ZeroGPT and GPTZero Are Separate Tools With Launches Dated 17 Days Apart

The web address says which product produced a score, and the logo doesn't: GPTZero lives at gptzero.me and ZeroGPT at zerogpt.com, two separate products one letter-swap apart. GPTZero's own comparison page answers the question in a single line: completely separate tools with similar names ¹. A reading from one can't be taken as a reading from the other.

On a Tuesday this shows up as a forwarded screenshot: a percentage, a coloured bar, a product name half-cropped at the top, and a message asking what to do.

A cropped image rarely carries the address, so ask for the link. The same click shows two more tells, because GPTZero's page names its founders and a research lineage ¹ while ZeroGPT's home page gives a support email and a detection figure and names nobody ². Neither product hides any of that. None of it fits in a screenshot.

Seventeen days separate the launches, and the second name is the first one reversed. GPTZero dates its own launch to 1 January 2023 and ZeroGPT's to 18 January 2023, and says it came out of research at Princeton's NLP lab and was started by Edward Tian and Alex Cui. Its page calls ZeroGPT a copycat site capitalizing on brand confusion ¹.

That phrase comes from a competitor, and it reads like one.

The date has one outside check. The Internet Archive's first capture of zerogpt.com is dated 18 January 2023 ⁸, which shows the site was up that day and is consistent with GPTZero's account; a first capture is not a launch record. The motive has no such check, and it stays GPTZero's characterisation rather than a finding.

ZeroGPT doesn't return the comparison. Its home page has a detection figure and a contact email and says nothing about GPTZero or any link between the two ². Of the two companies, only GPTZero has published an answer to the naming question on the pages cited here.

Each publishes a detection figure and a false-positive figure for its own product. GPTZero states that its checker detects 95.7% of AI texts while wrongly predicting 1% of human texts as AI ¹. ZeroGPT's home page states 98.4% detection accuracy and a false positive rate under 1%, and it names nobody behind the figure ².

GPTZero says its number was measured on RAID, a third-party benchmark dataset ¹. ZeroGPT names no test set or method beside its number ². The University of Pittsburgh's teaching centre notes that headline claims in the 98 to 99 percent range are common among detector vendors, and that third-party evaluation of the same tools tends to find high false positive rates ⁵. The same page records that many of these companies decline to publish a false positive rate at all, and it states that detectors are not reliable enough to act as proof ⁵.

One independent test exists for one of the two. Bloomberg Businessweek ran 500 college application essays submitted to Texas A&M in the summer of 2022, before ChatGPT was released. GPTZero and Copyleaks falsely flagged 1% to 2% of them, in some cases claiming near-total certainty ⁴. That range sits at or just above GPTZero's own published 1% ¹. ZeroGPT was not part of the test.

Take a department reading 2,000 papers a term and assume, for the arithmetic, that all of them are human-written and that the rate Bloomberg saw on application essays carries over to them. Neither is measured. On those assumptions 1% to 2% is 20 to 40 wrongly flagged papers before anyone has been spoken to, and the false-positive arithmetic worked through shows how the count scales.

Nor is the error spread evenly. Seven detectors tested on 91 TOEFL essays by non-native English writers wrongly flagged more than half of them, an average false positive rate of 61.3% ⁶. That was 2023, and neither vendor page cited here shows the result has changed. A flag on a multilingual writer's essay deserves more caution, not less, and that case has its own rules.

GPTZero's FAQ figures come with a condition. By its own FAQ, 99.1% of human articles are classified as human and 98.4% of AI articles as AI, inside its high confidence band ³. The page labels those figures by band, so a reading quoted without its band has lost the condition that produced it. How often one paper lands inside the band, and what the reading is worth when it doesn't, is a separate question about GPTZero specifically.

Both vendor numbers are self-reported. Bloomberg's wasn't.

Because only one of the two companies has answered it. GPTZero keeps a comparison page that states the relationship and dates both launches, then reports its own detection figures on the same screen ¹. ZeroGPT's home page doesn't address the question at all ². The clearest account found comes from an interested party.

Vendor benchmarks have a shape worth knowing. GPTZero also publishes a comparison against Pangram, a detector made by a different company, scoring itself at 99.6% accuracy with a 0.13% false positive rate, and its own product comes out ahead in it ⁷.

A table where the publisher wins is evidence about the table.

None of that makes the disambiguation wrong. The separate-products answer sits on a dated page, and ZeroGPT's launch date matches the Internet Archive's first capture of its address ⁸, so the date has a check that does not depend on GPTZero. The accuracy figures beside them are a different kind of claim, and they get a different level of trust.

Four details, plus the confidence band where the tool shows one, make a detector reading re-checkable a week later: the product name, the web address, the date, and the word count of the text that was scored. A bare percentage has none of them, and even the teacher who ran it will struggle to reproduce it. GPTZero's own FAQ says its results should not be used to punish students ³.

Length is easy to lose. A pasted paragraph gives a detector less to read than a full essay. The paste box at human.olive.is, which takes up to 1,000 words, states its own rule: Human won't score anything under 50 words, and it treats a reading on 50 to 149 words as weak.

Running the same text through the same tool twice isn't a second opinion. What does carry information is the work itself: a draft history with dates on it, the notes and sources, and a few minutes talking through the argument. A detector at either address sees none of that.

Numbers from the maker of a detector are still the maker's numbers, and Human's are no exception. In its own in-house measurement on held-out sets, as of September 2026 and not an independent audit, it records 0 false flags in 1,928 human documents it had never seen; the statistical ceiling on that is 0.155%. This is an estimate from our detector. Treat a flag as a reason to look closer, not as a finding. The figure describes documents, never students, and it predicts nothing about what either of the other two addresses would return on the same paper.

Write the product name into the gradebook note before the number goes anywhere.