Does StealthGPT bypass Turnitin? On the only independent test that exists, no — the same StealthGPT-humanized document went from 0% to 72% likely-AI once Turnitin’s bypasser detection shipped in August 2025, the largest swing of any tool tested.
Key Takeaways
- In the only independent test of Turnitin’s bypasser detection, StealthGPT went from 0% to 72% likely-AI on the same document — the largest swing of any tool tested (Blommerde, Aug 2025).
- StealthGPT’s counter-evidence is a 30-test case study from July 2024 — published 14 months before Turnitin shipped the detector that caught it.
- Its headline “99% AI-detection bypass rate” lives only in a meta tag, not in visible page copy, with no sample size, date or method attached.
- Its own FAQ lists the detectors it beats — GPTZero, Originality, ZeroGPT, Winston AI — and Turnitin is not on that list.
- A comparison table on its site rates its Turnitin performance “Strong” based on “published independent tests” that are never cited or linked.
- The two rigorous studies everyone cites — Pangram’s benchmark and the UChicago BFI paper — never tested Turnitin at all.
- In August 2026 StealthGPT was openly recruiting 30–40 paid creators at $1,150+/month to target US college students, and asking them to upvote the recruitment post.
Want to bypass Turnitin in 2026? Grab the free prompt pack.
Get the exact text-humanization prompts I use to drop an AI score by hand — copy, paste, submit. Free, straight to your inbox.
Send me the free prompts →Does StealthGPT bypass Turnitin?
Not according to the one measurement anybody has published. And StealthGPT is an unusually clean case to examine, because both sides of the argument are on the record with numbers — they are just fourteen months apart, and only one of them was taken after Turnitin changed.
The structural problem is the same one that makes this whole question hard: Turnitin has no public API and no self-serve tier. Only institutional licence-holders can test it. That is why a tool can be benchmarked against GPTZero a dozen times and against Turnitin essentially never. We hit the identical wall on Undetectable AI, GPTHuman and SuperHumanizer.
This page reviews the published record. It reports no first-party test, and every claim below is dated and attributed.
What does StealthGPT claim, and where?
The headline number is a “99% AI-detection bypass rate” — and it is worth knowing where that sentence actually lives. It sits in the site’s meta description, the text search engines display, rather than in the visible page copy. No sample size, no test date, no per-detector breakdown, no methodology.
A homepage carousel goes further, claiming “Turnitin: 98% human average (0/30 detections)”. That item carries no date. A comparison table elsewhere on the same page rates StealthGPT’s Turnitin performance as “Strong” against rivals rated “Partial”, footnoted as being “based on published independent tests and user reviews as of mid-2026.” Those independent tests are never named, cited or linked. We looked for them.
Pricing runs $9.99/week for 50 requests and $19.99/week for 100, both currently promoted at $1 for a first week, plus a student offer at 99 cents. There is no free humanizing tier beyond a 100-word box. Our full StealthGPT review covers the product itself.
What did the only independent Turnitin test find?
In late August 2025, Dr Tadhg Blommerde — an Assistant Professor at Northumbria University with institutional access to Turnitin Originality — re-uploaded a set of humanized documents that had all previously scored 0% likely-AI, to see what Turnitin’s newly shipped bypasser detection made of them.
StealthGPT’s document came back at 72% likely-AI, the largest change in the run. The results were reported by EdTech Innovation Hub in September 2025.
| Tool | Before update | After update | Read |
|---|---|---|---|
| StealthGPT | 0% | 72% | Largest swing of any tool tested |
| Groby | 0% | 67% | Second-largest |
| GPTHuman | 0% | 31% | Caught, but less severely |
| StealthWriter | 0% | 1–19% | Unspecified low band |
| Refrazy | 0% | 1–19% | Unspecified low band |
| Easy Essay | 0% | 0% | Unchanged |
Two caveats belong with that number and are usually dropped. It is one document, one pass, one tester, with no sample text published, so nothing is reproducible. And the tester was careful not to oversell his own result.
“The new AI bypasser detector is an improvement, but it’s not perfect… Totally accurate AI detection is a myth. If this feature worked well, every single document we tested would have got a 100% likely AI score.”Dr Tadhg Blommerde, Northumbria University — September 2025
He also predicted the tools would respond with updates quickly. Nobody has published a retest since, so whether that happened is genuinely unknown.
Why doesn’t StealthGPT’s own case study answer this?
Because of when it was run. StealthGPT published a Turnitin case study reporting 30 tests — ten each across three engines, using 2,000-word essays — with scores clustered at 100% human. On its own terms that is a far larger sample than the independent test, and the company deserves credit for publishing a method at all.
The date undoes it. The case study is dated 3 July 2024. Turnitin shipped its dedicated AI bypasser detection on 27 August 2025 — roughly fourteen months later. The case study is an accurate description of a detector that has since been rebuilt specifically to catch what it was measuring.
Worth noting: StealthGPT’s Turnitin blog post has been modified as recently as 2026 and still does not mention the bypasser update at all.
What about the two big studies everyone cites?
Both are real, both are serious, and neither one touches Turnitin. This gets misreported constantly, so it is worth being exact.
