An abstract image of a teacher standing between two giant icons: one representing an AI detector (robotic eye scanning text) and the other representing a human brain glowing with creativity. The teacher leans toward the brain, symbolizing trust in human judgment.

The Problem with TurnItIn and other AI Detectors in Education

Over the last few years, many educators have turned to AI detection software to check for student use of AI in their assignments. One such tool, Turnitin.com, is widely known for its plagiarism detection, but it now includes an AI detection feature as well. But, is the AI Detection feature effective? And is it fair?

In a recent episode of the Educational Duct Tape Podcast, I discussed this with my friend and fellow educator, Molly Klodor, who raised important concerns about its effectiveness and fairness.


Turnitin’s AI Detector: Useful or Problematic?

For years, educators have used Turnitin to detect plagiarism by identifying direct copying from internet sources. But its AI detector is less straightforward. Molly pointed out that, while Turnitin claims its tool is highly effective, the study they cite in those claims involves a small sample size of undergraduate students, not younger or more diverse students.

In fact, Adam Sparks highlighted on X that, while Turnitin’s AI detector performed the best of the 14 detectors in the study, it may be far less reliable than it claims: “In this study, Turnitin was right about 79% of the time. Put another way, it was wrong about 1/5th of the time!“

On the bright side, in this study at least, Turnitin did not have any results that the researchers believe would’ve led to “False Accusations.” In other words, while it missed a AI-generated text more often than you’d hope, it never identified human-written text as being AI-generated.

Avoiding those false accusations is essential but, on the other hand, Sparks noted that Turnitin failed to recognize AI writing almost 30% of the time. This means that in nearly a third of cases, students using AI tools could pass undetected. Interestingly, each of those undetected cases from the study occurred when AI-generated text was modified before being submitted. And get this: it doesn’t have to be modified by a human. Have ChatGPT write something for you and then ask it to modify it and Turnitin and other AI Detectors will be less likely to notice AI’s involvement.

A classroom scene with a teacher looking at a computer screen showing Turnitin’s AI detection results, split-screen style: one side glowing green for plagiarism detection, the other side with a confusing, glitchy AI detection bar, symbolizing uncertainty.
Image generated using Canva AI Image Generator (Pro) from the prompt “A classroom scene with a teacher looking at a computer screen showing Turnitin’s AI detection results, split-screen style: one side glowing green for plagiarism detection, the other side with a confusing, glitchy AI detection bar, symbolizing uncertainty.” August 17, 2025.

Most strikingly, Sparks pointed out that the very study Turnitin cites includes a warning: “The systems we tested should not be used in academic settings.” Despite this, Turnitin and others continue to promote their AI detectors for classroom use.

Molly and I discussed how this kind of error rate could have serious consequences in a classroom setting.

The first consequence is having AI-generated work slip through without the teacher noticing. And if your goal is to assess your learners’ mastery, as it should be, that’s a missed opportunity for an accurate assessment. And with Turnitin leading the pack of tools (in the study they cited) with a 79% success rate (plus a warning from the researchers themselves), the degradation of the quality of your assessments of mastery is evident.

The second and, in my opinion, bigger consequence is false accusations of student cheating or, as studies often label them, false positives. In this scenario, a student writes their own essay and the AI detector mistakenly identifies it as AI-generated. At that point, it’s not that big of a deal unless the unthinkable happens: the teacher trusts a website more than they trust their own learner. *shudders*

Again, I should point out, this is a scenario where Turnitin performed well (in this study). While GPT Zero had a likelihood of false accusations (which it defines as “a ratio of false positives and partially false positives to all cases”) of 50% and 5 other tools had more a 0% false accusation rate (Check for AI, Crossplag, GPT-2 Output Detector Demo, Writeful GPT Detector, and Writer), Turnitin and 7 other tools (Compilatio, Content at Scale, DetectGPT, Go Winston, OpenAI Text Classifier, PlagiarismCheck, ZeroGPT) had 0% false accusation rates.

And, listen, that’s pretty good news. But I’m not going to look at a study where 14 different tools looked at 18 undergrad-written essays and had an average false accusation rate of 6.75% (😦) and say “OK, go ahead and use an AI detector on my 10-year-old’s writing assignment.” Seriously, while Turnitin’s false accusation rate was 0%, the average for the 14 tools was to falsely accuse 1 of of every 18 students of having cheated with AI. (6.75% of 18 = 1.21)

And I’m even less likely to condone the use of them for learners other than my own kids, and here’s why:


Impact on Non-Native English Speakers and Neurodivergent Students

While Turnitin reports that “care was taken during dataset construction to represent statistically under-represented groups like second language learners [and] English users from non-English speaking countries,” studies have found that AI detection tools can misidentify work from students whose writing may not match typical patterns—particularly non-native English speakers (source: AI Bias Against Non-Native English Speakers) and neurodivergent students (source: Stanford’s Study on Bias Against Neurodivergent Writers). This leads to false accusations, which unfairly targets these specific groups of students, raising concerns about equity in the classroom.

A diverse classroom with students of different ethnicities and one student with visible neurodivergent cues (e.g., noise-canceling headphones, fidget tools). Their essays float above them, some falsely flagged as AI with warning icons, others passing unnoticed.
Image generated using Canva AI Image Generator (Pro) from the prompt “A diverse classroom with students of different ethnicities and one student with visible neurodivergent cues (e.g., noise-canceling headphones, fidget tools). Their essays float above them, some falsely flagged as AI with warning icons, others passing unnoticed.” August 17, 2025.

