Is ZeroGPT Accurate? What Independent Tests Show (2026)

Not as accurate as its homepage suggests. ZeroGPT advertises 98.4% detection accuracy and a false-positive rate under 1% (ZeroGPT), but it doesn't explain how it measured those figures. Independent evaluations that included it found much more uneven results: 59% of documents classified correctly in a 2023 peer-reviewed test, and anywhere from 95.4% down to 0.3% of AI text caught in the RAID benchmark, depending on how the text was generated. It can give you a quick first read on long, unedited English prose, but a ZeroGPT score is not proof that anyone used AI.

Below: what ZeroGPT claims, what each independent evaluation actually measured, public examples of false positives, what causes them, and a simple way to sanity-check any detector yourself.

What ZeroGPT claims (checked September 2026)

ZeroGPT's homepage lists "98.4%" detection accuracy, a false-positive rate shown as under 1%, "20+ LLM families covered", and support for "all the languages with the highest accuracy rate of detection." It credits "DeepAnalyse™ Technology," a "multi-stage methodology" trained on "extensive text collections from the internet, educational datasets, and our proprietary synthetic AI datasets." Results highlight "every sentence written by AI" next to a gauge showing the percentage of AI in the text. The free checker takes up to 15,000 characters.

The FAQ is a little more careful. It says ZeroGPT is "pushing toward >98% on internal evaluations," and it concedes that "highly polished, formulaic, or low-entropy writing can resemble AI."

Neither page says which texts, which AI models, what length or what threshold produced those numbers, so nobody outside the company can check them. That doesn't make them false, but it does make them vendor claims rather than evidence. The authors of RAID, the largest public detector benchmark, open their paper with the same warning: many detectors "claim to detect machine-generated text with extremely high accuracy (99% or more)," yet their tests found "current detectors are easily fooled by adversarial attacks, variations in sampling strategies, repetition penalties, and unseen generative models" (Dugan et al., 2024).

What independent evaluations found

These are the evaluations we could open that report a result for ZeroGPT specifically.

A few things to keep in mind when you read that table:

• Most of it describes older versions. ZeroGPT's homepage now lists models such as GPT-5 that didn't exist when the academic studies ran, so those results describe earlier versions of the tool.
• The famous "20.51%" figure is a vendor test. You'll see it repeated on many sites. It comes from a blog post written by the CEO of DecEptioner, which sells AI humanizing tools. The page doesn't name the AI models or text sources it used, and it notes that "results can vary with different topics, writing styles, and newer model outputs." It's a data point, not a study.
• The failure mode varies. In the two academic tests that counted false accusations, ZeroGPT made none but missed a lot of AI text. RAID shows the misses cluster in particular ways of generating text.

ZeroGPT was also one of the seven detectors in Liang et al. (2023), the Stanford study in which the detectors, on average, labeled 61.22% of 91 human-written TOEFL essays as AI. The paper reports that average rather than a separate figure for each tool, so we don't quote one for ZeroGPT.

How ZeroGPT compares on the same benchmark

RAID is the only evaluation here that ran ZeroGPT alongside other commercial detectors on identical text, with each detector's threshold set so that it wrongly flagged 5% of RAID's human text (Dugan et al., Tables 4, 5 and 7).

What stands out:

• How text is generated changes everything. On the same chat models, ZeroGPT caught 92.1% of text generated one way and 65.8% of text generated another way. The authors found that "random sampling with a repetition penalty consistently makes output generations very difficult to detect."
• "All languages" isn't the same as accurate in all languages. On Czech and German news, ZeroGPT caught about half the AI text at the 5% false-positive target.
• False positives were low, but the highest of the four. At a plain 0.5 cutoff, ZeroGPT wrongly flagged 1.42% of RAID's human text. The authors note the closed-source detectors "seem to be calibrated fairly well, with none having an FPR above 1.7%."
• One caveat. RAID marks ZeroGPT as "unable to achieve the target FPR," so its row isn't a perfect like-for-like comparison. And all four tools have been updated since 2024.

Public examples of ZeroGPT false positives

The best-documented example comes from Ars Technica on 14 July 2023, which showed ZeroGPT flagging a section of the Book of Genesis as "88.2 percent AI-generated." The same article showed GPTZero calling part of the US Constitution "likely to be written entirely by AI."

The explanation applies to ZeroGPT too. GPTZero's creator Edward Tian told Ars that the Constitution "is a text fed repeatedly into the training data of many large language models," so detectors that look for text a language model would likely produce end up flagging it. Famous, heavily quoted text looks predictable to a model that has seen it many times.

You'll also find plenty of screenshots of ZeroGPT scoring famous documents as AI on social media. We only cite cases a news outlet documented, because a screenshot can't show which version was used or what text was pasted.

For context, even OpenAI couldn't make this work reliably. Its own AI-text classifier correctly identified 26% of AI-written text while labeling human text as AI 9% of the time, and OpenAI withdrew it on 20 July 2023 "due to its low rate of accuracy" (OpenAI).

