Independent studies, all run before that update, found edited AI text often lowered detection rates. Nobody outside Turnitin knows which humanizers get caught, so I wouldn't bet a grade on one.
| Question | What the sources say (read 29 October 2026) |
|---|---|
| Does Turnitin detect humanizers? | Turnitin says its English model detects AI text modified by "leading" bypassers, since 27 August 2025 |
| Does it detect paraphrasers? | Yes, according to Turnitin, since 6 December 2023; it names QuillBot as an example word spinner |
| Which humanizers? | Turnitin will not name them; it says naming them "would make it easier for students to evade our system" |
| Languages | English only for paraphrase and bypasser detection |
| How it appears in the report | Folded into the "AI-generated" category and highlighted in blue |
| What humanizer vendors claim | Phrasly: content "that passes authenticity checks"; Undetectable AI: "scoring as human on every major detector" |
| What independent research found | Edits lowered detection rates in 2023 tests; a 2024 study found Turnitin had the largest accuracy drop of seven detectors |
| Is the score proof? | No. Turnitin says it should not be the sole basis for adverse action |
Can Turnitin detect AI humanizers?
Turnitin's answer is yes, for English text. Its AI writing detection FAQ says the detector can "identify instances where AI-generated text may have been modified by AI paraphraser or bypasser (also called humanizers) tools to evade detection." It says this runs automatically on every submission at institutions that have AI writing detection enabled, with no setting to change.
The claim is specific about what it covers:
- AI text that was then modified. The model looks for text that was "likely AI-generated" and then changed by a tool. It does not claim to detect a humanizer run on text a person wrote.
- English only. The FAQ says "our AI paraphrasing and bypassing detection capabilities are only available for English submissions." Spanish, Japanese and the other supported languages get standard AI detection only.
- Leading tools, unnamed. Turnitin says the model "has been trained and tested to detect leading paraphraser and bypasser tools" but declines to list them.
So when someone asks whether Turnitin catches a particular humanizer, the honest answer is that Turnitin does not say. It says it trains against the popular ones and keeps updating.
My take: Turnitin's refusal to name tools is deliberate, and it works in Turnitin's favor. A student cannot know which side of the line a tool sits on this term, so every "it passed for me" post online describes a model that may already have changed.
Turnitin's AI bypasser and paraphrase detection
Turnitin added the two capabilities in stages. Its AI writing model release notes record the history:
| Date | Change Turnitin documented |
|---|---|
| 6 December 2023 | Detection of likely AI-generated text "even if it may have been paraphrased using an AI word spinner" |
| 16 July 2024 | Scores from 1% to 19% replaced by an asterisk to reduce false positives |
| 27 August 2025 | AI bypasser detection added; the "AI-generated only" category now includes text "that may have been modified by an AI bypasser tool" |
| 14 October 2025 | Model update to improve recall while keeping a low false positive rate |
| 12 February 2026 | Model update to improve recall while keeping a low false positive rate |
| October 2026 | New English model architecture to improve accuracy at transitions between human and likely AI text, per the FAQ |
Turnitin announced the bypasser feature in a 27 August 2025 press release. Its chief product officer, Annie Chechitelli, described humanizer companies as "a new category" of "cheating providers" that "exist to profit from students' misuse of AI." The release says the feature is available to institutions that license Turnitin Originality or the iThenticate 2.0 AI add-on, and that it "has been trained and tested to support interactions in English only."
In the report itself, Turnitin's guide to using the AI Writing Report says AI text "further modified by an AI bypasser, AI-paraphrasing tool or AI word spinner, such as Quillbot" is highlighted in blue as part of the AI-generated category. An instructor does not see a separate "humanizer" label. The AI writing indicator is also separate from the similarity score, which measures text that matches other sources; my Turnitin similarity score guide explains how to read that number.
Two rules from the release notes matter for anyone reading an old report. Updates "will not retroactively update previously-generated AI writing reports," and a paper is only rescored if it is resubmitted. A submission that scored low in 2024 could score differently on today's model.
For how the underlying classifier works, see my guide to how AI detectors work. For the rest of Turnitin's AI checks, including model coverage and file rules, read does Turnitin check for AI.
What changes between a paraphraser, a humanizer and your own edit
People use these words loosely, so here is how I separate them.
| Kind of change | What it does to the text | What Turnitin says it targets |
|---|---|---|
| Your own editing of your own draft | Changes ideas, structure, examples and wording because you rethought them | Nothing; this is ordinary writing |
| Grammar and spelling fixes | Corrects errors, leaves sentences mostly intact | "In most cases" not flagged, per Turnitin's Grammarly tests |
| AI paraphraser or word spinner | Swaps words and reorders clauses in existing text | AI text paraphrased this way, since December 2023, English only |
| AI humanizer or bypasser | Rewrites AI text to change its statistical profile, often marketed against detectors | AI text modified this way, since August 2025, English only |
Example. This is a pair I wrote myself to show the difference between a word-swap and a real edit. Neither came from a tool.
- Original AI-style sentence: "Social media has had a significant impact on the mental health of teenagers in recent years."
- Word-swap version: "In recent times, social platforms have had a considerable effect on adolescents' psychological well-being."
