Does Ryne AI Bypass GPTZero? Check Its 100,000-Text Study (2026)

Published:

Updated:

Disclosure. Detection Drama earns affiliate commission from Ryne AI, and Ryne is the sponsor link on our YouTube channel. This page reaches an unfavourable conclusion for that relationship and we publish it anyway. No link to Ryne AI on this page is an affiliate link; outbound links are rel=nofollow citations. The promoted call-to-action above is a StealthWriter affiliate placement carrying its own disclosure, unrelated to this review.

Does Ryne AI bypass GPTZero? It claims 99.8%, and it published a study titled “We Ran 100,000+ Texts to prove it” — a study that does not contain the 100,000 texts.

Detection Drama · 2026 Verdict

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 →
Free plan to start · No credit card · Paid from $20/mo
Or get the free prompt pack first →
Affiliate link — I earn a commission if you upgrade. I only point at tools I actually run.

Key Takeaways

  • Ryne AI publishes a blog post titled “Why GPTZero is not reliable anymore. We Ran 100,000+ Texts to prove it.” We read it. The 100,000 texts appear in the headline and the meta tags, and nowhere in the body.
  • There is no sample description, no source corpus, no dates, no models, no thresholds and no raw data. The post is a literature review of other people’s studies, written by Ryne’s CEO.
  • It also misattributes the best-known figure in the field: it reports 61.3% of non-native essays falsely flagged as if it were GPTZero’s rate. That number is an average across seven detectors, and no per-detector figure was ever published.
  • Ryne quotes two different GPTZero numbers on its own site — 99.9% on the homepage, 99.8% on the humanizer page. Neither carries a method.
  • Its own page answering this exact question is dated 15 December 2024, last touched May 2025, and is bylined to an author titled “AI Writer”. It contains no percentage and no screenshot.
  • GPTZero is free to check yourself — 10,000 characters, no account. Ryne shows no GPTZero screenshot anywhere.
  • Refunds need all three: within 3 days, under 500 coins, under 1 AI report — then a 30% processing fee. Renewals are non-refundable.

By Vlad Ivanov — publisher of Detection Drama, tracking AI-detector and humanizer behaviour across 245 published tests and teardowns. Last updated: August 2026

Does Ryne AI bypass GPTZero - 2026 evidence review
Detection Drama · 2026 Verdict

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 →
Free plan to start · No credit card · Paid from $20/mo
Or get the free prompt pack first →
Affiliate link — I earn a commission if you upgrade. I only point at tools I actually run.

Does Ryne AI bypass GPTZero?

Unknown, and unusually, that is not because the vendor is quiet. Ryne AI is one of the most publication-heavy companies in this market. It names ten detectors, attaches percentages to four of them, runs roughly thirty-one competitor and detector teardowns, and has a dedicated post answering this exact question.

None of it contains a test you can check. That is what this page is about.

We report no first-party test of our own. Everything below was fetched from ryne.ai and GPTZero directly on 3 September 2026, and every claim is dated and attributed — including where the source has a commercial stake, ours included.

What does Ryne actually claim about GPTZero?

Two things, and they do not match each other.

The homepage leads with “Bypass Turnitin & GPTZero with a 99.9% success rate.” The humanizer page breaks the same claim down per detector and gives GPTZero 99.8%, alongside Turnitin at 100%, Originality.ai at 99.8% and Copyleaks at 99.7%.

Where it appearsGPTZero figureSample sizeDateMethod
Homepage headline99.9%nonenonenone
Humanizer page table99.8%nonenonenone
“Does Ryne AI Bypass GPTZero?” postno figurenoneDec 2024none
“100,000+ texts” studyno figurenot describedMar 2026none

Source: ryne.ai homepage, humanizer page and blog, read directly 3 September 2026.

A tenth of a percentage point is not the issue. The issue is that a company precise enough to distinguish 99.8% from 99.7% between detectors has published no sample size, corpus, date, model or threshold behind any of them — and quotes itself inconsistently one page apart.

What is inside the 100,000-text study?

What is actually inside Ryne AI's claimed 100,000-text GPTZero study

This is the centrepiece of Ryne’s case against GPTZero, and the largest claimed sample in this entire market. The post is titled “Why GPTZero is not reliable anymore. We Ran 100,000+ Texts to prove it.” It is dated 27 March 2026 and written by Syed Rajin, Ryne’s CEO.

We read it in full, looking specifically for the study.

The 100,000 texts appear in the headline and in the page’s meta tags. They do not appear in the article. There is no description of the corpus, no account of where the texts came from, no models named, no dates, no thresholds, no explanation of how GPTZero was queried at that volume, and no table, chart or dataset drawn from any original test.