What drives false positives

Detectors estimate how predictable a text's wording is. They don't know who wrote it. That puts several kinds of honest writing at risk:

• Polished or formulaic prose. ZeroGPT's own FAQ names "highly polished, formulaic, or low-entropy writing." Lab reports, legal summaries and textbook-style definitions fit that description.
• Short text. ZeroGPT recommends "at least 150–200 words" and says "longer text (500–1,000+) improves stability." A single paragraph gives any detector very little to go on.
• Non-native English. In Liang et al., simplifying the vocabulary of US students' essays raised the average false-positive rate from 5.19% to 56.65%. Plain, careful English can look "predictable." We cover the debate in AI detection and ESL students.
• Well-known text. Quotations, scripture, legal boilerplate and anything widely copied online, as the Ars examples show.
• Machine translation. Weber-Wulff's team found that false-positive risk across the 14 tools rose sharply for machine-translated human writing.
• Editing. A 2026 study of 135,389 manuscripts found that professional editing pushed some detectors' scores toward "AI" and others toward "human," with false-positive rates from 0.0% to 100.0% across 13 research detectors (Park et al.). ZeroGPT wasn't one of them, but the mechanism is general.

More cases, with sources, are in our explainer on AI detection false positives.

Free vs. paid: does upgrading help?

ZeroGPT's paid plans start at $13.99 a month, or $9.99 a month billed annually (list prices on its pricing page, checked September 2026). They raise your limits and add batch uploads.

What they don't come with is any documented accuracy difference. ZeroGPT's pricing page doesn't say paid plans use a different or better model, and none of the evaluations above compared free and paid results. So if you pay, you're paying for capacity and workflow, not for accuracy anyone has shown. If you need a second opinion, a detector from a different vendor tells you more than an upgrade.

How to sanity-check any AI detector yourself

You don't need a lab to learn how a detector treats your kind of writing. You need a small set of texts where you already know the answer.

1. Collect human samples you're sure of. Five pieces you wrote yourself long before chatbots were common, such as old essays or reports, in the same genre and language as the work you're worried about.
2. Add AI samples on the same topics. Five drafts from the chatbot you'd expect, plus five of those drafts after you've edited them by hand.
3. Match the real length. Test pieces as long as the document you'll actually submit. A detector can behave very differently on one paragraph than on a full paper.
4. Run each text twice. If the same text gets noticeably different scores on two runs, you've learned something about how much weight any single score deserves.
5. Count both kinds of mistake. How many of your human samples were flagged? How many AI samples slipped through? One false flag out of five human samples is a warning sign. Five clean results don't prove the tool is safe, because the sample is tiny.
6. Repeat with a second detector from a different vendor. Passages that both tools flag deserve a closer look. Disagreement means the result is uncertain.
7. Keep screenshots with dates and tool names. If a score is ever used against you, a record of how the same tool treated your older writing is useful context.

This works for any tool, including SupWriter's own AI detector, which gives one score from one model with sentence highlights. It isn't the detector your school uses, and like every detector it can be wrong (how we test).

When should you actually use ZeroGPT?

ZeroGPT is reasonable for:

• Quick spot checks on long English text you already have questions about
• Screening large volumes of content where a person will review everything that gets flagged
• A second opinion alongside a detector from a different vendor

ZeroGPT is not reliable for:

• Academic integrity decisions. No evaluation above supports using a single score that way, and ZeroGPT publishes no method for its accuracy figure.
• Short text. ZeroGPT itself recommends at least 150–200 words.
• Non-English text. "All languages" is a support claim, not an accuracy claim. In RAID, ZeroGPT caught only about half the AI text in Czech and German news.
• Proof. A percentage is a model's estimate, not evidence of how a text was written.

The bigger picture

ZeroGPT isn't uniquely bad. In the academic tests above it avoided false accusations better than some rivals, and in RAID it performed close to the others on some kinds of text. The problem is the gap between a precise-sounding "98.4%" and results that swing from excellent to useless depending on the text.

If a score could affect someone's grade, job or reputation, the question to ask isn't "what did the detector say?" It's "what evidence of how this was written do we have?" Drafts, version history and notes answer that. A detector score doesn't. If you're the one who was flagged, our guide on proving your writing is human walks through the evidence to gather.

FAQ

Is ZeroGPT really 98% accurate?

ZeroGPT says so, but it doesn't publish how the figure was measured. Its FAQ describes the target as "pushing toward >98% on internal evaluations." Independent tests found 59% of documents classified correctly in March 2023 (Weber-Wulff et al.) and results from 95.4% to 0.3% depending on the text in RAID.

Can ZeroGPT detect GPT-5 and Claude?

ZeroGPT's homepage lists GPT-5, GPT-6, Gemini, Grok, Claude, DeepSeek and LLaMa models among those it detects. We couldn't find an independent evaluation of ZeroGPT on those newest models. The published studies tested it on earlier ones, such as ChatGPT, GPT-4 and Claude 2.

Is ZeroGPT biased against non-native English speakers?

ZeroGPT was one of seven detectors in the Liang et al. study, where the detectors on average labeled 61.22% of human-written TOEFL essays as AI. That study used detector versions from early 2023, and later research is mixed; we lay out both sides in AI detection and ESL students. Treat any score on a non-native writer's work with extra caution.

Should I pay for ZeroGPT Plus?

Only if you need the higher limits or batch uploads. ZeroGPT doesn't claim the paid plans are more accurate, and no independent test has compared free and paid results.

Is ZeroGPT the same as GPTZero?

No. They are different companies with different models, and they often disagree. On RAID's human text at a plain 0.5 cutoff, GPTZero flagged 0.00% and ZeroGPT 1.42% (Dugan et al.). For GPTZero's own record, see is GPTZero accurate?

---

Check a draft with SupWriter's AI detector before you submit. It's a pre-check from one model, not proof either way. 300 words free, no credit card required.