Does GPTHuman bypass Turnitin? On the only independent test that exists, no — Turnitin scored GPTHuman-humanized text at 31% likely-AI once its bypasser detection shipped in August 2025, up from 0% under the previous model.
Key Takeaways
- Exactly one independent test of GPTHuman against real Turnitin exists: Turnitin scored GPTHuman-humanized text at 0% before its bypasser update and 31% likely-AI after (Blommerde, 27 Aug 2025).
- That put GPTHuman mid-table — better than StealthGPT (72%) and Groby (67%), worse than the tools that stayed under 20%.
- GPTHuman names Turnitin twice on its homepage but publishes no bypass percentage at all — no sample size, no date, no method.
- Its guarantee reads “Bypass all AI Detectors guarantee*”. The asterisk has no footnote, and no refund policy is published anywhere on the pricing page.
- GPTHuman is absent from both serious humanizer benchmarks — Pangram’s 20-tool table and the DAMAGE paper’s 19-tool study.
- A tester who is GPTHuman’s own affiliate published a result showing its output flagged 100% AI-generated by Pangram.
- The one measurement anyone has is nearly a year old and has never been repeated. Nobody has retested GPTHuman against Turnitin since.
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 GPTHuman bypass Turnitin?
Not cleanly, and the evidence base is thinner than the marketing implies. One person has tested GPTHuman against real Turnitin and published the result. One. Everything else you will find is a vendor blog, a competitor’s teardown, or a proxy detector standing in for a product the reviewer could not access.
That scarcity is structural rather than suspicious. Turnitin has no public API and no self-serve tier, so the only people who can test it are institutional licence-holders. It is the reason this question stays open while the same tools get benchmarked against GPTZero a dozen times over. We hit the same wall on Undetectable AI and SuperHumanizer.
What makes GPTHuman different is that a real measurement does exist. It is just not the one the vendor talks about.
What does GPTHuman actually claim?
It names Turnitin twice on its homepage. Verbatim: humanized text is protected “from false positives on leading AI detection tools, including GPTZero and Turnitin,” and its models are “constantly re-trained and fine-tuned to stay ahead of the evolving AI detection landscape, including GPTZero, Turnitin, Originality.ai, Copyleaks, and more.”
What it never does is put a number on it. There is no bypass rate, no success percentage, no sample size and no test date anywhere on the site. In a market where every rival leads with a figure, the absence is conspicuous — and, to be fair, more honest than inventing one.
Pricing is a 300-word free trial, then $15/month for 25,000 words, $25/month for 60,000, and $49/month for unlimited — capped at 2,500 words per output. Our full GPTHuman review covers the product itself.
What did the only independent Turnitin test find?
On 27 August 2025, Dr Tadhg Blommerde — an Assistant Professor at Northumbria University with institutional access to Turnitin Originality — re-uploaded humanized documents that had previously scored 0% likely-AI, to see what Turnitin’s newly shipped bypasser detection made of them.
GPTHuman’s document came back at 31% likely-AI. The result was reported by EdTech Innovation Hub alongside the rest of his run.
| Tool | Before update | After update | Read |
|---|---|---|---|
| StealthGPT | 0% | 72% | Largest swing of any tool tested |
| Groby | 0% | 67% | Second-largest |
| GPTHuman | 0% | 31% | Mid-table — caught, but least badly of the three |
| StealthWriter | 0% | 1–19% | Turnitin returned an unspecified low band |
| Refrazy | 0% | 1–19% | Same unspecified low band |
| Easy Essay | 0% | 0% | Unchanged |
Three caveats belong with that table, and most pages quoting it leave them out. It is one document per tool, a single pass, with no sample text published, so nothing is reproducible. The flagged content carried no bypasser-specific label — it was simply classified as AI-generated. And the tester said so himself.
“The new AI bypasser detector is an improvement, but it’s not perfect… Totally accurate AI detection is a myth.”Dr Tadhg Blommerde, Northumbria University — September 2025
Why doesn’t the vendor’s own test settle it?
GPTHuman published its own Turnitin result in July 2026: a 300-word block that scored 100% AI before humanizing and 0% after. Read the method and it stops being evidence. Sample size is one. The Turnitin run was performed by an unnamed “colleague… a professor at a university.” No source text, no prompt, no settings, no way for anyone to reproduce it.
A second vendor page is harder to defend. It concludes that “the best AI humanizer for Turnitin in this test was GPTHuman, full stop” — in an article whose testing only ran GPTZero. Turnitin was never involved. A third page publishes GPTHuman’s own internal detector scores as evidence of Turnitin performance, while conceding that unless you are a teacher or administrator, “you don’t have direct access to its originality report.”
That last admission is the honest one, and it applies to the whole category. A tool’s built-in checker is not Turnitin and has a commercial reason to show you green.