What the post does contain is a summary of other people’s research: Perkins et al. on false positive and false negative rates, Liang et al. at Stanford on non-native English writers, and press coverage of OpenAI’s discontinued classifier. Those are real studies. They are also not Ryne’s, and none of them involves 100,000 texts run by Ryne.

The charts embedded in the post visualise those cited figures. The article closes with the line “The only people who disagree are the ones selling the alarm” — on a page published by a company selling the cure, with no conflict-of-interest disclosure anywhere in the body.

Does it get the borrowed numbers right?

Not all of them, and the one it gets wrong is the one everyone gets wrong.

The post reports 61.3% of non-native English essays being falsely flagged, presented in a way that reads as GPTZero’s own false positive rate. The underlying source is Liang et al. (2023), which ran seven detectors over 91 TOEFL essays and reported roughly 61% as the average across all seven. No per-detector breakdown was ever published. That figure has never been GPTZero’s, and cannot be, because the number does not exist at that granularity.

We apply that rule to ourselves throughout this series, including on pages where the misquote would have helped our argument. It is worth applying here.

What does Ryne’s own answer page say?

Ryne has a post titled, exactly, “Does Ryne AI Bypass GPTZero?” — the same question as this page. Its state is the finding.

It is dated 15 December 2024 and last modified 22 May 2025. It quotes no percentage. It shows no screenshot. It names no GPTZero model version and describes none of GPTZero’s current result behaviour. Its claim is that “Ryne AI’s Humanizer remains the only solution capable of reliably bypassing this enhanced detection”, and it concludes “If you’re looking for a reliable solution that stays ahead of detection technology, Ryne AI’s Humanizer is your ultimate choice.”

And the byline reads “Elana Davies — AI Writer.” A company whose product exists to make AI text read as human has published its own answer to this question under a byline that states the author is an AI. We are not treating that as damning; we are noting that it is the only methodology disclosure on the page.

GPTZero has shipped multiple model revisions since December 2024 — the current build stamps results as Model 4.4b. A page written twenty-one months ago describes a detector that has been rebuilt repeatedly since.

Can you check GPTZero yourself?

GPTZero is free to test - 10,000 characters with no account

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.

This is what makes the missing screenshot conspicuous. Ryne shows GPTZero results for nothing, anywhere on its site, despite GPTZero being the one detector on its list that requires no licence, no institution and no payment to demonstrate.

Has any detector company tested Ryne AI?

Read that table carefully, because it is easy to misread. The column is headed “bypassing ability on naive AI detection methods” — it rates those nine tools against weak detectors, not against GPTZero. The caption states GPTZero’s own position outright: “These methods are all ineffective against GPTZero due to the four-tiered red teaming approach.” That paper is also a GPTZero preprint, not peer-reviewed, and GPTZero has an obvious interest in the conclusion.

None has. Ryne is absent from the appendix table in GPTZero’s February 2026 paper, which names nine bypass services individually. It is absent from the nineteen tools in Pangram’s DAMAGE paper and from the twenty in Pangram’s August 2025 benchmark. Pangram’s July 2026 paper anonymises thirteen humanizers as Commercial A to M, so inclusion there can be neither confirmed nor excluded.

For the same question against the other two detectors we cover, see Ryne AI and Turnitin and Ryne AI and Pangram. The pattern is consistent across all three.

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.

One figure to handle carefully. “GPTZero only catches 46% of humanized text” circulates constantly in humanizer marketing. It comes from Russell, Karpinska and Iyyer and it is one cell of one table: the o1-Pro humanized condition, n=30, where the “humanizer” was a research prompt the authors wrote themselves rather than any commercial tool. In the same table GPTZero scored 100% on paraphrased text and 85.3% overall at a 0.7% false positive rate. We will not quote the 46% without those qualifiers.

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 do the terms say if it does not work?

They say you are unlikely to get your money back. A refund requires all three conditions simultaneously: a request within 3 days, fewer than 500 coins used, and fewer than 1 AI report used. An approved refund then carries a 30% processing fee, and renewals are non-refundable outright.

Ryne AI L.L.C is registered in the Meydan Free Zone, Dubai, licence number 2528619, with terms governed by Meydan Free Zone rules and UAE federal law.

Worth noting the regulatory direction of travel: the FTC has already ordered an AI-detection vendor to stop making accuracy claims without “competent and reliable evidence…at the time it is made.” That standard is about substantiation at the moment of the claim, and it does not obviously distinguish between selling detection and selling evasion.

Methodology. This page reports no first-party test. On 3 September 2026 we read ryne.ai’s homepage and humanizer page, then fetched the “100,000+ texts” post and the “Does Ryne AI Bypass GPTZero?” post at source and audited each against a fixed checklist: claimed sample size, stated corpus, models, dates, thresholds, raw data or tables, author, publication and modification dates, and conflict-of-interest disclosure. We traced every borrowed statistic in the 100,000-text post back to its cited study and checked the attribution. Detector-side facts are quoted from GPTZero’s own FAQ, developers page and February 2026 preprint, and from Pangram’s papers and benchmark post. Pricing, refund and corporate-registration details are quoted from Ryne’s own terms. Our affiliate and sponsorship relationship with Ryne is disclosed at the top.

