rel=nofollow citations. The promoted call-to-action above is a separate StealthWriter affiliate placement with its own disclosure. Our companion page on Turnitin reached an unfavourable conclusion about the same product.Does Undetectable AI bypass ZeroGPT? The vendor never claims it does — and the peer-reviewed study it cites on its own website is the best evidence that something like it would.
The only tool that bypassed Pangram and Turnitin in 2026.
Most humanizers clear one detector and get caught by the other. StealthWriter is the one that gets past both — run Ghost 5.2 Pro at level 7–8, section by section, and re-check before you submit.
Try StealthWriter →Key Takeaways
- Undetectable AI never claims to beat ZeroGPT. We read its homepage, humanizer page, detector page and pricing at source: ZeroGPT is not named once, on a site that names GPTZero, Originality.ai, Copyleaks and Writer.com by name and grades them.
- But its own detector page cites a peer-reviewed study to prove its detector is best — and the same table in that study contains ZeroGPT’s number.
- Kar et al. (2024) found ZeroGPT scored 95.03% on raw AI text and 32.31% once a paraphraser had been over it. A rewrite cut the score by roughly two thirds.
- Undetectable AI sells a paraphraser. The study it quotes to sell its detector is the strongest published evidence that its other product works on ZeroGPT.
- Handle it honestly: that study is four documents, and the paraphrasers were QuillBot, Grammarly and ChatGPT — not Undetectable AI. It shows ZeroGPT is fragile, not that this tool beats it.
- The same table is worse for the other detector. GPTZero fell to 29% and 38% on two of the three paraphrases, below ZeroGPT on both.
- Two rival detector vendors do rank this tool hardest of the field to catch — GPTZero rates it the only “High” of nine, Pangram caught it 90.3%, lowest of twenty. Neither figure is about ZeroGPT.
Does Undetectable AI bypass ZeroGPT?
No published test of this exact pairing exists. Not from the vendor, not from a detector company, not from us.
What makes this cell unusual is where the nearest evidence sits. It is on Undetectable AI’s own website, in a citation the company put there to sell a different product.
This page reports no first-party test; everything below was fetched on 21 September 2026.
What does the study on its own site show?
Undetectable AI’s detector page cites a peer-reviewed paper to establish that its detector is the most accurate available: Kar, Bansal, Modi and Singh (2024), in the Indian Journal of Psychological Medicine. The quote it pulls is accurate — the study does find Undetectable AI and Sapling detecting all three paraphrased texts with 100% accuracy.
The study tested ten free detectors. ZeroGPT is one of them, in the same table.
| Detector | Original AI text | QuillBot paraphrase | Grammarly paraphrase | ChatGPT paraphrase |
|---|---|---|---|---|
| ZeroGPT | 95.03% | 32.31% | 96.74% | 90.61% |
| GPTZero | 97% | 29% | 38% | 97% |
| Undetectable AI | 100% | 100% | 100% | 100% |
| Sapling | 100% | 100% | 100% | 100% |
| Copyleaks | 100% | 0% | 100% | 100% |
| Scispace | 61% | 29% | 42% | 26% |
Read the ZeroGPT row on its own. Raw AI text: 95.03%. The same content after one paraphraser: 32.31%. A rewrite cut the score by roughly two thirds, and 32.31% on ZeroGPT’s gauge reads as predominantly human.
What are the limits of that evidence?
Substantial, and we are going to state them before anyone quotes the 32.31% without them.
The study is four documents. One ChatGPT-written article on electroconvulsive therapy, plus three paraphrased versions of it. Peer review does not turn n=4 into a benchmark, and the authors do not claim it does.
The paraphrasers were not Undetectable AI. They were QuillBot, Grammarly and ChatGPT. So the table shows what generic rewriting does to ZeroGPT, not what this specific tool does to it. Anyone citing this as proof that Undetectable AI beats ZeroGPT is citing it wrongly, and that includes anyone citing this page.
The result is not even consistent within the study. One paraphraser out of three collapsed ZeroGPT. The other two did not — 96.74% and 90.61% are clean catches. The honest summary is that ZeroGPT is vulnerable to some rewrites and not others, which is a much more useful thing to know than a single headline number.