Pangram’s August 2025 benchmark scored 20 humanizers and put StealthGPT at 95.6% detection accuracy — mid-table, below Undetectable AI’s 90.3% and TwainGPT’s 92.7% in terms of evading it. Three caveats: Pangram is a commercial detector vendor benchmarking tools it competes against; the post discloses no sample size, source texts or false-positive rate; and the detector it measures is Pangram, not Turnitin. It is also a roughly twelve-month-old number against a model Pangram has since superseded twice.
The UChicago BFI working paper (Jabarian and Imas, August 2025) is the most rigorous StealthGPT test in existence — 1,992 human passages with matched AI corpora, run through StealthGPT’s default rewrite endpoint. It found Pangram largely robust to humanizing while GPTZero “largely loses its capacity to detect AI-generated text.” It evaluated Pangram, Originality.ai, GPTZero and RoBERTa. Turnitin appears in it zero times.
The honest summary: StealthGPT reliably beats GPTZero, partially beats Originality.ai on short text, and fails against Pangram — which we cover in does StealthGPT bypass Pangram. None of that tells you what Turnitin will do. The divergence itself is the point, and we track it in how much AI detectors disagree.
What does the Reddit footprint actually show?
Something more useful than testimonials. On 19 August 2026, StealthGPT posted a recruitment ad seeking 30–40 paid creators at “$1,150+/month”, explicitly wanting people who “know how to target US college/university students” — and instructing applicants to “Upvote this post for visibility.” That is a dated, first-party record of an active paid-amplification programme aimed at students.
Meanwhile its own brand subreddit is effectively dormant: of the most recent 25 posts, 24 date from 2024, with the last vendor post in November of that year. And in a live August 2026 thread where students swapped methods for beating Turnitin, StealthGPT — the market leader by user count — was not mentioned once. The recommendations were manual rewriting and translation round-trips.
Read those three facts together and the pattern is clear enough: paid promotion is buoyant, organic advocacy is not.
How should you read a bypass guarantee?
Four questions do most of the work. How old is the number? Anything measured before September 2025 describes a Turnitin that no longer exists. Was the detector real or a proxy? Only institutional licence-holders can run Turnitin; everyone else substitutes. Who published it? Vendor, competitor and affiliate all carry a stake. What does the guarantee actually promise? “Money back if you get detected” sounds firm until you look for the claim window, the evidence requirement and the process — none of which are published here.
Then reduce the risk you control. Check before you submit rather than after — see the best pre-submission check for Turnitin AI risk and the Turnitin self-check walkthrough. Keep your drafting history; it costs nothing and is the strongest defence against a false accusation, which matters most for the writers detectors treat worst — see ESL writers and AI detection and false positives for neurodivergent students. And strip the obvious tells by hand first: what to remove before using an AI humanizer, then what makes an essay sound too polished.
For the wider picture, humanizers and Turnitin bypasser detection and bypass rates after humanization collect what the update actually caught.
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 StealthGPT bypass Turnitin?
Not on the only independent evidence available. When Turnitin’s AI bypasser detection shipped in August 2025, a StealthGPT-humanized document that had previously scored 0% likely-AI was re-scored at 72% likely-AI. That was the largest change of any tool in the test.
What does the 0% to 72% actually measure?
The same document, scored twice: once by Turnitin before its bypasser update and once after. It is not a raw-AI-versus-humanized comparison. It measures what changed on Turnitin’s side, on identical text.
Isn’t there a StealthGPT case study showing 0 detections?
Yes, and the date is the problem. StealthGPT published a case study reporting 30 tests across three engines with essentially no Turnitin detections — but it is dated 3 July 2024, roughly 14 months before Turnitin shipped the bypasser detection that later flagged its output at 72%. It describes a version of Turnitin that no longer exists.
Where does the 99% bypass rate claim come from?
StealthGPT’s own site — specifically its meta description, not its visible body copy. No sample size, test date, detector breakdown or methodology is attached to it anywhere we could find.
Does StealthGPT say it beats Turnitin?
Inconsistently. Its homepage carousel claims a “Turnitin: 98% human average (0/30 detections)” with no date, and it links a 2024 blog post titled around beating Turnitin. But its own FAQ answer listing which detectors it outsmarts names GPTZero, Originality, ZeroGPT and Winston AI — and omits Turnitin.
What about the Pangram and UChicago studies?
Neither tested Turnitin. Pangram’s August 2025 benchmark scored 20 humanizers against Pangram and put StealthGPT at 95.6% detected. The UChicago BFI paper used StealthGPT as its humanizer across 1,992 passages, but evaluated Pangram, Originality.ai, GPTZero and RoBERTa. Turnitin does not appear in it once.
Is the money-back guarantee reliable?
It is stated but not specified. The homepage FAQ says subscriptions come with a money-back guarantee if you get detected. We found no published claim window, evidence requirement, exclusion list or refund process on the homepage, pricing page or affiliate page.
Has anyone retested StealthGPT against Turnitin since 2025?
Not that we could find, from anyone — independent, academic or commercial. The tester who ran the original predicted the humanizers would update quickly to defeat the new detector. Nobody has checked whether they did.