- A real edit by the author: "The study I used surveyed 1,200 students, and the link it found was with late-night scrolling, not total screen time. That is the part I want to argue about."
The word-swap keeps the same generic claim in the same order, which is exactly the pattern a paraphrase detector is trained on. The real edit adds a source, a specific finding and a point of view. That is the version a teacher wants, and it is also the only one that is actually yours.
What independent studies show
Independent research on edited AI text is mostly from 2023, before Turnitin's bypasser update. It shows that detection has been fragile, and that both sides keep changing.
Perkins and colleagues, 2024. The team tested seven detectors, including Turnitin, on 805 samples in September and October 2023, including AI text modified with simple manipulation techniques. Mean accuracy was 39.5% on unedited AI text and fell to 22.1% after manipulation. Turnitin detected 61% of unedited AI samples, second of seven, but had the largest drop, 42.1%. The authors also noted that some edits that lowered detection produced text that "would very likely receive poor marks." The paper is in the International Journal of Educational Technology in Higher Education.
Krishna and colleagues, NeurIPS 2023. A research paraphraser cut DetectGPT's detection rate from 70.3% to 4.6% and also weakened GPTZero, OpenAI's classifier and a watermark. The same paper, Paraphrasing evades detectors of AI-generated text, but retrieval is an effective defense, showed that checking text against a stored log of model outputs caught 80% to 97% of the paraphrased text while flagging 1% of human text.
Sadasivan and colleagues. Their Transactions on Machine Learning Research paper, Can AI-Generated Text be Reliably Detected?, found that repeated paraphrasing lowered detection across several detector types with only slight loss of quality, and that watermarks can be spoofed to make human text look like AI.
Weber-Wulff and colleagues, 2023. In a 14-tool study that included Turnitin, the authors found that "content obfuscation techniques significantly worsen the performance of tools," and concluded that the tools are "neither accurate nor reliable."
What these studies do not show: how today's Turnitin model performs on today's humanizers. As of 29 October 2026 I know of no published independent study of Turnitin's post-August 2025 bypasser detection. Until one exists, I treat both the vendor claims and Turnitin's claims as claims, and so should you.
| Study | When the tests ran | Main finding on edited AI text |
|---|---|---|
| Perkins et al., 2024 | September to October 2023 | Mean accuracy fell from 39.5% to 22.1%; Turnitin had the largest drop, 42.1% |
| Krishna et al., NeurIPS 2023 | 2023 | Paraphrasing cut DetectGPT from 70.3% to 4.6%; retrieval caught 80% to 97% |
| Sadasivan et al., TMLR | 2023 | Repeated paraphrasing lowered detection across detector types |
| Weber-Wulff et al., 2023 | 2023 | Obfuscation "significantly" worsened all 14 tools |
Can Turnitin detect StealthWriter, Phrasly, Undetectable AI?
Turnitin does not say, for any named tool. Its FAQ declines to list the bypassers it trains against. Here is what each vendor's own home page says, read on 29 October 2026, and what is actually documented.
| Tool | What the vendor says | What Turnitin documents |
|---|---|---|
| StealthWriter | Rewrites AI-generated content "so it reads naturally human," with a built-in AI detector to scan before and after | No tool named; detection of "leading" bypassers in English |
| Phrasly | Humanized content "that passes authenticity checks"; a premium tier advertised with "industry-leading pass rates" | No tool named |
| Undetectable AI | Rewrites that sound like you "while scoring as human on every major detector" | No tool named |
These are the vendors' marketing claims, and each vendor measures "passing" with its own checker or third-party detectors it chooses. None of them can show what a school's Turnitin report will say, because only instructors can see it. My Undetectable AI review, StealthWriter review and Phrasly AI review cover those products' plans and documented limits. HumanRank does not rank humanizers by whether they pass Turnitin; my ranking of AI humanizers uses price, free limits, languages and privacy terms instead.
Turnitin claims run through the rest of the category too. Grubby AI's marketing promises to refund your subscription if Turnitin flags its output, as my Grubby AI review explains. BypassGPT sells itself as an "AI detection remover" with a promise to beat Turnitin, covered in my BypassGPT review. Surfer's humanizer page names Turnitin too, while its own FAQ concedes that more advanced detectors "might" catch humanized text, per my Surfer AI Humanizer review. None of these vendors publishes a Turnitin report to back the claim.
The University of Melbourne gives students a blunt warning about this market. It says many free checker sites are "aimed at exploiting student anxieties", claiming to find AI in human work so they can sell humanizer tools, and that these tools "usually use AI and are often picked up by the University's detector tool."
Why "Turnitin-proof" claims are risky
A "Turnitin-proof" or "100% human score" claim is a vendor promise that no one outside the vendor can check against a real Turnitin report. Several facts from the sources above make it fragile.
- Students cannot test against Turnitin. Turnitin's AI score is visible only to instructors and administrators, and Turnitin sells no individual licenses. A "passed" result on a vendor's own checker is a different model with different thresholds.
- The target moves. Turnitin lists model updates in October 2025, February 2026 and October 2026. A result from one month says little about the next, and a resubmitted paper is rescored on the current model.
