AI Detection with Turnitin

Written in consultation with the exam committees of Wageningen University.

Contents:

Introduction

Now that generative AI (GenAI) tools have become so popular, GenAI detection tools are increasingly used in educational contexts. These tools are said to be able to determine whether any given text was written by a human or by a GenAI tool, and are thus used by many examiners to detect AI fraud. However, how accurate they are (and will be in the future) is a question up for debate. To determine in what way we should use these tools in education – if at all – we conducted research on the (current) accuracy of Turnitin’s AI detection score. 

There are many more (paid and unpaid) AI detection tools out there, and although there will of course be differences in accuracy, we assume the accuracy of Turnitin and other tools to be comparable. In addition, these other tools often lack transparency on what is done with the materials uploaded, which is why we discourage the use of AI detection tools other than Turnitin.

In the following paragraphs, we will first provide a step-by-step overview of how to use Turnitin’s AI detection tool. In the following attachement, we explain how we arrived at these recommendations by summarising our approach, findings, and conclusions. 

Note! AI detection is a temporarily solution. Please reach out to [email protected] to discuss the best way to integrate AI in your course and assignments.  

Rules and steps to follow

Based on the findings of the research below, the examining board proposes the following steps in AI-detection:

  1. Read the submitted work critically yourself. Develop your own opinion.
  2. Check whether the student has acknowledged AI use, and to what degree. Is this in line with the assignment?
  3. Check the references for validity and existence of publications.
    (See if the Plagiarism checker highlights references. If so: Likely existing reference (top figure). If not: Likely hallucinated reference (bottom figure) This is not a guarantee, however.)
  1. Ask for version histories of the submitted work.
    1. Version history requires the submission of previous file versions. Upon receiving these files, check in the file properties when the file was created & modified.
    2. Previous versions of files can be obtained via the OneDrive “Version history” option (when a file was maintained in OneDrive) by the student.
  2. (Optional) Use Turnitin to see if your suspicions match flagged paragraphs by Turnitin.
  3. If suspicions remain: Forward the case to your Examining Board. These can recommend proceeding with step 7.
  4. If suspicions are present: Perform an additional oral verification without access to AI.

You can also request advice from the Examining Board at an earlier stage.

Turnitin reliability report

Summary of the report

In this study, the accuracy of Turnitin’s AI detection was evaluated for different AI use cases, ranging from fully human written to completely AI-generated. For fully human written samples, circa one in 75 (1.5%) of all samples surpassed the 20% detection limit of Turnitin, with an average Turnitin detection score of 36% among positives. After improving the spelling and grammar of these human written samples with Grammarly, 3% of all samples were flagged as likely AI-written. The percentage of samples flagged by Turnitin when rewritten with different Large Language Models is 57%, with a detection percentage of 74% among positives. When fully AI-written, an average of 30 percent of samples were not detected, raising serious concerns about the reliability of Turnitin’s AI detection. Moreover, significant differences in AI detection percentages and flagging percentages were observed between models for both rewriting and generation.

Full report

You can access the full report here:

attachment Turnitin Report 23-12-2025.pdf