so it can score like AI. ZeroGPT's own FAQ acknowledges that polished or low-entropy writing can resemble AI.

Is ZeroGPT biased against non-native English speakers?

A 2023 Stanford study tested seven detectors, ZeroGPT included, and found they flagged about 61% of non-native speakers' TOEFL essays as AI on average, while getting US eighth graders' essays nearly all right. That figure is an average across tools, not ZeroGPT alone, but it points to a real risk.

Is a ZeroGPT score enough to prove someone used AI?

No. ZeroGPT presents its result as a likelihood and tells educators to use it within a fair review process. Detectors disagree and misfire on some kinds of human writing, so a score should start a conversation, not end one.