- Detection is only one check. Instructors compare work with earlier writing, ask students to explain their argument, and can review drafts and version history. A humanizer changes none of that; my guide on whether teachers can detect AI covers those checks.
- Edited text can read worse. Perkins and colleagues found some edits that lowered detection also made the writing poor enough to lose marks.
- Your text goes to a third party. Melbourne warns that putting work into free online tools can give away your intellectual property, and that some business models reuse the text. My guide on whether AI humanizers are safe covers vendors' license and refund terms.
- The integrity question does not change. If a course bans AI writing, using any tool to hide it is a breach of the rules, whether or not a detector flags it. My guide on whether using an AI humanizer is cheating quotes real policies, and my guide to do AI humanizers work reviews the research.
Legitimate uses for humanizer and rewriting tools exist. Here is where I think they fit and where they don't:
| Situation | My view |
|---|---|
| Marketing, SEO or product copy you drafted with AI | A fair use; edit for accuracy and your brand voice |
| Work emails and reports where AI help is allowed | A fair use; read the result before you send it |
| Smoothing phrasing when English is not your first language | A fair use where your course or employer allows it; disclose if asked |
| Coursework where the policy allows AI with disclosure | Follow the disclosure rule exactly; keep your drafts |
| Coursework where AI writing is banned | Not acceptable, with or without a detector |
| Answering a false flag on your own writing | The wrong tool; use your drafts and version history instead |
My guide to humanizing versus paraphrasing explains what these tools change in a text. For school work, the rule that matters is your course's AI policy.
If you're flagged
A Turnitin flag starts a conversation; it is not a finding. Turnitin itself says "the percentage on the AI writing indicator should not be used as the sole basis for action," and the University of Melbourne says "an AI writing detection report alone is not sufficient evidence for an allegation."
- Read your institution's policy first. Process and deadlines are set by your school, not Turnitin.
- Ask for the report and its date. Students cannot see the AI report, but an instructor can share a PDF. The date tells you which model version produced the score.
- Gather process evidence. Drafts, version history, notes, outlines and the sources you read show how the work was built.
- Be ready to explain the work. Melbourne says you may be asked how you developed the argument, which sources you used and how you reached your conclusion, and that this conversation "is not an allegation of academic misconduct."
- Be honest about any AI use. If you used a tool your course allows, say so and show how. If you used one it does not allow, a student union advisor can help you understand the process before you respond.
- Get support. Most universities have a student advocacy service that can attend meetings and explain appeal deadlines.
My take: if you wrote the work, your drafts will defend you far better than any tool. If you didn't, the honest conversation is the one that limits the damage.
My guide to AI detector false positives covers why human writing gets flagged and how to respond, and my Turnitin AI detection review sets out Melbourne's formal process step by step. More detection guides are in the HumanRank guides hub.
How we checked this
- Read on 29 October 2026: Turnitin's AI writing detection FAQ, Using the AI Writing Report, the AI writing model release notes and the 27 August 2025 bypasser press release.
- Vendor pages read: the home pages of StealthWriter, Phrasly and Undetectable AI, quoted as claims.
- Studies read: Perkins et al. (2024), Krishna et al. (NeurIPS 2023), Sadasivan et al. (TMLR) and Weber-Wulff et al. (2023).
- University guidance: the University of Melbourne's advice for students on Turnitin and AI writing detection.
- Method: I attributed each claim to the party making it and kept my own judgement in the "My take" lines. The site's general approach is on the HumanRank ranking methodology page.
Frequently asked questions
Can Turnitin detect AI humanizers? Turnitin says its English detector has flagged AI text modified by leading bypasser tools since 27 August 2025. It does not name the tools, and I know of no independent study of the current version yet.
Does Turnitin detect QuillBot paraphrasing? Turnitin's report guide names QuillBot as an example of an AI word spinner whose changes to AI-generated text it can detect, in English, since December 2023.
Can Turnitin detect StealthWriter or Phrasly? Turnitin does not say for any named tool. Both vendors make their own claims about passing detectors; those claims are marketing, not independent results.
Does humanized text show up differently in a Turnitin report? No separate label appears. Turnitin folds likely AI text that was paraphrased or bypassed into the AI-generated category and highlights it in blue.
Does Turnitin detect humanizers in other languages? No. Turnitin says paraphrase and bypasser detection is available for English submissions only.
Is there a free Turnitin AI humanizer? No. Turnitin does not make a humanizer. It sells detection to institutions, and its chief product officer has called humanizer companies cheating providers.
Can a humanizer make my own writing look less like AI? It may change a score, but a false flag is better answered with drafts and version history than with more rewriting. Your process evidence is what an integrity review weighs.
Is using an AI humanizer against school rules? It depends on your course. Schools set their own AI rules, and using any tool to hide AI use that a course bans is an integrity breach.
REVIEW NOTEI checked this guide on 29 October 2026 against Turnitin's own documentation and press releases, the claims on three humanizer vendors' home pages, four independent studies and University of Melbourne guidance. I explain what each source says and give my own view. I do not rank tools by whether they pass a detector, and I never will.