Is Turnitin's AI detection accurate?

What Turnitin says

Turnitin’s AI writing indicator reports the share of a submission its model judges likely to be AI-written. Turnitin states a document-level false positive rate under 1 percent for documents with more than 20 percent AI writing, and it publishes a page titled “Understanding false positives within our AI writing detection capabilities.” That is a vendor figure, measured on the vendor’s test sets, for the case it describes.

Turnitin also reports how much AI-heavy writing it sees. Its data from February 24, 2026 put 15 percent of submissions above 80 percent AI for October 2025 to February 2026, against 3 percent at launch in 2023, and 19.4 percent of US higher education submissions were heavy AI in Q2 2026.

Turnitin is also direct about the limits. Its guide to the AI Writing Report says the model “may not always be accurate (it may misidentify both human and AI-generated text) so it should not be used as the sole basis for adverse actions against a student.” Scores between 1 and 19 percent are shown only as an asterisk, with no highlights, which Turnitin says is to avoid false positives at the low end.

Source: Turnitin blog and press materials, checked 2026-09-08; guides.turnitin.com, “Using the AI Writing Report” and “AI writing detection model”, checked 2026-10-22.

What independent evidence says

Outside Turnitin’s own testing, the picture is less tidy.

• Some writers are flagged more often. Liang and colleagues at Stanford (Patterns, 2023) found that GPT detectors flagged 61.3 percent of TOEFL essays by non-native English writers as AI-generated. That study tested detectors generally, not Turnitin alone, but it shows how a model trained on typical prose treats simpler, more formulaic English.
• Students report false accusations unevenly. Common Sense Media (2024) found 20 percent of Black teens said they had been falsely accused of AI use, against 7 percent of white teens.
• Universities have switched it off. Vanderbilt (August 2023), the University of Waterloo (September 2025, after an internal test flagged human text as 100 percent AI), Curtin University (January 1, 2026) and others have turned the indicator off. Australian Catholic University faced about 6,000 allegations, roughly 90 percent AI-related, a large share of them dismissed (ABC News, October 2025).
• Courts have weighed in. In Newby v. Adelphi University (2026 NY Slip Op 26021), a 100 percent Turnitin flag was read as human by Grammarly and ZeroGPT; the court found the finding “without valid basis” and ordered the record expunged.

Sources: Liang et al., Patterns, 2023; Common Sense Media, 2024; Vanderbilt, Waterloo and Curtin announcements; ABC News Australia; Newby v. Adelphi, Justia; all checked 2026-09-08.

Why a small error rate still matters

Even taking the vendor figure at face value, a small rate across a large number of papers produces real students wrongly flagged. A department that reads thousands of essays a year will see some, and each one is a person facing an accusation with no way to cross-examine a model. The false positive rates page walks through the arithmetic and every published number.

How to read a Turnitin AI score

1. Treat it as a reason to look, not a result. Write down the number and what else, if anything, raised the question.
2. Look at how the paper was written. A process record or a document history tells you something the score cannot.
3. Check every citation against the source.
4. Compare with the student’s earlier work.
5. Talk with the student about the paper before anything formal. The step by step guide has a first message.

What not to do

• Do not send the student a screenshot of the score.
• Do not run the paper through other detectors until one agrees.
• Do not penalize on the score alone.

What answers the question a score raises

A score says something about the finished text. What settles a case is evidence about how the text was produced. When students write in EssayGuard, every keystroke, revision, pause and paste is recorded with server-side timestamps, and periodic snapshots cross-check the record so a fabricated history does not read as writing. Each essay is marked Approved, Needs Review or Flagged with a plain summary, and you can replay the session. EssayGuard gives no AI percentage and does not judge whether prose sounds machine-made; a person makes every decision.

A record has limits too. It shows how text entered the editor, not whose ideas it carries. It gives the conversation with the student something concrete to look at. EssayGuard is free for the rest of this semester; set up one assignment and compare what the record shows with what a score would have told you.