1 of 37

AI Text Detectors

(aka AI Plagiarism Detectors & AI Content Detectors)

Designed by Torrey Trust, Ph.D.

​

College of Education

University of Massachusetts Amherst

@torreytrust | torrey@umass.edu

This work is licensed under CC BY NC 4.0, meaning that you can freely use, remix, and share it as long as you give attribution and do not use it for commercial purposes.

NOTE: This slide deck is a work in progress and will continue to be updated as new resources, research, and ideas are published.

Published April 2023 | Updated Nov 2023

2 of 37

Table of Contents

3 of 37

AI Text Detectors

The launch of ChatGPT motivated the development of new AI text detectors, such as GPTZero, ZeroGPT, Sapling, Copyleaks, and, more recently, Turnitin, to identify potential cases of plagiarism.

These tools were designed to determine whether text was written by humans or by AI (or both).

Screenshot of Copyleaks AI Content Detector page.

4 of 37

AI Text Detectors

Many of these tools boast a 97% or higher accuracy rate; but, when put to the test, they are not always as accurate as they claim.

Screenshot of the Sapling AI Detector results for human-written text.

5 of 37

AI Text Detectors

Screenshot of the Draft & Goal results for human-written text.

Screenshot of the Copyleaks results for human-written text.

6 of 37

AI Text Detectors

7 of 37

AI Text Detectors

8 of 37

How do AI Text Detectors Work?

Each detector seems to work differently.

Some detectors were trained on datasets of human-written text and AI-written text to predict the probability that any given text was written by AI.

Others work by looking for indicators within the text, such as linguistic patterns (Copyleaks, 2023).

Screenshot from Sapling AI Content Detector

Screenshot from CrossPlag

9 of 37

How do AI Text Detectors Work?

Screenshot from Copyleaks “ChatGPT and AI Content Detection” page

Screenshot from Draft & Goal

10 of 37

How do AI Text Detectors Work?

Many of the AI plagiarism detectors are not fully transparent about how their tool works. For instance, Turnitin’s tool runs text against their “AI detection model” but does not describe how that model was designed or works.

Screenshot from Turnitin FAQ page

11 of 37

Limitations of AI Text Detectors - Accuracy

Many of these tools explicitly state that their results may NOT be accurate and that they should NOT be used as a “primary decision-making tool” (OpenAI, 2023).

Screenshot from Copyleaks Terms of Use

12 of 37

Limitations of AI Text Detectors - Accuracy

Screenshot from OpenAi’s “New Classifier for Indicating AI-Written Text” page

Screenshot from GPTZero’s homepage

Screenshot from Sapling AI’s homepage

13 of 37

Limitations of AI Text Detectors - Accuracy

Worth a read: Can AI-Generated Text be Reliably Detected?

14 of 37

Limitations of AI Text Detectors - Accuracy

Worth a read: Testing of Detection Tools for AI-Generated Text

15 of 37

Limitations of AI Text Detectors - Accuracy

Worth a read: AI vs academia: Experimental study on AI text detectors’ accuracy in behavioral health academic writing (2024)

16 of 37

Limitations of AI Text Detectors - Evasion

Worth a read: GenAI Detection Tools, Adversarial Techniques and Implications for Inclusivity in Higher Education

17 of 37

Limitations of AI Text Detectors - Evasion

Worth a read: How effective are Turnitin, ZeroGPT, GPTZero, and Writer AI in detecting text generated by ChatGPT, Perplexity, and Gemini?

18 of 37

Limitations of AI Text Detectors - Evasion

Students can prompt AI writing tools to write text that will not be caught by AI Text Detectors.

Screenshot from ChatGPT can help you fool OpenAI’s anti-cheating tool

19 of 37

Limitations of AI Text Detectors - Evasion

AI tools are being developed to write/revise text so that it will not be caught by AI Text Detectors.