And the comparison cuts the other way too. On the Grammarly paraphrase, ZeroGPT held at 96.74% while GPTZero dropped to 38%. The reflexive assumption in this market is that ZeroGPT is the weaker of the two. On this particular evidence it is not, and we are not going to suppress that because it complicates the page.
Where is ZeroGPT in Undetectable AI’s own marketing?
Nowhere. We read the homepage, the AI Humanizer page, the AI Detector page and the pricing page at source and searched each for “ZeroGPT”. It does not appear on any of them.
That absence is louder here than it would be elsewhere, because this vendor is unusually willing to name rivals. Its detector page runs a comparison grid that grades GPTZero at 77.2% accuracy with a B-, Originality.ai at 87.9% with a C, Writer.com at 62% with a D, and Copyleaks with an A, against 100% detection and 0% false positives for itself.
A company that will publish letter grades for four named competitors has left one of the world’s most-used free detectors off the board entirely. We are recording the asymmetry, not guessing at a reason for it.
The humanizer page, meanwhile, names no detector at all. Its strongest wording is that the free version “may be flagged by detectors” and that a free trial lets you “bypass detectors with 99% confidence” — a figure with no study, corpus, date or detector attached.
What do rival detector companies say about this tool?
More than about any other humanizer in this series, and it is the most credible kind of evidence available here: concessions from companies that sell against it.
Appendix M of GPTZero’s February 2026 paper names nine bypass services and rates each. Undetectable is the only one marked “High”. Pangram’s August 2025 benchmark caught its output 90.3% of the time, the lowest of all twenty tools scored.
Now the fine print that most coverage drops. GPTZero’s column header reads “bypassing ability on naive AI detection methods”, and its caption states the nine are “all ineffective against GPTZero.” Neither figure is about ZeroGPT. And 90.3% still means caught nine times in ten. Pangram’s DAMAGE paper separately rates this tool’s output quality L3, the lowest tier — which may be the same fact from the other side, since the paper notes less fluent output can be harder to detect.
Can you check ZeroGPT yourself?
ZeroGPT is free and needs no account. You can paste up to 15,000 characters per check at zerogpt.com right now, which is more per scan than GPTZero’s signed-out tier allows.
A genuine current ZeroGPT result shows a percentage gauge giving the share of the text it reads as AI, plus sentence-level highlighting of the passages it flagged. The paid tiers add automatically generated PDF reports.
So for this detector there is no access excuse. Any vendor claiming to beat ZeroGPT can demonstrate it in a minute, for nothing, and so can you.
Which is the practical conclusion of this whole page. Nobody has tested this pairing, the study that gets closest is four documents old enough to predate the current versions of both products, and you can run the exact test that matters — your text, today, on the tool you are actually using — for nothing, in the time it takes to paste it.
How reliable is ZeroGPT itself?
ZeroGPT advertises “98.4% Detection accuracy” and a “<1% False positive rate” across “20+ LLM families”. No study, sample size, corpus or date is cited for any of those three figures anywhere on the site.
The independent record is less flattering and older. ZeroGPT was one of the seven detectors in Liang et al. (2023), which ran 91 TOEFL essays by non-native English writers through all seven and reported a 61.22% average false positive rate against near-perfect accuracy on US eighth-grade essays. The per-detector results appear only as a graph, with no individual figure stated numerically, so that average is not ZeroGPT’s own score — but ZeroGPT is inside it.
One caution about sourcing. Search results for ZeroGPT’s accuracy are dominated by blogs run by humanizer vendors — companies that sell the product for defeating it. We have not used those as evidence anywhere in this cluster, and neither should you.
What should you take from this?
This is the rare cell where the evidence leans toward the tool, and it leans for a reason that has nothing to do with the tool’s marketing.
The vendor makes no ZeroGPT claim. Two rival detector companies rank it hardest of the field to catch, on questions that are not about ZeroGPT. And the one peer-reviewed table anywhere in this picture — the one the vendor itself put on its website — shows ZeroGPT losing roughly two thirds of its score to a paraphraser that was not even this product.
