rel=nofollow citations. The three reviews cited against it are published by companies selling competing humanizers, and we say so each time they appear. The promoted call-to-action above is a StealthWriter affiliate placement with its own disclosure.Does UnAIMyText bypass GPTZero? Its own test says 3% of its output gets caught. Three other tests of the same product say 90 to 100%.
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
- UnAIMyText is the most methodologically transparent vendor in this category. It publishes sample sizes, the models behind its corpus, its exact humanizer settings, a timing window and a limitations section. Almost nobody else does.
- Its July 2026 round reports GPTZero catching 96% of raw AI text and 3% after Ultra mode — a claimed 97% pass rate.
- Three separate reviewers tested the same product and got 90.4%, 100% and 100% detected. One of those ran in the same month as UnAIMyText’s own round.
- All three reviewers sell competing humanizers, so treat them as interested. They also published screenshots. UnAIMyText published none — the only image file on the entire site is the logo.
- Its corpus is built from 2024-era models — GPT-4o, GPT-4o mini, Claude 3.5 Haiku, Gemini 1.5 Flash — run in July 2026 against detectors calibrated on far newer output.
- The “monthly” benchmark has two months in it. The August row still read pending when we checked on 4 September 2026, and there is no September row.
- GPTZero is free to screenshot — 10,000 characters, no account. That is what makes the missing image conspicuous rather than merely absent.
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 →Does UnAIMyText bypass GPTZero?
This one is unusual, and it is worth saying why before anything else.
Most vendors in this series shout a bypass percentage in marketing and quietly disclaim it in their terms. UnAIMyText does the opposite. It prints the caution itself, on its own homepage, directly beneath its own results table: “no humanizer can promise a specific score on every detector, every time, and you should be sceptical of any tool that claims otherwise.” Then it publishes a benchmark with a disclosed method.
And three independent tests found the opposite of what that benchmark reports. Working out which to believe is the whole exercise here.
This page reports no first-party test. Everything below was fetched on 4 September 2026 and every claim is dated and attributed, including where the source has a commercial stake.
How much does it actually publish?
More than most detector companies publish about themselves, and it deserves credit for that before the criticism starts.
Its detector-tests page discloses 50 samples per round; the exact model split generating the corpus — 20 from ChatGPT-4o mini, 15 from GPT-4o, 10 from Claude 3.5 Haiku, 5 from Gemini 1.5 Flash; the content-type breakdown across 20 academic essays, 15 business samples, 10 blog posts and 5 creative pieces; a 250 to 600 word range; the humanizer configuration down to Creativity 70%, Context Preservation 80%, Emotional Tone Neutral, Formality Formal; that all detectors ran on 3 July 2026 within a four-hour window “to minimize model drift”; that error samples were discarded and replaced; and that samples are not reused between rounds. It adds a five-point limitations section conceding that no tool prevents detection perfectly and that silent detector updates shift results month to month.
| Detector | Raw AI detected | Standard mode | Ultra mode | Claimed pass rate |
|---|---|---|---|---|
| GPTZero | 96% | 15% | 3% | 97% |
| Turnitin | 94% | 18% | 2% | 98% |
| Copyleaks | 90% | 20% | 4% | 96% |
What do the independent tests say?
Three reviewers have tested UnAIMyText against GPTZero and published results. All three found it caught.
| Source | Date | GPTZero result | Method | Screenshots |
|---|---|---|---|---|
| UnAIMyText (self) | 3 Jul 2026 | 3% detected | 50 samples, method published | none |
| Brus | 13 Jul 2026 | 90.4% detected | 20 text pairs, 3 detectors, 3 LLM judges | yes |
| Evan | 3 Dec 2025 | 100% detected | 3 samples, all three modes | yes |
| Verva (ex-TwainGPT) | 13 Nov 2025 | 100% detected | not disclosed | not stated |
We are not treating the reviews as neutral. Verva sells a rival humanizer and says so in its conclusion; the other two sit on a forum operated by Clever AI Humanizer, also a rival. Adverse interest runs in both directions on this page.
