# Does Turnitin check for AI, and what does the instructor actually see?

Source: https://forum.global100.org/q/does-turnitin-check-for-ai-and-what-does-the-instructor-actually-see/
Site: Global 100 Forum, category Turnitin, Canvas and schools
Published: 2026-09-18. Updated: 2026-09-24. Replies: 1.

## Question by Maya Lindqvist (staff), 2026-09-18

Yes, if the institution licenses and enables it. The instructor sees a percentage of the prose sentences the model thinks were AI-generated or AI-paraphrased, plus highlighted passages, inside the Similarity Report. Students do not see it. The caveat: Turnitin hides any score under 20% behind an asterisk because its own testing found those scores unreliable.

## What is switched on, and for whom

Turnitin's AI writing detection is not in every Turnitin account. Its [AI writing detection FAQ](https://guides.turnitin.com/hc/en-us/articles/28477544839821-Turnitin-s-AI-writing-detection-capabilities-FAQs) says the feature is only available to customers that license Turnitin Originality, or the AI writing add-on for iThenticate 2.0. Administrators can enable or disable it, and Turnitin does not process submissions for AI writing where it is off. Through Moodle, Blackboard or Canvas the indicator appears inside the Similarity Report, not the LMS interface. It is not available through Gradescope.

## What the report shows

Turnitin's [guide to the AI Writing Report](https://guides.turnitin.com/hc/en-us/articles/22774058814093-Using-the-AI-Writing-Report) describes the pieces. The headline is an overall percentage of "qualifying text": prose sentences in long-form writing. Bullet points, tables, code, poetry and annotated bibliographies are not counted, which is why the percentage often does not match the amount of highlighted text. Below it, a breakdown bar shows where flagged passages fall, and those sentences are highlighted in the document. For English submissions the model also flags AI text it believes was later run through a paraphraser or bypass tool, and counts that inside the same percentage.

The report needs a file of at least 300 words of prose, under 30,000 words, in English, Spanish, Japanese or Arabic. Anything else returns an error state rather than a score.

## The 20% rule and the asterisk

In May 2023 Turnitin's chief product officer [published an update](https://www.turnitin.com/blog/ai-writing-detection-update-from-turnitins-chief-product-officer) after running 800,000 pre-ChatGPT papers through the detector. It found a higher incidence of false positives when the model detected less than 20% AI writing, so it added an asterisk to those scores, raised the minimum word count from 150 to 300, and changed how the first and last sentences were aggregated. The same post put the sentence-level false positive rate at roughly 4%, with most of those sentences sitting next to genuine AI text.

The current guide goes further: for reports generated after 8 July 2024, scores between 1% and 19% are not shown at all. The indicator displays `*%` and no highlights. A score of 0% means no qualifying text was flagged.

## What Turnitin says about being wrong

The FAQ states a document-level false positive rate under 1% for documents with more than 20% AI writing, checked against 700,000 pre-ChatGPT papers before each model release, and spells out the trade: to keep false positives low the model will miss some AI text, so a document showing 50% could contain 65%. It also lists the writing most likely to trip a false positive: repetitive text, text with little structural variation, and paraphrase that adds no new ideas. Grammarly's spelling and grammar fixes are generally not flagged; its generative rewriting features likely are. It repeats that the percentage should not be the sole basis for action. What a student should do with a flag is covered in [its own thread](/q/what-should-a-student-do-if-an-ai-detector-flags-their-work/).

**What the Turnitin AI report shows and what it does not**

| Shows | Does not show |
| --- | --- |
| A percentage of qualifying prose sentences flagged as likely AI-generated or AI-paraphrased | Which AI tool, model or paraphraser was used |
| Highlighted sentences and a breakdown bar of where they fall | Any score between 1% and 19% (replaced by `*%` since July 2024) |
| Separate indicator states: 0%, `*%`, 20 to 100%, error, not enabled at submission | Bullet points, tables, code, poetry or bibliographies in the count |
| A downloadable PDF the instructor may choose to share | Anything to the student by default |
| Aggregated AI scores at institution level, for administrators | A finding of misconduct; Turnitin says that decision is the instructor's |

## What that means in practice

- **Ask whether it was even on.** If the institution has not licensed Turnitin Originality, there is no AI indicator, whatever a student has been told.
- **A percentage is not a passage count.** Highlights and percentage diverge whenever the paper mixes prose with lists or tables.
- **The report is a starting point, by design.** The [accuracy thread](/q/how-accurate-are-ai-text-detectors-really/) covers the independent tests; Turnitin's own documents already tell instructors to treat the score as a prompt for a conversation.

## Reply 1 by Maya Lindqvist (staff), 2026-09-24

One detail instructors often miss: the FAQ says short documents are scored almost "all or nothing" because there is only a single segment and no overlap. A 400-word reflection that mixes a student's own sentences with a pasted AI paragraph can come back as entirely AI. If a flagged submission is near the 300-word floor, ask for the draft history or an earlier version before treating the percentage as a proportion, because at that length it is closer to a binary guess than a measurement.

---
Cite as: Global 100 Forum, "Does Turnitin check for AI, and what does the instructor actually see?", https://forum.global100.org/q/does-turnitin-check-for-ai-and-what-does-the-instructor-actually-see/, accessed 2026-10-11.