Why Turnitin AI Detection Mistakenly Flags Human Work as AI

Discover how a PhD students original work was wrongly flagged as AI by Turnitins new update that made their detections much more strict.

As AI-powered tools become more prevalent in detecting academic dishonesty, instances of false positives have been reported. One such case involves a Ph.D. student at Stanford who was detected for 100% AI on Turnitin, despite the text being human-written.

Stanford University
Stanford University

Your professor accused you of using ChatGPT for cheating. You did nothing. Now what?

Don’t worry. We’ll give you a detailed explanation of how Turnitin operates and the potential negative consequences it may have in an academic setting.

In just over a year since ChatGPT’s launch, teachers are now supplied with AI-powered tools that can detect if students are using AI to cheat. By August, Turnitin, a company specializing in plagiarism detection, had already run its AI detector on over 70 million assignments. Turnitin is confirmed to have a 4% false positive rate, but it’s assumed that this number is much higher after Turnitin’s recent update. This means that over 280 thousand assignments have been incorrectly flagged as AI-generated, which causes a great amount of concern for students.

Turnitin’s Recent Update

Earlier this week, on December 5th, 2023, Turnitin updated its AI detection algorithms, resulting in an increase in the number of false positives. Phrasly, a content generation & humanization tool, has received several reports from students almost immediately after the update from students claiming their own written work has been flagged as 100% AI. Our team has quickly stepped in to help students avoid these false positives by using the Phrasly.AI humanization feature.

TurnItIn false-positive from human-generated text written by a student

Turnitin claims to promote academic integrity on its website. However, the recent updates to their AI detection system have gone too far, causing legitimate student work to be flagged as AI-generated. This poses a serious problem in educational institutions as it could lead to students being unfairly penalized with academic probation or failing grades for assignments they did not cheat on. The false positive rate is assumed to be over 10% after this update, which is very high, especially when the majority of educational institutions across the globe trust Turnitin to give accurate results.

Turnitin’s False Positives

Our team at Phrasly.AI has obtained an instructor account with Turnitin. This account allows us to view what teachers see after a student submits their assignment. We are using this for research purposes and to demonstrate that Turnitin’s new update is excessively strict.

After receiving reports from a Ph.D. student at Stanford University that their professor accused them of generating AI-generated work despite the work being human-written, we found the Turnitin results surprising. We’ve double-checked their text with TurnItIn, and it seems to be 100% AI-generated, even though the work is human-written.

Ph.D. Student at Stanford University’s assignment being flagged incorrectly by Turnitin.

Turnitin is confident that this text is AI-generated, even though it isn’t, as shown in the image above. After the new update, Turnitin’s system has increasingly flagged human-generated content as AI-generated. This phenomenon is becoming more common, raising the concern of student’s work being flagged as AI incorrectly.

As mentioned earlier, this is not the first reported case of students’ work being incorrectly flagged by Turnitin. In fact, there are dozens of students who have reportedly been disciplined by their schools for no reason.

Take a look at this example: a college senior at UC Davis named William Quarterman was accused of cheating on his history exam after his professor used an AI detection tool called GPTZero, which alleged that his answers were generated by AI.

Quarterman denied the accusations and suffered panic attacks while appearing before the university’s honor court. He was eventually exonerated when he presented proof that he did not use AI.

While GPTZero is known to have its set of false positives, it is not as widely used as Turnitin, which is built into its platform and checks papers for AI by default. These false positives can have life-changing outcomes, especially when the student’s future is on the line. Compared to plagiarism detection, where teachers are informed where the content originates from, there is no way to actually prove if a student is using AI to write their papers.

How to Prevent False Positives

Turnitin’s recent update may require students to be more cautious about the sources and text they use to avoid AI detection. Here are some ways to avoid your content being incorrectly flagged as AI:

1. Avoid Using Grammarly

Grammarly, an AI-based typing assistant founded in 2009, leverages AI for grammar correction. While Grammarly may be effective in making text more coherent and double-checking your work, it can be flagged as AI when submitting your work to Turnitin.

We understand that Grammarly can be an extremely useful tool, and we even use it ourselves. However, if you want to avoid detection, we suggest you avoid using Grammarly’s dedicated AI generation features. If your text was indeed written by a human and you use Grammarly, the probability of your content being flagged as AI is low, but never zero.

2. QuillBot Doesn’t Bypass AI Detection

QuillBot, a company acquired by CourseHero in 2021, is a very popular tool to paraphrase your content for free. Students often use QuillBot to avoid plagiarism detection, but it is not foolproof against AI detection. It is a common misconception that QuillBot bypasses AI detection when it does not at all.

These assumptions have originally come from quite a few TikTok posts, which consist of people claiming that QuillBot bypasses AI detection. Through TikTok, millions of users have been misled by these claims, resulting in students’ work being flagged as 100% AI on Turnitin.

TikTok misleading student’s that QuillBot bypasses AI detection, when it does not
3. Don’t Copy and Paste Wikipedia Content

This may be an odd claim, but some AI detectors actually use Wikipedia as part of their training data to differentiate between human and AI-detected content. Don’t believe us? Test it yourself.

We are using a Wikipedia post about the United States for this example. When we put the text into an AI detector such as GPTZero, the text is flagged as 89% AI. Keep in mind that these texts were written way before ChatGPT was released.

Wikipedia text being flagged as AI

While Wikipedia may be a good source of information, we suggest that you don’t directly copy and paste it. Paraphrasing the text using QuillBot may bypass plagiarism, but it will not be fool-proof against AI detectors.

4. Use the Phrasly.AI Platform

While AI detectors are trained on AI-sounding text, Phrasly.AI leverages custom AI models that are trained on 300k+ human-written essays. This way, Phrasly writes like a human and bypasses AI detection.

In the AI-detected text above, we’ve inputted it into Phrasly’s humanization feature. As you can see in the image below, Phrasly’s paraphrased content is marked as 100% human, according to GPTZero.

AI-Written text inputted into Phrasly.AI’s humanization feature

If you want to reliably bypass plagiarism & AI detection, we suggest using the Phrasly.AI humanization feature. This way, you never have to worry about your text getting incorrectly flagged as AI-written.

Conclusion

The recent update to Turnitin’s AI detection system poses a significant risk to students, as there is now a higher chance of their text being falsely flagged as AI-generated. However, there are ways to avoid this. By following the tips mentioned above, you can take steps to bypass AI detection and ensure that your work is accurately assessed.