What separates them is what they put on screen. Brus’s test also recorded grammar scoring at 10.4 out of 20 — second worst of eleven tools — and word-count inflation beyond 30% on 9 of 20 samples, findings that are independent of the detection question and would be odd to fabricate alongside it.
Why does the missing screenshot matter so much here?
Turnitin sells no individual licences and shows students no AI report, so nobody outside an institution can produce a genuine Turnitin screenshot. GPTZero removes that excuse entirely. Anyone can paste up to 10,000 characters into gptzero.me right now with no account, no card and no signup; a free account raises the per-scan limit to 150,000 characters with a 10,000-word monthly quota, and paid tiers start at $9.99 a month.
So when a humanizer claims to beat GPTZero and shows you nothing, that is a choice, not a constraint.
It also means you can check your own text. A genuine current GPTZero result shows a three-class classification — human, AI or mixed — a confidence category rather than a bare percentage, a probability breakdown across all three classes, and sentence-by-sentence highlighting. An image showing only “X% AI”, or one displaying perplexity and burstiness, is either fabricated or dates to roughly 2023.
We extracted every image reference across the homepage and both test pages. The only image file on the site is the logo. Every pass rate is rendered as HTML text in a table.
For Turnitin that would be forgivable, since no vendor can lawfully obtain a Turnitin AI report. For GPTZero it is not. A single scan costs nothing, takes a minute, needs no account, and would produce exactly the artefact that settles the dispute above. A company that publishes its creativity slider setting to two significant figures has chosen not to publish the one image anyone could verify.
How old is the corpus being tested?
Older than the detectors it is tested against, which is the quiet problem underneath the disagreement.
All 50 samples come from GPT-4o, GPT-4o mini, Claude 3.5 Haiku and Gemini 1.5 Flash — 2024-era models — and were run in July 2026 against detectors retrained repeatedly on newer output in the interim. Detectors generally find older model output easier to identify, which if anything should push the detected figure up rather than down.
Two smaller inconsistencies sit alongside it. The prose describes testing in Standard mode while every headline figure quoted is Ultra. And the monthly history contains two months: June and July 2026. The August row still read pending when we checked on 4 September 2026, and no September row exists — so the benchmark has now missed two consecutive cycles.
Has any detector company tested it?
None has. UnAIMyText is absent from the appendix table in GPTZero’s February 2026 paper, which names nine bypass services individually; absent from the nineteen tools in Pangram’s DAMAGE paper; and absent from the twenty in Pangram’s August 2025 benchmark. We also ran a full census of all 41 pages on the site and found Pangram mentioned zero times, which we cover in does UnAIMyText bypass Pangram.
How reliable are the detector numbers on either side?
GPTZero’s own FAQ concedes: “Our classifier is not trained to identify AI-generated text after it has been heavily modified after generation.” Its developers page simultaneously claims a “Paraphraser Shield” means “even if AI content has been altered to look more human-like, GPTZero can detect it.” Both statements are currently published.
The two leading detector vendors publish opposite numbers on the same question. GPTZero’s February 2026 paper reports 93.5% recall on 1,000 texts run through nine bypasser services, and puts Pangram at 49.7%. Pangram’s DAMAGE paper reports GPTZero at 60.04% on humanized text, and 34.53% at GPTZero’s own default threshold. Each benchmark shows its publisher winning, and neither is peer-reviewed.
On false positives, GPTZero was one of seven detectors in Liang et al. (2023), which reported a 61.22% average false positive rate across those seven on 91 TOEFL essays. No per-detector figure was ever published, so that number is not GPTZero’s. GPTZero now claims it has cut its own TOEFL false positive rate to 1.1% — a figure no third party has verified.
What should you take from this?
UnAIMyText behaves better than most of this market. It publishes a method, states its limitations, and warns you off believing bypass guarantees, including implicitly its own. None of that is common and none of it should be dismissed.
