rel=nofollow citations. The promoted call-to-action above is a StealthWriter affiliate placement carrying its own disclosure, and is unrelated to this review.Does SuperHumanizer bypass Copyleaks? Nobody has tested it and the vendor has never claimed it — so the only number that bears on the question comes from the study Copyleaks puts at the top of its own evidence page.
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
- SuperHumanizer has never mentioned Copyleaks. Not on its homepage, not on its pricing page, not in its Terms, and not in the public disclosure where it did name ZeroGPT and GPTZero as development targets.
- No detector company has tested it either. It is absent from GPTZero’s nine named services, from Pangram’s nineteen-tool audit and from Pangram’s twenty-tool benchmark.
- So there is no claim to check and no test to report. What is left is the detector — and Copyleaks is the best-evidenced detector in this entire series, with eighteen third-party citations including real peer-reviewed work.
- Read that list closely, though. It counts a press-release wire item and a four-document consumer blog test alongside journal articles, and its flagship citation is misattributed to Cornell University.
- The four authors of that paper are at Toronto Mississauga, Maranatha Christian University and Escuela Superior Politécnica del Litoral. arXiv is run by Cornell Tech. Hosting a preprint is not writing one.
- And the same paper contains the number the marketing skips: Copyleaks resilience went from 100% before a QuillBot paraphrase to 50% after.
- That is the honest answer to the title question. Nobody has tested this pairing, but the one published paraphrase-resilience figure Copyleaks has is 50% — and you can scan 25,000 characters against it for free, without an account.
Does SuperHumanizer bypass Copyleaks?
This cell is empty, and saying so plainly is more useful than filling it.
We read superhumanizer.ai’s homepage, pricing page and Terms of Service at source and searched each for “Copyleaks”. It does not appear on any of them. Nor does it appear in the vendor’s public development disclosure — the one place SuperHumanizer has ever been specific about detectors, where it named ZeroGPT and GPTZero as tools it built the product against.
No detector company has filled the gap either. SuperHumanizer is absent from the nine bypass services named in GPTZero’s February 2026 paper, from the nineteen humanizers audited in Pangram’s DAMAGE paper, and from the twenty scored in Pangram’s August 2025 benchmark.
So there is no vendor claim to audit and no independent test to report. This page reports no first-party test either; everything below was fetched on 21 September 2026. What is left is the detector, and on that side there is a great deal to say.
What does the vendor actually promise?
Something broad enough to cover Copyleaks and narrow enough to commit to nothing.
The homepage says humanized text “can easily pass AI detection tools like Turnitin, GPTZero, and Originality.ai”. The pricing tiers — Basic at $12 a month for 15,000 words, Pro at $19 for 50,000 — each carry “Bypass All AI Detectors including GPTZero”. A homepage chart claims 97% against 68% for “Basic Humanizers” and 45% for “Other Tools”, with no source, method, sample size or date for any of the three bars.
How good is Copyleaks compared with the rest?
Here is where Copyleaks is genuinely unlike every other detector in this series, and it deserves saying before the criticism: it has the strongest third-party evidence base of any detector we have covered. Its blog maintains a running citation list of eighteen items, and a real share of them is peer-reviewed academic work rather than vendor content — Walters in Open Information Science (2023), Chaka in the Journal of Applied Learning & Teaching (2023 and 2024), Kar et al. in the Indian Journal of Psychological Medicine (2024), and an internal evaluation by Boston University’s own AI task force. Chaka’s 2024 study of thirty detectors found Copyleaks one of only two that correctly classified every human-written student essay.
No other detector in this project can point to anything close. Now the parts of that list that do not hold up.
