We can't run Turnitin ourselves: it sells only to schools and universities, and students never see the AI score. So this page sticks to what Turnitin's own documentation says, plus the independent studies that tested it. Every source is listed at the end with the date we read it.
What Turnitin says it detects
| Date | What changed (Turnitin's own release notes) |
|---|---|
| Dec 6, 2023 | Detects likely AI text even after it has been run through an "AI word spinner" (paraphraser). |
| Jul 16, 2024 | Reports gain a separate category for AI text that was then AI-paraphrased. Turnitin's own example of such a tool is QuillBot. |
| Aug 27, 2025 | Adds detection of "the likely use of AI bypasser tools", meaning humanizers. English submissions only. |
| Feb 12, 2026 | English model updated to catch more AI text while keeping false positives low. |
| Jul 2026 | Turnitin says it replaced several models with a single model. |
| Aug 4, 2026 | The separate paraphrase category is removed: paraphrased and bypassed AI text now shows as plain "AI-generated" in the report. Detection itself continues. |
The August 2026 change matters if you read older advice. Reports no longer show a separate colour for "AI-paraphrased" text. Turnitin's reasoning, in a post by its chief product officer, was that AI editing is now built into everyday writing tools, so splitting "generated" from "modified" invited over-reading. The same post repeats that paraphraser and bypasser detection is English-only.
Which humanizers does it catch?
Turnitin won't say. Its FAQ declines to name the tools it detects, on the grounds that naming them would help people evade it. So any page claiming "Turnitin catches X but not Y" is guessing, unless it shows a real Turnitin report, which only an instructor can produce.
On Grammarly, Turnitin's FAQ says ordinary spelling and grammar fixes mostly weren't flagged in its tests on human writing, but text produced by Grammarly's generative features (drafting, paraphrasing, summarising) will likely be flagged. We found no Turnitin statement about QuillBot's separate "AI Humanizer" product.
What the teacher sees, and what you don't
- Students don't see the AI score. Turnitin's FAQ says the AI indicator and report aren't visible to students; an instructor can download and share the PDF.
- Scores under 20% are hidden. Since July 2024, a result between 1% and 19% shows as an asterisk with no highlighting, to cut false positives.
- Minimum length. English AI reports need at least 300 words of prose. Lists, bullet points and code aren't analysed, so the percentage is of the qualifying prose, not the whole file.
- Short texts are close to all-or-nothing. Turnitin says that on documents of a few hundred words, the prediction tends to be all or nothing, so a mixed essay can come back as entirely AI.
- Turnitin Clarity is a separate, paid writing space some schools add. It shows instructors the writing process (pasting, typing, draft history), not just the final text. That makes pasting in rewritten AI text visible in a way a detector score alone isn't.
How accurate is it on humanized text?
Turnitin publishes a target, not a measured detection rate for humanized text. Its target is a false-positive rate under 1% at document level, for documents with more than 20% AI writing, and it says it accepts missing some AI text to keep that low: by its own account, a document scored at 50% "could contain as much as 65%" AI writing.
Independent evidence is thin and dates quickly:
- A peer-reviewed 2026 study (Vrije Universiteit Brussel, International Journal for Educational Integrity, June 2026) ran 160 English papers through Turnitin, GPTZero, Copyleaks and Pangram. Turnitin scored every human paper correctly, but put every fully AI paper in the 0-20% band, and was right on half of the "humanised" papers. Two caveats: the humanising was done with a GPT-4o prompt, not a commercial humanizer, and the texts were generated in May 2025, before Turnitin's bypasser update.
- A Notre Dame preprint (arXiv, August 2026) found that after rewriting with Undetectable AI, fewer than 4% of AI texts stayed flagged. It tested Pangram and GPTZero, not Turnitin, so it is often mis-cited as a Turnitin result.
We found no independent study that tests today's Turnitin against commercial humanizers. Anyone selling a humanizer "guaranteed to pass Turnitin" has no public evidence for that claim.
What this means for you
- If your school uses Turnitin with AI detection on, assume humanized English text can be flagged, and that you won't see the score before your instructor does.
- Turnitin itself says its score should not be the only basis for a misconduct decision. If you're flagged for work you wrote, ask what else the decision rests on and offer your drafts and version history.
- If your school bans AI rewriting tools, no humanizer changes that.
Sources (all read Oct 8, 2026)
- Turnitin Guides, "AI writing detection model" release notes (entries dated 2023-12-06, 2024-07-16, 2025-08-27, 2025-10-14, 2026-02-12, 2026-05-05): guides.turnitin.com/hc/en-us/articles/28294949544717
- Turnitin Guides, "Turnitin release notes" (entry dated 2026-08-04): guides.turnitin.com/hc/en-us/articles/27251688507533
- Turnitin Guides, "Turnitin's AI writing detection capabilities FAQs" (languages, file rules, false-positive target, student visibility, Grammarly, July 2026 model change): guides.turnitin.com/hc/en-us/articles/28477544839821
- Turnitin blog, "How Turnitin is simplifying AI detection for educators and publishers", Aug 4, 2026: turnitin.com/blog
- Turnitin Educator Network, product team post launching AI bypasser detection (Aug 27, 2025): turnitin.forumbee.com
- Turnitin press release, Turnitin Clarity launch, Mar 4, 2025: turnitin.com/press
- Van Vlasselaer, Van Droogenbroeck & Spruyt, "Who wrote this? Evaluating the reliability of AI detection tools in higher education", International Journal for Educational Integrity 22:16, Jun 29, 2026 (correction Oct 6, 2026): doi 10.1007/s40979-026-00226-w
- Karr, Khvatskii, Hua & Chawla, "Why AI Detection Fails for Academic Integrity", arXiv 2608.11256, Aug 2026 (preprint)