What should you take from this?

Ryne AI might well beat GPTZero. Plenty of humanizers partly do, and the strongest independent study in the field found GPTZero missing roughly half of humanized text from one commercial tool. Nothing here shows Ryne failing.

What this page shows is that the evidence Ryne offers does not survive being opened. The headline sample of 100,000 is not in the study. The answer page is twenty-one months old, has no number in it, and is bylined to an AI. The two percentages on the site disagree with each other. The one borrowed statistic we could trace is misattributed. And no screenshot exists for the one detector on the list that anybody can run for free.

Do not buy a percentage. Paste your own text into gptzero.me, free, no account, and read what it actually returns — then keep your drafting history regardless of the number, because that is the only thing that survives a detector being wrong.

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 Ryne AI review. Also in this cluster: WriteHuman, StealthGPT, Clarity Bubble.

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.

Grab the free prompts at detectiondrama.com →

Frequently asked questions

Does Ryne AI bypass GPTZero?

Ryne says it does, at 99.8% or 99.9% depending which page you read, and publishes nothing that would let you check either figure. No sample size, no source texts, no test date, no model, no screenshot. GPTZero has never tested Ryne AI, and no independent party has published a result on it. The honest answer is that nobody outside the company has measured it, which is not the same as saying it fails.

What is the 100,000-text study?

A blog post dated 27 March 2026, written by Ryne’s CEO Syed Rajin, titled “Why GPTZero is not reliable anymore. We Ran 100,000+ Texts to prove it.” We read it in full. The figure of 100,000 appears in the title and the page metadata and is never mentioned again. The body describes no corpus, no method, no dates, no models and no thresholds, publishes no table or dataset from any original test, and instead summarises published research by other people — Perkins et al., Liang et al. at Stanford, and press coverage of OpenAI’s discontinued classifier.

So is the study wrong?

The cited research is real; the framing is the problem. A post headlined with an original sample size of 100,000 should contain that study. This one contains a reading list. The distinction matters because the headline is what gets quoted, and it is currently the largest claimed sample in this market.

What is the misattribution?

The post reports 61.3% of non-native English essays being falsely flagged in a way that reads as GPTZero’s own false positive rate. The underlying study is Liang et al. (2023), which tested seven detectors on 91 TOEFL essays and reported roughly 61% as the average across all seven. No per-detector breakdown was ever published, so that number is not GPTZero’s and never was. It is the most frequently misquoted statistic in this field, and we hold the same rule against ourselves.

Doesn’t Ryne have a page answering this exact question?

It does, and its condition is the point. The post titled “Does Ryne AI Bypass GPTZero?” is dated 15 December 2024, last modified 22 May 2025, and carries the byline “Elana Davies” with the job title “AI Writer”. It quotes no percentage, shows no screenshot, names no GPTZero model version, and describes none of GPTZero’s current result behaviour. GPTZero has shipped several model revisions since that post was written.

Has any detector company tested Ryne AI?

No. Ryne is absent from the appendix table in GPTZero’s February 2026 paper, which names nine bypass services individually. It is absent from Pangram’s DAMAGE paper, which named nineteen tools, and from Pangram’s August 2025 benchmark of twenty. Pangram’s July 2026 paper anonymises thirteen humanizers, so its presence there can be neither confirmed nor excluded.

Can I test this myself?

Yes, and it costs nothing. GPTZero takes up to 10,000 characters with no account and no card, and a free account raises that to 150,000 characters. Paste your own humanized output and read the actual result rather than a vendor’s percentage. A genuine current GPTZero result shows a three-class breakdown across human, AI and mixed, a confidence category, and sentence-level highlighting.

What does it cost, and can I get a refund?

A free tier gives 100 coins with a 250-word limit per run, English only, and no AI report. Paid plans are $29.99, $39.99 and $109.99 a month, cheaper annually. Refunds require all three conditions at once: a request within 3 days, fewer than 500 coins used, and fewer than 1 AI report used. An approved refund then carries a 30% processing fee, and renewals are not refundable at all. Ryne AI L.L.C is registered in the Meydan Free Zone, Dubai, licence number 2528619.

About the author. Vlad Ivanov publishes Detection Drama, a site dedicated to AI-detection and text-humanization testing, with 245 published teardowns of detectors and humanizer tools including Turnitin, Pangram, GPTZero and Ryne AI. Profile: LinkedIn. This page reviews published evidence and reports no first-party test; corrections with a source are welcome.