Introduction

One saved ZeroGPT result is more instructive than a confident-looking average: ai-marketing-01, a text known to have been AI-written before it was scanned, displayed 28.7% AI. Other AI-written inputs in the same recorded run displayed much higher numbers. A low detector score therefore cannot clear a document of AI involvement. Nor can a high score, by itself, establish who wrote or edited a disputed passage.

This review is for editors, content teams, and researchers deciding how to use ZeroGPT as a screening signal. It is not a verdict about a student’s honesty or an author’s process. The September 2, 2026 saved run included six documented human excerpts and six known AI-generated excerpts across editorial, marketing, and technical writing. The public method and individual outputs show the inputs, labels, source links, screenshots, and limits. The same campaign also covered Sapling and Undetectable AI; this page focuses on what the ZeroGPT results mean for an editorial decision.

ZeroGPT’s official detector page markets the product as a free AI checker and makes broad accuracy claims. Those are the vendor’s statements, not findings established by our twelve inputs. The review below separates what was recorded from what a buyer must still check on the live product.

Who It Is Best For

ZeroGPT may be useful to a reviewer who wants an initial signal and has a documented path for checking the text itself. A low score can prompt a look at drafts and revision history; a high score can prompt the same look. Either result should be treated as a reason to investigate, not as an automatic accusation or approval. Teams that already preserve source documents and permit writers to explain their work are better placed to use such a signal fairly.

It is a poor choice as the sole gate for grading, hiring, rejecting submissions, or declaring content “SEO safe.” The saved 28.7% result was on text whose AI origin was known in advance. If the origin had been unknown, that interface number alone could not have revealed it. Conversely, a high display on a genuine human draft would require examination, not punishment. The six human examples here are historical public-domain prose, so their low scores cannot establish how ZeroGPT treats contemporary essays, edited marketing copy, or multilingual material.

Before adopting any detector, test it on material resembling your actual workflow. Obtain permission where needed; do not upload sensitive client, student, or employee text until the service’s data-handling terms fit your policy. Record the exact input and date alongside the output. The useful question is whether the result leads to a fair next step, not whether the interface supplies a persuasive percentage.

What the Saved Results Show

The experiment used twelve independently labeled excerpts, two of each label in three writing categories. The human texts have preserved public-domain source links; the AI texts were labeled before scanning. Each percentage below is a saved interface display for that particular input, not a calibrated probability of authorship. A newer product version, different passage length, or edited text could behave differently.

Saved input	Known label	ZeroGPT display
`ai-editorial-01`	AI-written	78.4%
`ai-editorial-02`	AI-written	100.0%
`ai-marketing-01`	AI-written	28.7%
`ai-marketing-02`	AI-written	83.3%
`ai-technical-01`	AI-written	100.0%
`ai-technical-02`	AI-written	100.0%
`human-editorial-01`	Human source	8.0%
`human-editorial-02`	Human source	0.0%
`human-marketing-01`	Human source	0.0%
`human-marketing-02`	Human source	0.0%
`human-technical-01`	Human source	0.0%
`human-technical-02`	Human source	0.0%

Five known AI inputs displayed 78.4% or above; the sixth was the 28.7% marketing input. All six saved human excerpts displayed 0.0%–8.0%. This is a description of a small, dated set, not a population accuracy rate, a false-positive rate, or proof that one category of writing causes a particular score. The record does not isolate whether wording, length, model, formatting, or other factors explain the outlier. It also does not test the behavior of every plan or the handling of confidential text.

The practical lesson is narrower and stronger than a ranking: a reviewer cannot set an “AI below this number means human” cutoff from this run. The text at ai-marketing-01 had a known origin regardless of its display. Someone interpreting an unknown document would need independent information about its creation. Keep the full verbal output where available, not merely the number, and repeat a trial after substantial product changes.

Pricing

The vendor’s pricing page presents a free entry point and enterprise positioning. The saved official-page metadata does not establish today’s exact price, word allowance, API quota, renewal conditions, or which features belong to each plan. We did not purchase and audit every subscription tier. Check the official website for the latest pricing.

If considering a paid plan, ask what problem the payment solves. A higher allowance can be useful for a team processing many documents, but it does not turn a detector result into proof. Compare the permitted volume, data handling, export or record-keeping needs, and cancellation terms with the actual editorial workflow. The value of any tier depends on how its output is used, not only on the number of scans it permits.

Pros

All twelve saved submissions returned a recorded display, so this small run has no missing ZeroGPT score.
The six documented human excerpts received low displays in this run. That pattern is useful to inspect, while remaining too narrow to establish a general false-positive rate.
Five of the six known AI inputs received much higher displays than the human examples. For those exact inputs, the interface supplied a visible signal that could prompt review.
A public-facing checker is described on the vendor’s site. It may be straightforward for a team to run a small, consented trial before considering a paid plan.

These are observations about the saved set or clearly attributed vendor descriptions. They are not a claim that ZeroGPT is the best detector or that its current version will repeat the same scores.

Cons

The known AI input ai-marketing-01 displayed just 28.7%. A workflow that automatically clears lower-scoring content would have misinterpreted this example.
Twelve inputs are not enough for a population performance estimate. There are only six human sources and six AI sources, and no systematic coverage of edited, mixed, multilingual, or very short passages.
The human material is historical public-domain text. Modern workplace and school writing may have different properties; the recorded results do not resolve that uncertainty.
A percentage can look more precise than the evidence warrants. The saved archive cannot show that a 28.7% display means a 28.7% chance of AI authorship.
Pricing, current account limits, privacy behavior, and support quality were not measured in the recorded run. The vendor’s marketing language should be checked independently before purchase or policy use.

Alternatives

The Sapling detector review examines another product from the same saved campaign and records where its results on documented human text require caution. The Undetectable AI detector review explains a different display pattern on the same labeled inputs. Those pages are useful for comparing observed outputs, not for declaring a universal winner. Our public benchmark data makes the sample-level evidence available so a reader can inspect the differences directly.

There is also a non-product alternative: preserve revision history, citations, notes, and author communication, then use a detector only as optional context. For consequential decisions, that process is more defensible than treating one vendor’s percentage as an authorship certificate. If a team does compare products, it should repeat the comparison with its own consented, representative material and current versions rather than choosing solely from a twelve-input archive.

Final Verdict

ZeroGPT is reasonable to trial as an initial screening interface if a team is prepared to verify current terms and interpret results cautiously. The September 2, 2026 archive shows low displays on six historical human excerpts and high displays on five known AI excerpts. It also shows a crucial exception: 28.7% on one known AI marketing input. The exception makes a simple cutoff especially hard to defend; the other results do not erase it.

We cannot infer why that input received a lower display, calculate a meaningful population accuracy from this small set, or promise the live product behaves identically now. Use the result as a prompt to review the actual document and its history. Do not use it to certify SEO quality or to make an unappealable authorship judgment. Last official-page check: September 26, 2026. Last saved experiment: September 2, 2026. Check the official website for the latest pricing.

Affiliate Disclosure

This article may contain affiliate links. A commercial relationship does not change the recorded scores, our refusal to infer a general accuracy rate, or our conclusion about the 28.7% example. AI assisted with organizing the archived evidence and drafting this article; the sample inputs, saved interface outputs, and source checks were not invented. Readers can inspect the method and raw data and the vendor’s detector and pricing pages directly.