We catalogued all eighteen entries by what kind of object each one actually is. The result is a real academic core with a soft edge:
| What the source actually is | Count | What it can establish |
|---|---|---|
| Peer-reviewed journal article | 7 | The load-bearing part of the list, and it is substantial |
| Preprint, not peer reviewed | 3 | Real research, not yet checked by reviewers |
| Student thesis | 2 | Supervised work, narrow scope, rarely replicated |
| Conference paper | 1 | Lighter review than a journal |
| Institutional evaluation (Boston University) | 1 | A buyer’s internal assessment, which is its own kind of useful |
| Book or edited volume | 1 | Depends entirely on the chapter |
| Consumer tech article, 4 documents | 1 | Very little |
| Press-release wire item | 1 | Nothing. This is not a study. |
| Venue not stated | 1 | Uncheckable as listed |
Seven peer-reviewed journal articles is a genuinely strong showing — more than every other detector in this project combined. The criticism is not that the evidence is thin. It is that a press release and a four-document blog test are counted in the same total, which makes the number bigger and the average weaker.
What is wrong with the headline claim?
Copyleaks advertises “over 99% accuracy and an industry-low .03% false positive rate”, and for English specifically 99.97% on human text and 99.20% on AI text.
And that same paper contains the figure that actually bears on bypassing, which the marketing does not quote.
Copyleaks resilience, before paraphrasing: 100.00%. After a QuillBot paraphrase: 50.00%.Orenstrakh, Karnalim, Suárez and Liut, arXiv:2307.07411, July 2023
Half the detections gone to one commodity paraphraser. For context the same table puts GPTZero at 70% before and 20% after, so Copyleaks still finishes ahead — but “best in the room” and “holds up” are different claims, and only the first is supported. That result is three years old and both products have changed since, which cuts both ways: it is the only published paraphrase-resilience figure Copyleaks has, and nobody has refreshed it.
Why does that 50% matter for this question?
Because it is the closest thing to an answer that exists, and because of what produced it.
The paraphraser in that test was QuillBot — a mainstream writing tool, not a purpose-built bypass service, used at its ordinary settings. Halving a detector’s resilience did not require a $19-a-month humanizer. It required a free paraphrase.
| Detector | Resilience before paraphrase | After QuillBot paraphrase | Detections lost |
|---|---|---|---|
| Copyleaks | 100.00% | 50.00% | Half |
| GPTZero | 70.00% | 20.00% | Over two thirds |
Two things follow, and they pull in opposite directions.
Against the detector: Copyleaks’ own flagship citation shows it losing half its detections to commodity paraphrasing, and its marketing quotes the accuracy rows from that paper without the resilience row.
Against the tool: none of this is evidence that SuperHumanizer specifically does anything to Copyleaks. QuillBot is not SuperHumanizer. A tool with no published test, no vendor claim and no independent score does not inherit another product’s result. Anyone reading this page as proof SuperHumanizer beats Copyleaks is reading it wrong, and that includes anyone quoting it.
Can you check Copyleaks yourself?
Copyleaks is the most generous free tier of any detector in this project. Its own page states you can “scan up to 25,000 characters without logging in, a higher limit than most major AI detection tools” — against 15,000 at ZeroGPT and 10,000 at GPTZero. A free account unlocks additional monthly credits.
A genuine current Copyleaks result separates the human-written and AI-written parts of a mixed document rather than returning one number, and surfaces what it calls “AI Phrases”: text “displaying the statistical frequency with which it originates from AI, rather than human-generated sources.”
So there is no access excuse on this detector either, and less of one than anywhere else. Any vendor claiming to beat Copyleaks can demonstrate it on 25,000 characters, for nothing, without an account.
Which is the practical end of this page. There is no published test of this pairing, the detector’s only paraphrase-resilience figure is three years old, and the version of the question that actually affects you — your text, today, on the settings you are using — costs nothing and takes a minute.
What should you take from this?
That an empty cell is information, if you say what is missing rather than filling it with adjectives.
SuperHumanizer has never claimed to beat Copyleaks. No detector company has tested it. The one figure in the vicinity belongs to a paper about a different paraphraser, published three years ago, cited by the detector itself under a university name that is not on it.