20 of 37

Limitations of AI Text Detectors - False Accusations

21 of 37

Limitations of AI Text Detectors - False Accusations

22 of 37

Limitations of AI Text Detectors - False Accusations

23 of 37

Limitations of AI Text Detectors - False Accusations

24 of 37

Limitations of AI Text Detectors - False Accusations

Students who are using Grammarly are getting FLAGGED by Turnitin and other AI text detectors!

25 of 37

Limitations of AI Text Detectors - False Accusations

Using GenAI tools to polish (improve) human-generated text leads to false positives!

26 of 37

Limitations of AI Text Detectors - Edited Text

Many of these tools are not trained to detect AI-written text that has been edited/remixed by humans (e.g., unless a student submits AI-written text as is without edits, the AI plagiarism detector results may not be accurate).

Screenshot from OpenAi’s “New Classifier for Indicating AI-Written Text” page

27 of 37

Limitations of AI Text Detectors - Language

Many of these tools are only trained on English language text and are likely unreliable for text written in other languages.

Screenshot from Turnitin’s FAQ page

28 of 37

Limitations of AI Text Detectors - Privacy

Most of these tools collect a lot of data from users and use, sell, or share that data to make money.

Hint: Make sure to read the privacy policy before using these tools!

Screenshot from Sapling AI’s Privacy Policy.

29 of 37

Limitations of AI Text Detectors - Free Labor

Some of these tools rely on feedback from users to improve their accuracy. This feedback (based on free human labor) can be used to improve the tool; and if the tool is sold or charges money for access to use it, the company is making money off this free labor.

Screenshot from Draft & Goal 2023

30 of 37

Evaluating Student Text with AI Text Detectors

The use of these tools to evaluate student text can increase students’ anxiety and stress (both of which have been found to inhibit learning), while also creating an atmosphere of distrust.

​

This is especially important to consider given the high false positive rate of these tools.

  • How would you feel if you wrote a blog post or journal article that was flagged as AI text and removed from consideration for publication, even though you knew you wrote it without the assistance of an AI writing tool?

Against the use of GPTZero and other LLM-detection tools on student writing (Gegg-Harrison, 2023)

31 of 37

Evaluating Student Text with AI Text Detectors

32 of 37

Evaluating Student Text with AI Text Detectors

33 of 37

Evaluating Student Text with AI Text Detectors

Do you really need to spend your time running students’ work through AI text detectors?

34 of 37

Evaluating Student Text with AI Text Detectors

Should you inform students if you plan to use AI Text Detectors to evaluate their text?

Consider that students immediately own the copyright to anything they record or write down (unlike a patent, students do not need to file a request with the government to copyright their work).

If you upload their text to one of these tools without their knowledge or permission, are you violating their intellectual property rights?

35 of 37

Using AI Text Detectors in Education

If you plan to use AI Text Detectors to evaluate student work:

  • Review and discuss your class Academic Integrity Policies with your students. Let them know when it is okay to use an AI writing tool (like ChatGPT, Quillbot, Grammarly), if at all, and when it is not okay.
    • For example, students could use ChatGPT to help with brainstorming ideas, rephrasing sentences, or improving their writing quality, but if they copy any text word-for-word directly from ChatGPT into their own work, that is considered plagiarism.
    • Check out the Classroom Policies for AI Generative Tools document. (Lance Eaton)
  • Be transparent - let students know how and why you will be using these tools.
  • Do not trust the accuracy of the results - instead, use the results as a means of starting a conversation with students about their writing (rather than automatically failing them based on the results).
  • Give students a chance to revise their work - Students, like the rest of society, are still grappling with the appropriate use of AI writing tools. If they are caught submitting AI-generated text, give them a chance to revise and resubmit their work rather than automatically failing them.
  • Give students a chance to explain their use of AI - You might ask students to respond to the following prompt for each written assignment submitted or if their work is flagged by an AI text detector:
    • “Did you use any AI tools to aid your thinking, writing, or learning on this paper? If so, how?”

36 of 37

Learn More

(this resource was created by SFCC professionals)

37 of 37

Still Concerned About Student Plagiarism?

Explore this one-pager: Technology is NOT the Solution to Cheating.