I do want to point out, again, that Turnitin performed perfectly on the 9 essays in this study that were “human-written in a non-English language with a subsequent AI/machine translation to English,” but again there were only 9 of them and they were undergraduate students. And, despite the perfect score of Turnitin (Compilatio and Content at Scale and also scored perfectly) the average for the 14 tools was 69% accuracy for those essays that were written by humans in a non-English language and then machine-translated to English. And here’s the kicker: the average false accusation rate of the 14 tools for the translated essays was 11.1%, with GPT Zero having a false accusation rate of 66.67%. (ouch) So, while GPT Zero would suggest accusing 6 of these 9 non-English speakers of academic dishonesty, the average of the 14 tools suggests accusing 1 of the 9. And, while 1 is better than 6, it ain’t 0.


What If It Was a Different Situation?

Here in Ohio, I have to renew my teaching license every five years. And, as a part of that renewal, I have to have my fingerprints taken for a BCI (Bureau of Criminal Identification) and FBI (Federal Bureau of Investigation) background check.

What if a group of convicted criminals did the background check, but used fake fingerprints and the test missed 41% of the criminals?

A futuristic fingerprint scanner at a security checkpoint. On one screen, a person with a clean record is wrongly flagged as a criminal, while on another, an actual criminal slips through with a green check mark. The metaphor highlights detection flaws.
Image generated using Canva AI Image Generator (Pro) from the prompt “A futuristic fingerprint scanner at a security checkpoint. On one screen, a person with a clean record is wrongly flagged as a criminal, while on another, an actual criminal slips through with a green check mark. The metaphor highlights detection flaws.” August 17, 2025.

That’s the average number of “undetected cases” in the study that Turnitin (who had 27.8%, the second best) cited.

What if 6.76% of the educators with clean backgrounds were falsely accused of having criminal histories?

That’s the average false accusation ratio in the study that Turnitin cited.


Why AI Detection Can’t Be the Only Solution

Imagine how educators would feel if they were falsely accused of having criminal records.

This hints at the emotional impact on students. Imagine having a teacher that you trust, respect, and revere accuse you of cheating . . . when you put your full, honest work into an assignment.

An abstract image of a teacher standing between two giant icons: one representing an AI detector (robotic eye scanning text) and the other representing a human brain glowing with creativity. The teacher leans toward the brain, symbolizing trust in human judgment.
Image generated using Canva AI Image Generator (Pro) from the prompt “An abstract image of a teacher standing between two giant icons: one representing an AI detector (robotic eye scanning text) and the other representing a human brain glowing with creativity. The teacher leans toward the brain, symbolizing trust in human judgment.” August 17, 2025.

Turnitin even shows the student the AI-generation score before they submit it. What would it feel like to see your hard work immediately flagged as AI-generated? This could be particularly traumatizing, especially for students already struggling with academic pressure.

Molly and I both agree: AI detection should not be the final word in determining academic integrity. These tools are guessing based on patterns, not identifying any hidden marker in the text. There’s no “tracker” embedded in AI-generated writing—just an algorithm making an educated guess. And like any guess, it can be wrong.

Fortunately, like other tools, Turnitin has admitted that inaccurate results are a possibility and the even published a guide for how to handle those situations to avoid falsely accusing learners of cheating.

While Molly suggested not using AI detection as the final word in this decision, if you’ve listened to the episode of the Educational Duct Tape Podcast, you’ll know that I came down more firmly: it likely shouldn’t be any word in the decision, worse yet the final one.


The Real Conversation Teachers Should Be Having

Rather than relying heavily on these tools, teachers should foster discussions with students about their writing process, their use of resources, and how they can ethically engage with AI tools. Teachers should conference with students throughout the writing process, formatively assess their writing and comprehension regularly, and see patterns in student writing.

A teacher and student sitting at a desk together, engaged in a warm, thoughtful discussion with papers and laptops open. Speech bubbles filled with handwritten notes and doodles float above them, symbolizing authentic dialogue and formative feedback.
Image generated using Canva AI Image Generator (Pro) from the prompt “A teacher and student sitting at a desk together, engaged in a warm, thoughtful discussion with papers and laptops open. Speech bubbles filled with handwritten notes and doodles float above them, symbolizing authentic dialogue and formative feedback.” August 17, 2025.

If you’re already using plagiarism detectors like Turnitin, go ahead and continue to use them for their intended purpose—identifying direct copying from the web. But when it comes to AI detection, be cautious. It won’t catch all of the AI-generated work and, more importantly, the stakes are too high for students to face false accusations based on flawed technology.


Final Thoughts

As AI continues to shape the educational landscape, we need to approach new technologies with a balance of caution and curiosity. AI detectors like Turnitin’s may play a role in managing academic integrity, but they can’t replace the nuanced formative assessments that can inform educators’ understanding of their students’ work. Ultimately, teaching students about ethical writing practices will always be more impactful than what any algorithm can determine.

A balanced scale floating in space: one side with an AI chip glowing blue, the other side with a stack of handwritten notebooks glowing yellow. The background is a futuristic classroom blending technology and traditional learning.
Image generated using Canva AI Image Generator (Pro) from the prompt “A balanced scale floating in space: one side with an AI chip glowing blue, the other side with a stack of handwritten notebooks glowing yellow. The background is a futuristic classroom blending technology and traditional learning.” August 17, 2025.

Have thoughts on AI detection in the classroom? Let’s discuss!

If you’d like to hear my full conversation with Molly tune into this episode of the Educational Duct Tape Podcast.

Published by

Jake Miller

Jake is the host of the Educational Duct Tape podcast, the #EduGIF Guy, a Tech Integration Coach, speaker, Former STEM, Math & Science Teacher, and a presenter.