What that figure does establish is narrow and worth keeping: the best-evidenced detector in this market has exactly one published result on rewritten text, and it is 50%. Everything else in Copyleaks’ considerable evidence base measures raw AI output, which is not the thing anyone using a humanizer is submitting.
So test it yourself. Copyleaks gives you 25,000 characters free without an account, which is more room than any rival, and then run the same text through a second detector — 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 SuperHumanizer review, and the same question against Turnitin and GPTZero. Also in this cluster: ZeroGPT, Pangram.
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 SuperHumanizer bypass Copyleaks?
Nobody has published a test, including the vendor and including us. SuperHumanizer does not mention Copyleaks anywhere on its website or in its Terms, and it did not name Copyleaks in the public disclosure where it did name ZeroGPT and GPTZero as tools it built the product against. No detector company has tested SuperHumanizer against anything. The honest answer is that this cell is empty, and the useful information is on the detector side.
Has SuperHumanizer ever claimed to beat Copyleaks?
No. Its homepage claims humanized text can easily pass AI detection tools like Turnitin, GPTZero, and Originality.ai. Its pricing tiers say Bypass All AI Detectors including GPTZero. Copyleaks is named in neither, which is unusual given that Copyleaks is one of the most widely deployed detectors in education. The generic Bypass All AI Detectors phrase would cover it, but a phrase that covers everything establishes nothing.
How good is Copyleaks, really?
On raw AI text, better evidenced than any other detector in this series. Its citation list runs to eighteen items and includes peer-reviewed work: Walters in Open Information Science, Chaka in the Journal of Applied Learning and Teaching, Kar et al. in the Indian Journal of Psychological Medicine, plus an internal evaluation by Boston University’s own AI task force. Chaka’s 2024 study of thirty detectors found Copyleaks one of only two that correctly classified every human-written student essay.
What is wrong with that citation list?
It mixes categories. Alongside the journal articles it counts a press-release wire item from a financial-content syndication site and a consumer tech article that tested four documents, plus several student theses and unreviewed preprints. Those are not the same kind of object as peer-reviewed research, and presenting them in one undifferentiated list of third-party studies inflates the apparent weight of the whole.
Is the Cornell attribution accurate?
No. The page reads: According to Cornell University, the answer is Copyleaks AI Detector. In July 2023, a study published on arXiv by Cornell Tech researchers declared Copyleaks AI Detector the most accurate. The paper is arXiv:2307.07411, and its four authors are at the University of Toronto Mississauga, Maranatha Christian University and Escuela Superior Politecnica del Litoral. None is at Cornell. arXiv is operated by Cornell Tech, which makes Cornell the host, not the author. The study itself is real and its finding does favour Copyleaks.
What does that study say about paraphrased text?
The part the marketing does not quote. Its resilience test put Copyleaks at 100.00% before a QuillBot paraphrase and 50.00% after. GPTZero went from 70% to 20% in the same table, so Copyleaks still finished ahead, but half its detections were lost to one commodity paraphraser. That result is from July 2023 and both products have changed since. It is also the only published paraphrase-resilience figure Copyleaks has, and nobody has refreshed it.
Can I test Copyleaks myself for free?
Yes, and more generously than any other detector we cover. Copyleaks states you can scan up to 25,000 characters without logging in, against 15,000 at ZeroGPT and 10,000 at GPTZero. A free account unlocks more monthly credits. A genuine result separates the human-written and AI-written parts of a mixed document rather than returning one number, and flags AI Phrases showing the statistical frequency with which each passage originates from AI.
What do SuperHumanizer’s Terms commit to?
Not to detection outcomes, against Copyleaks or anything else. The Terms state that output is not guaranteed to be always undetectable by AI detection tools, and that the service does not guarantee output will meet expectations including detection outcomes. They are dated 23 September 2025, operated by Kamranali Limited, and the refund is seven days with fewer than 2,000 words used, which matches the 7-Day Money Back Guarantee advertised on the pricing page.