But a method without data is a description of an experiment, not a result. Its 3% figure cannot be reproduced by anyone outside the company, three separate testers reported between 90% and 100% detection instead, the closest of those ran ten days after its own round, the corpus is two years older than the detectors judging it, and the monthly benchmark has stopped updating.
On this evidence the honest answer is that the claim is unverified and the available independent testing runs against it. You can settle it for your own text in about a minute: paste your output into gptzero.me, free, no account, and read the three-class result rather than anybody’s table.
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 UnAIMyText review, and the same question against Turnitin and Pangram. Also in this cluster: Ryne AI, WriteHuman, StealthGPT.
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 UnAIMyText bypass GPTZero?
The published record disagrees with itself, sharply. UnAIMyText’s own July 2026 round reports GPTZero detecting only 3% of its Ultra-mode output, a 97% pass rate. Three independent reviews of the same product report 90.4%, 100% and 100% detected. Both sides have a commercial stake. The difference is that the three reviews published screenshots and UnAIMyText published none.
Why call it the most transparent vendor if the numbers are disputed?
Because transparency of method and quality of evidence are different things, and UnAIMyText genuinely has the first. Its detector-tests page discloses 50 samples per round, the exact model split generating the corpus, the content-type breakdown, the humanizer configuration down to creativity and formality settings, a 250 to 600 word range, a four-hour testing window stated as being there to minimise model drift, that error samples were discarded and replaced, and that samples are not reused between rounds. It adds a five-point limitations section. Most vendors in this series publish a percentage and nothing else.
What did the independent reviews find?
Three tests, all adverse. Verva, formerly TwainGPT, published on 13 November 2025 and reported 100% AI on GPTZero. A reviewer named Evan published on 3 December 2025 after testing three ChatGPT samples across all three UnAIMyText modes, and reported GPTZero flagging every single humanized output at 100% AI. A reviewer named Brus published on 13 July 2026 after running 20 identical text pairs through three detection platforms and three LLM judges, and reported a 90.4% average AI detection on GPTZero with 15 of the 20 samples scoring 99 to 100%.
How do those square with the vendor’s own July number?
They do not, and the timing makes it stark. UnAIMyText’s July 2026 round ran on 3 July and reports 3% detected. Brus’s 20-pair test published on 13 July and reports 90.4% detected. Same month, same product, roughly 87 points apart. Something in one of those two designs is producing a result the other cannot reproduce, and neither party has released the raw outputs that would settle it.
Should the competitor reviews be trusted?
Not uncritically. Verva sells a competing humanizer and closes its review by recommending its own product. The other two are published on a forum run by Clever AI Humanizer, another competitor. Adverse interest cuts both ways here. What separates them is evidential: all three show detector screenshots, and Brus’s test also documents grammar scoring and word-count inflation independent of the detection question. UnAIMyText’s site contains no detector screenshot at all.
Is the benchmark being kept up to date?
Not currently. The history table lists June 2026 and July 2026 with 50 samples each. The August 2026 row read pending when we checked on 4 September 2026, and no September row exists. The most recent actual round was run on 3 July 2026, so a benchmark described as monthly has gone two cycles without an update.
What else is worth knowing about the corpus?
Its age. The 50 samples are generated by GPT-4o, GPT-4o mini, Claude 3.5 Haiku and Gemini 1.5 Flash, all 2024-era models, then tested in July 2026 against detectors that have been retrained repeatedly on newer output. There is also a mode inconsistency: the prose describes testing in Standard mode while every headline figure quoted is Ultra. And there is no raw data of any kind, only percentages rendered as HTML text.
Has any detector company tested UnAIMyText?
None has. It is absent from the appendix table in GPTZero’s February 2026 paper, which names nine bypass services individually, and absent from Pangram’s DAMAGE paper and its August 2025 benchmark of twenty tools. We also ran a full census of all 41 pages on unaimytext.com and found the word Pangram zero times, which we cover in the Pangram version of this question.
