ChatGPT-3.5 en plagiaat (Engelse opsom)

‘n Voorskou van ‘n artikel in druk oor ChatGPT-3.5 en die moontlikheid om plagiaat te bevorder. Engelse opsomming.

Titel: A Pro-active strategy to mitigate plagiarism with ChatGPT-3.5 – an explorative study

Almost immediately after OpenAI (2022) have launched ChatGPT on the 30th of November 2022 online, and free of charge, innovators (Rogers 1983), or those who have adopted this innovative technology first, flooded the internet with reports regarding the quality of its texts. The minority described ChatGPT as a stupid machine, based on single prompts, but the majority found ChatGPT’s answers of such a high quality that it could have passed various post-graduate examinations, if human. These reports spurred the adoption of ChatGPT, and it reached the 100 000-user milestone within a record-breaking period of two months (Reuters 2023).

Soon lecturers worldwide cautioned that ChatGPT could promote plagiarism on a large scale. This possibility must be taken seriously as research conducted after the launch of the internet and word-possessing tools showed that these innovative technologies made it easier for students to plagiarise. They only needed to find information on the internet, copy and paste it into word-processing tools, save it, and submit it as their own work. In an attempt to mitigate plagiarism, developers launched plagiarism detectors such as Turnitin to help educators to detect instances of plagiarism. These plagiarism detectors mitigated instances of plagiarism, but promoted a new type of academic dishonesty, namely, to pay others to complete assignments on their behalf.

ChatGPT makes it even easier to plagiarise as it only need a prompt (question) to search its large corpus of data, collected from the internet, within seconds to generate text-based outputs. Thus, students only need to copy these answers with the press of a button, paste it in their word processing tools, save it and submit it as their own work. Although the abuse of ChatGPT is not (yet) regarded as plagiarism, it is a new form of plagiarism as the student did not conduct the research, write the text, or refer to ChatGPT as the original author. Different from previous plagiarism detection practices, it is a daunting task to identify instances of ChatGPT-plagiarism.

ChatGPT does not copy and paste from the internet or its training data. It writes word by word using algorithms to statistically predict the next word in a sentence while taking previous words into account. Thus, ChatGPT’s texts are original and so authentic that Turnitin does not find similarities with existing texts. Therefore, developers hastily launched artificial intelligence detectors (AI-detectors) which educators can use to determine if a text was AI- or human-written. However, these AI-detectors are not reliable enough to accuse a student of ChatGPT-plagiarism. The detection of ChatGPT-plagiarism is also hindered by another group of developers who have launched new tools which students can use to alter ChatGPT’s texts in such a way that it reduces the risk of being detected.

Thus, ChatGPT needs the attention of academics (King en ChatGPT 2023). Although ChatGPT do not intentionally plagiarise, it can commit reference plagiarism as it makes references up (Senekal 2023; Van Staden 2023). But it can promote plagiarism if students submit its answers as their own work. It is reasonable to assume that students will abuse ChatGPT based on the quality of its texts (Bommarito II en Katz 2022; Kelly 2023; Kung et al. 2023; Terwiesch 2023) and critical thinking skills. According to Susjnak (2023) ChatGPT’s critical thinking skills can seduce students to abuse ChatGPT, especially as it needs minimum inputs to generate realistic answers. Jiao e.a. (2023) found that its translation skills good, and Aydin and Karaarslan (2022), Mellon, Bailey, Scott, Breckwoldt and Miori (2022); Alshater (2023) and Dowling and Lucey (2023) found it useful during various phases of the research process. Marti published within the first month a book entirely written by ChatGPT (Marti and ChatGPT 2023). However, Senekal (2023) and Van Staden (2023) does not regard ChatGPT as a reliable research assistant. As ChatGPT can be abused, and policing strategies are inefficient, lecturers need to understand how ChatGPT-plagiarism can be mitigated.

The purpose of this research was to develop a pro-active strategy lecturers can apply to mitigate ChatGPT-plagiarism in their classrooms. As AI-chatbots differ, and I have assumed that the majority of students will be using the free ChatGPT-3.5 model, I have specifically focused on developing strategies for this model. Due to the novelty of ChatGPT, and insufficient knowledge regarding its ability to promote plagiarism, I have conducted explorative research. I have followed a novel approach, namely, to identify its limitations to better understand how plagiarism can be mitigated in classrooms. As I have found that three factors, namely usefulness, input expectation (amount of effort needed to use the technology), and work-related expectations (how it can be used to improve work) impact on the adoption decision, I have assumed that these factors can also impact on the decision to adopt ChatGPT for unethical purposes.

I have followed a mixed methods approach to collecting data. The qualitative data was collected while I have conducted structured interviews with ChatGPT to explore its limitations. As ChatGPT can remember within a chat but cannot carry the new learning over to future chats, I have started a new chat to explore each of the limitations. Both quantitative and qualitative data were collected while I have used Turnitin, three AI-detectors, namely AI Text Classifier, GTP-2 Output Detector en GPT-Zero (AI-detectors), and ChatGPT to detect if the answers were AI-written.

The first finding was that most of the limitations identified in literature could limit the usefulness of ChatGPT. Its knowledge was limited to events before November 2021; thus, it could not answer questions related to events after that date. It did not know much about subjects not often discussed on the internet, and hallucinated (made up or lied) facts convincingly. It did not perform well when prompted to discuss Afrikaans literature (human sciences), and hallucinated links to YouTube videos. It could also not create a PowerPoint presentation but provided guidelines for the structure and content of such a presentation. Although it could argue different points of view, it was sympathetic towards a convicted murder (Oscar Pistorius), but not towards a South African fighting against farm murders (Steve Hofmeyr). This might indicate a problem with its training as it can indicate that ChatGPT leans to the left. These limitations can have an impact on the usefulness of ChatGPT in classrooms.

The second finding was that Turnitin, the three AI-detectors and ChatGPT could not identify all of the AI-texts. In fact, Turnitin and the three AI-detectors regarded ChatGPT’s texts as human-written. Even ChatGPT could only identify three of the seven AI-texts as its own work and was convinced that it was not the author of the rest.

Based on the findings, I have suggested a few pro-active strategies for mitigating the abuse of the free version of ChatGPT-3.5 in classrooms. Following this route, rich learning opportunities can be provided as students can learn from own experience that ChatGPT’s limitations, mistakes, and hallucinations can have a negative impact on their results if submitted without improving the texts. The pro-active strategies can be implemented from a learning-oriented approach to assessment, which is based on three principles, namely (a) to design learning tasks, rather than assessment tasks, (b) to provide feedback promptly, and (c) to involve students as peer assessors. Following this route, higher education can send employable graduates into the workplace as they have learned how to use artificial intelligence, such as ChatGPT, responsibly.

Keywords: ChatGPT; ChatGPT-plagiarism; higher education; plagiarism combatting; strategies