What do other detectors say about the same tool?
They disagree sharply, which is the most useful thing about them. The most striking result comes from a tester who is GPTHuman’s own affiliate: in a July 2026 four-tool comparison carrying GPTHuman affiliate links and an explicit disclosure, GPTHuman’s output was flagged 100% AI-generated by Pangram, while Quetext put it at 80% likely human. Publishing a failure you are paid to promote is worth more than a dozen glowing reviews.
A competing humanizer reported Originality.ai flagging GPTHuman output at 97% AI — treat that as adversarial, since the publisher sells a rival. Another rival tested it against “a system built to act like Turnitin,” which is a proxy, not Turnitin, and produced no numbers at all.
The spread — 31% on Turnitin, 100% on Pangram, 97% on Originality, 80% human on Quetext — is the real finding. We track that divergence in how much AI detectors disagree, and the Pangram-specific question in does GPTHuman bypass Pangram.
What do actual users report?
Very little, and what exists is not flattering. The humanizer corner of Reddit is heavily seeded — we found a subreddit created in December 2025 whose GPTHuman posts all follow one listicle template and all rank GPTHuman first, one of which pads its comparison with tool names that do not appear to exist.
Filter that out and the organic record is thin. The clearest first-hand account comes from a student working on a dissertation in May 2026, who reported plainly that GPTHuman “couldn’t get past it either — both detected ai.” The same user flagged the astroturf problem unprompted, noting she kept seeing bots promoting a different tool. A separate August 2026 complaint describes a refund being refused after the company acknowledged the problem, which sits awkwardly beside a guarantee with no published refund policy.
How should you read a humanizer guarantee?
Four questions strip most of this market down in under a minute. Is there a number, and what is behind it? GPTHuman publishes none, which is at least not a fabricated one. Was the detector real or a proxy? Only institutional licence-holders can run Turnitin; everyone else is substituting. How old is the test? The one real GPTHuman measurement is from August 2025 and has never been repeated. Who published it? Vendor, competitor and affiliate all have a stake — though as this case shows, an affiliate publishing a failure is a strong signal.
Then reduce the risk you can actually control. Check before you submit, not after — see the best pre-submission check for Turnitin AI risk and the Turnitin self-check walkthrough. Keep your drafting history, which costs nothing and is the strongest defence against a false accusation; it matters most for the writers detectors treat worst, covered in ESL writers and AI detection and false positives for neurodivergent students. And fix 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 on which tools the bypasser update actually caught, see humanizers and Turnitin bypasser detection and bypass rates after humanization.
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 GPTHuman bypass Turnitin?
Not reliably, on the only evidence that exists. In the single independent test on record, Turnitin scored GPTHuman-humanized text at 31% likely-AI once its bypasser detection shipped in August 2025 — down from 0% under the older model. That is a partial failure, not a clean pass, and it has not been retested since.
What does the 31% actually mean?
It is Turnitin’s confidence that 31% of the submitted text is AI-generated. It is not a bypass rate and not a pass/fail. The same document had scored 0% on Turnitin before the bypasser update, so the number measures what changed on Turnitin’s side, not what GPTHuman did differently.
Does GPTHuman claim a specific success rate?
No. Unusually for this market, it publishes no percentage at all. It says “Guaranteed undetectable AI” and names GPTZero, Turnitin, Originality.ai and Copyleaks, but attaches no number, sample size, test date or methodology to any of it.
What is GPTHuman’s guarantee worth?
Read it literally. The pricing page says “Bypass all AI Detectors guarantee*” and the asterisk leads nowhere — there is no footnote. No refund policy is published. The only remedy stated anywhere is the homepage line: if your content is flagged, “we’ll rewrite it for free.”
How much does GPTHuman cost?
There is a 300-word free trial. Paid plans are $15/month for 25,000 words (1,000 per output), $25/month for 60,000 words (1,500 per output) and $49/month for unlimited words (2,500 per output), with roughly 45% off on annual billing.
Has GPTHuman been tested by academics?
It has not appeared in either of the two serious humanizer benchmarks. Pangram’s August 2025 table scored 20 tools and GPTHuman is not among them; the DAMAGE paper studied 19 humanizers and paraphrasers and does not mention it once.
Why do different detectors give such different answers?
Because they are different models trained on different data with different thresholds. The same GPTHuman output has been reported at 31% likely-AI by Turnitin, 100% AI-generated by Pangram and 97% AI by Originality.ai. A single score from any one tool is not proof of anything.
Is the Turnitin result still current?
No, and that is the honest limitation. The test dates from 27 August 2025. Turnitin has shipped updates since, GPTHuman advertises “constantly re-trained” models, and nobody has published a retest. Anyone quoting a 2026 GPTHuman-vs-Turnitin number is quoting something that does not exist.