Better odds than its rivals, on thin evidence, against a detector that is unreliable in both directions. That is the honest position, and it is also why the check costs you nothing to run: paste the output into zerogpt.com, then paste the same text into a second detector, because agreement between two is worth more than a good number from one.
If your real worry is being wrongly flagged on your own writing, keep your drafting history — it costs nothing and it survives a detector being wrong. It matters most for the writers detectors treat worst: ESL writers and AI detection and false positives for neurodivergent students. Check before you submit with the best pre-submission check and the Turnitin self-check, and strip the obvious tells by hand first via what to remove before using an AI humanizer.
For the product itself see our Undetectable AI review, and the same question against Turnitin and GPTZero. Also in this cluster: SuperHumanizer.
Get the free prompt pack instead
The manual rewriting prompts that lower AI signals without a subscription — and without betting a submission on a number nobody has reproduced. Free, no card.
Frequently asked questions
Does Undetectable AI bypass ZeroGPT?
Nobody has published a test of this specific pairing, and the vendor does not claim it. ZeroGPT is not named anywhere on undetectable.ai. What exists is indirect and reasonably strong: the peer-reviewed study Undetectable AI itself cites on its detector page found ZeroGPT fell from 95.03% on raw AI text to 32.31% once a paraphraser had rewritten it. That measures ZeroGPT’s fragility against paraphrasing in general, not this product in particular.
What is the study on Undetectable AI’s own site?
Kar, Bansal, Modi and Singh, How Sensitive Are the Free AI-detector Tools in Detecting AI-generated Texts?, Indian Journal of Psychological Medicine, published online May 2024. It ran one ChatGPT-written article and three paraphrased versions of it through ten free detectors. Undetectable AI quotes it because its own detector scored 100% across all four texts. ZeroGPT is one of the other nine tools in the same table.
What exactly were ZeroGPT’s numbers in it?
ZeroGPT returned 95.03% AI on the original ChatGPT text, 32.31% on the QuillBot paraphrase, 96.74% on the Grammarly paraphrase and 90.61% on the ChatGPT paraphrase. So one paraphraser out of three collapsed it and two did not. That is a real published result and it is also four documents, which is the correct weight to give it.
How did GPTZero do in the same table?
Worse on two of the three. GPTZero scored 97% on the original text, then 29% on the QuillBot paraphrase and 38% on the Grammarly paraphrase, recovering to 97% on the ChatGPT paraphrase. On the Grammarly version ZeroGPT held at 96.74% while GPTZero fell to 38%. We report this because the reflexive assumption in this market is that ZeroGPT is the weaker tool, and on this particular evidence it is not.
Why does the vendor not mention ZeroGPT at all?
We do not know and will not guess at a motive. The fact is the asymmetry. Undetectable AI’s detector page publishes a comparison grid grading GPTZero, Originality.ai, Copyleaks and Writer.com against itself, with letter grades and accuracy figures. ZeroGPT, one of the most-used free detectors in the world, is absent from that grid and from every other page we read.
What do rival detector companies say about this tool?
More than about any other humanizer. GPTZero’s February 2026 paper rates its bypassing ability High, the only one of nine named services to get that rating. Pangram’s August 2025 benchmark caught it 90.3% of the time, the lowest detection rate of all twenty tools scored. Both are adverse-interest evidence from companies that sell against it. Neither measured ZeroGPT, and 90.3% still means caught nine times in ten.
Is ZeroGPT worth checking against at all?
As one input, yes, because it is free, takes 15,000 characters and needs no account. Not as your only input. Its own accuracy claims cite no study, it was in the seven-detector cohort that produced a 61.22% average false positive rate on essays by non-native English writers, and in 2023 it rated the United States Constitution 92.15% AI-generated. Agreement between two detectors is worth far more than a good number from one.
What would actually settle this?
A published test of Undetectable AI output against ZeroGPT, with the source text, the date, the model setting and the screenshot. Nobody has run one, including us. Until somebody does, the honest answer is that the strongest evidence here is about what paraphrasing does to ZeroGPT in general, and you can reproduce that part yourself in two minutes for nothing.
