AI humanizers and Turnitin: what the evidence shows

Check your institution's AI policy before using any writing tool. A humanizer does not make undisclosed AI use acceptable, and this site does not help anyone misrepresent authorship. MLA asks you to cite a generative AI tool for any content it created and to acknowledge editing uses in a note (MLA Style Center, citing generative AI, read 9 November 2026); APA asks you to describe the use in your Method section or introduction and to give a reference entry (APA Style blog, how to cite ChatGPT, read 9 November 2026).

No column below is a Turnitin result. The fields are the ones every page on this site uses, and the order is alphabetical, because no run has been published and no score exists to sort by.

What does Turnitin state it detects?

Turnitin's AI writing page says the checker helps educators identify when "an AI writing tool such as ChatGPT, text spinners, or AI bypassers (also called humanizers)" may have been used, read 9 November 2026. Its AI detector page adds that the indicator "shows an overall percentage of likely AI-generated text", including content "further modified using an AI paraphrasing or bypassing tool to evade detection", and that each sentence is scored and the scores averaged for the document (Turnitin, AI writing detection, read 9 November 2026).

Three stated numbers sit on those pages. Turnitin states a false positive rate below 1 percent for documents containing more than 20 percent AI-generated content, read 9 November 2026. Turnitin states that it tested nearly 2,000 writing samples from English language learners and found a false positive rate of 0.014 against 0.013 for native English writers on documents meeting its 300-word minimum, read 9 November 2026. Turnitin states that it "does not make a determination of misconduct" and provides data for educators to judge under their own policies, read 9 November 2026.

Two things are absent from Turnitin's pages. There is no published detection rate for humanized text, only the statement that detection includes it. And the under 1 percent false positive band applies to documents with more than 20 percent AI writing; below that line Turnitin makes no accuracy claim, which is why a small highlighted share should carry little weight on its own. Turnitin's own blog says the rate "is not zero" and asks instructors to assume positive intent, read 9 November 2026.

Not every institution kept the indicator on. Vanderbilt University disabled Turnitin's AI detector in August 2023, citing false positive concerns, in a public notice on its Brightspace site (Vanderbilt University, guidance on AI detection, 16 August 2023, read 9 November 2026).

Which is the best AI humanizer for Turnitin by your situation?

• Choose none of them for undisclosed AI use; the integrity notice above is the whole answer.
• Choose a tool with a mild named mode (Walter Writes AI's Simple, QuillBot's Basic, Phrasly's Balanced) if your institution permits AI-assisted editing and you want the smallest rewrite.
• Choose Grammarly AI Humanizer if your institution permits it and you already draft in Grammarly.
• Choose Clever AI Humanizer or HIX Bypass if your text runs past 1,000 words and must be processed whole.
• Choose a detector you can see, not a humanizer, if your worry is a false positive on your own writing: keep drafts with version history, and read the essays page for the disclosure rules.

What did the paraphrase studies find?

Four peer-reviewed or preprint studies tested what paraphrasing does to detectors that researchers could access. None tested Turnitin's current model, and that limit applies to every row below.

Krishna and colleagues reported that DIPPER dropped DetectGPT's detection from 70.3 percent to 4.6 percent at a constant 1 percent false positive rate (Krishna et al. 2023, arXiv 2303.13408, read 9 November 2026). Sadasivan and colleagues found that recursive paraphrasing significantly reduced detection rates across detector types while only slightly degrading text quality in many cases (Sadasivan et al. 2023, arXiv 2303.11156, read 9 November 2026). Weber-Wulff and colleagues included Turnitin in their 14 tools, tested in 2023 against the model of that time, and concluded that the available tools were neither accurate nor reliable (Weber-Wulff et al. 2023, International Journal for Educational Integrity, read 9 November 2026).

Two cautions follow. The studies measured detector readings, and "slightly degrades text quality" was the researchers' metric, not an editor reading an argument for dropped claims. And detector makers have moved since 2023: Originality.ai states that its 3.0.2 Turbo model of September 2025 reached up to 97 percent accuracy on the latest AI humanizers and bypassers at a 1.5 percent false positive rate, and that the model was retired in September 2026 for a newer one (Originality.ai, AI detector accuracy page, read 9 November 2026). A finding about a 2023 detector says nothing certain about a 2026 one.