ZeroGPT Review: How Accurate Is It Really? (2026)

On its homepage, ZeroGPT displays two numbers side by side: "98.4%" detection accuracy and a false positive rate under 1%. Also, in its terms of use, which were updated on 7 October 2026, it states: "No AI detector, including ZeroGPT, can determine authorship with absolute certainty in every case."

I read both pages on 3 November 2026, as well as the pricing page and the one head-to-head test that I found. I did not run any scans myself, so all of the results below are from the vendor themselves or another party that performed these tests.

My view: ZeroGPT is a free tool that is fine for a quick look at your own draft and a poor basis for accusing anyone of anything. Its own legal text agrees with that, and it is more careful than its homepage.

What ZeroGPT claims

The statistics section of the homepage reads "98.4% Detection accuracy", a false positive rate of below 1% and "35+ LLM families covered". The page states that the model was "trained on all languages and multiple LLM models" and that it supports "all the languages with the highest accuracy rate of detection". The models that the page recognizes include ChatGPT, Gemini, Grok and Claude.

The free checker accepts 15,000 characters per scan, according to the counter in the text box. The page states no word limit.

None of the accuracy figures link to a test, a dataset or a date. They are claims, not results, and I could not find a methodology page behind them.

What the terms of use say

Section 7 and section 23 of the terms are blunter than the homepage. A ZeroGPT score "does not constitute definitive proof" of who wrote a text. The terms list things ZeroGPT does not promise, including that "DETECTION RESULTS WILL ALWAYS BE CORRECT" and that "THE SERVICES WILL IDENTIFY ALL AI-GENERATED CONTENT." Outputs, they say, "SHOULD BE EVALUATED IN CONTEXT."

A sub-1% false positive rate and a disclaimer stating that the results will not always be correct can both be true. However, the first statement is used to sell the product while the second statement protects the company. If you are the individual whose essay was flagged by this AI tool, only the second statement is a promise to you.

One outside test, from a competitor

The only test that I was able to open was from GPTZero, a rival company, published on 20 July 2026 and modified on 30 July 2026. The test included 13 samples of texts: 6 that were written by AI (2 from each of Claude, ChatGPT and Gemini) and 7 human-written texts from prior to 2020. However, GPTZero did not state when the tests were performed.

The remaining four human texts (an academic paper, a Wikipedia article, a 1922 news item and a chapter of Moby Dick) scored between 0% and 4.2%.

Thirteen samples is a small test, and it comes from a company that sells a competing detector, so I would not quote it as proof of anything. What it does show is how far the percentage moves between texts: two Claude samples scored 56% and 29.9%.

A separate source puts ZeroGPT at 91% standardized accuracy in a 24-text study in the journal Information Research, published on 15 May 2026. I found that figure on Winston AI's help center, not in the paper, so treat it as secondhand. The same study scored Winston at 99%.

Pricing and refunds

The free plan costs $0 and needs no credit card. Personal and Business/EDU plans can be billed monthly or annually, with "40% off annually" advertised, and there is a custom enterprise tier and a pay-as-you-go API. The text of the pricing page I could read did not show the dollar amounts for the paid plans, so check the live page before you plan around a number.

The refund terms are the part to read before you pay. Section 14 of the terms of use allows a full refund "within ten (10) hours after the initial purchase", and 12 hours for annual plans. It is limited to one refund per customer, "Consumed usage is non-refundable" and "Renewal payments are non-refundable." Subscriptions renew automatically, and simply no longer using the service does not count as cancelling.

Put the calendar to work for you: if you subscribe, set a reminder to cancel straight away, since the renewal charge is the one you cannot get back.

Who it is for

ZeroGPT works as a free second opinion on text you wrote yourself. It is a weak fit for settling a dispute, for the same reason every single detector is: the score is a probability, and the sample test above shows how far it can swing. If a ZeroGPT result is the only thing standing between you and a grade, our guide to false positives in AI detectors covers what to bring to the conversation.

For the whole field in one place, Are AI detectors accurate? lists every published number, and the ZeroGPT bypass page explains how our humanizer handles text that ZeroGPT flags.

The short version

ZeroGPT is free, covers many models and gives a clear percentage. Its 98.4% accuracy and under-1% false positive claims come from ZeroGPT alone, its terms say results will not always be correct, and the one outside test I found shows wide swings. Use it on your own drafts, never as proof against someone else, and read the 10-hour refund window before paying.

If you want the longer treatment:

• Winston AI review
• Originality.ai review
• Is GPTZero accurate?
• How AI detectors work