The OpenAI tool “Chat Generative Pre-trained Tansformer (ChatGPT),” a popular artificial intelligence (AI) large language model (LLM), has recently gained attention for its potential applications and implications in various fields, including academic writing [1, 2]. Editors of Nuclear Medicine and Molecular Imaging (NMMI) are concerned about various issues that may arise from the use of ChatGPT in the research and publication process. To explore the potential benefits and problems of using ChatGPT in academic writing, a roundtable discussion was held on Saturday, April 22, 2023, from 9:15 to 9:45 a.m. at Conference Room 10 on the 5th floor of the Busan Port International Exhibition and Convention Center (BPEX) during the 2023 annual congress of the Korean Society of Nuclear Medicine (KSNM). I organized and moderated the meeting as the editor-in-chief of NMMI.

The participants were invited from among the members of KSNM who have a deep interest in using LLM, including ChatGPT (Fig. 1). The list of participants, in alphabetical order by name, includes Dr. Ari Chong (Department of Nuclear Medicine, Chosun University Hospital, Gwangju, S. Korea), Dr. Byeong-Cheol Ahn (Department of Nuclear Medicine, Kyungpook National University Hospital, Daegu, S. Korea), Dr. Dong-Woo Kim (Department of Nuclear Medicine, Severance Hospital of Yonsei University, Seoul, S. Korea), Dr. Hee-Seung Henry Bom (Department of Nuclear Medicine, Chonnam National University Hwasun Hospital, Hwasun, S. Korea), Dr. Hongyoon Choi (Department of Nuclear Medicine, Seoul National University Hospital, S. Korea), Dr. Jungsoo Oh (Department of Nuclear Medicine, Asan Medical Center, Seoul, S. Korea), Dr. Ki-Seong Park (Department of Nuclear Medicine, Chonnam National University Hospital, Gwangju, S. Korea), Dr. Mijin Yun (Department of Nuclear Medicine, Severance Hospital of Yonsei University, Seoul, S. Korea), and Dr. Suk Hyun Lee (Department of Nuclear Medicine, Hallym University Kangnam Sacred Heart Hospital, Seoul, S. Korea).

A roundtable discussion was held on Saturday, April 22, 2023, from 9:15 to 9:45 a.m. at Conference Room 10 on the 5th floor of the Busan Port International Exhibition and Convention Center during the 2023 annual congress of the Korean Society of Nuclear Medicine

There has been a lot of discussion about the possible roles of ChatGPT in research and paper writing. Participants introduced their experiences of using ChatGPT for scientific writing.

Ki-Seong Park: I recently used ChatGPT for my draft, and I employed it throughout the entire process, from generating research ideas to writing my personal statement. In the process, I experienced both the potential scope and limitations of ChatGPT. Here are some tips and limitations I encountered while conducting research with ChatGPT.

To generate research ideas using ChatGPT, I first asked it to explain basic concepts related to my research topic. For example, I planned to use FP-CIT PET to differentiate Parkinson’s disease from other conditions, and asked ChatGPT background questions about Parkinson’s disease and other diseases such as progressive supranuclear palsy and multiple system atrophy to activate its knowledge of the topic. When ChatGPT responded, I then asked about research topics in the field. However, research topics can span many different aspects, including political, social, economic, cultural, medical, and psychological, so it is helpful to specify each aspect when asking ChatGPT about potential topics.

ChatGPT could help me find literature. We should know that the free version, GPT3.5, has low reference accuracy, while the paid version, GPT4, has improved accuracy. The training data is only available through September 2021, so the most relevant or up-to-date literature might be missed. Therefore, we need to use other search engines to search for references.

When I entered findings in a table, I could move the table to Excel and insert it into ChatGPT. I described my findings in text format as ChatGPT does not yet support input in graph or picture formats. ChatGPT can create an abstract based on the results I entered, but the body of the paper may be limited by the token limit of ChatGPT. Instead, I can divide the content into sections such as Introduction, Materials and Methods, Results, Discussion, and Conclusion. I started by asking ChatGPT for an outline based on my paper title, keywords, or abstract, then edited and added the detailed content so that it makes up the body of my paper. The text generated by ChatGPT might not match the format commonly used in academic papers, so it required a lot of editing. In many cases, satisfactory answers were not immediately available, requiring repeated edits and revisions.

Dong-woo Kim: I mainly use ChatGPT to summarize research findings, write abstracts, and proofread my writing. I also had a lot of trouble finding references with ChatGPT. It provides a lot of references based on suggested keywords, but only about 1 out of 10 times can I actually find a relevant paper, and the DOIs are often inaccurate. As a simple example, it provided incorrect information about the year of establishment of KSNM. The overall outline is good, but the details are often inaccurate. It helps to correct English grammar and sentence structure, but it has technical limitations, so authors should carefully check the accuracy of the content before using it in their papers.

Jungsu Oh: One problem with ChatGPT is that it needs accurate input data to provide accurate results. For example, when I needed to find a PET/CT machine, I purposely asked for CT information, and ChatGPT directed me to a site about PET/CT as if it was about CT. It also repeatedly gave me the same wrong answer without admitting its mistake. Just as AlphaGo makes a mistake once with incorrect data and keeps making mistakes, similar errors occur when dealing with poorly trained data, which is why you need to input accurate data and factual information to get the right answer. Because of these errors, I do not think it is appropriate to use it for extended reproduction just yet.

Hongyoon Choi: Lately, I have been using ChatGPT more than Google search, and other LLMs are also popular. These tools, like ChatGPT, are widely used, similar to the Internet or Google, making it nearly impossible to restrict their usage. Actively discussing whether to use them might not be productive. Their applications can be diverse, from writing necessary sentences to drafting entire papers. ChatGPT generates high-quality sentences, helping overcome language barriers if used and edited appropriately. Concerns about reference errors were mentioned earlier, but existing systems already check for unsearchable items on Medline during the publication process. Moreover, it is unlikely that authors would use ChatGPT for an entire paper without any modifications. I primarily use LLMs for abstracts and paper reviews, and I believe more people will use them similarly.

Hee-Seung Henry Bom: It is beneficial to use ChatGPT when writing a thesis or report, but I wonder if we should emphasize concerns about the extended reproduction of false data.

Byeong-Cheol Ahn: I would like to focus on the issue of attribution when using ChatGPT. While some recent papers have recognized ChatGPT as an author, there are many opinions that do not recognize LLM like ChatGPT as authors. A recent editorial in Science stated that while ChatGPT contributes to the creation of a paper, it cannot be recognized as an author because it is not responsible for errors [1,2,3]. The journal PLOS One also stated that LLMs cannot be authors, and that if you use ChatGPT in your paper, you must clearly describe its use in the methods section [4]. Authors are responsible for ensuring there are no concerns about potential plagiarism. Use of AI-generated text without proper citation could be considered plagiarism [1]. I am sure other journals, not just PLOS One, have similar policies. To establish NMMI’s position on liability, it would be helpful to create an AI use checklist to clarify journals’ policies and require authors to disclose their use of LLMs in advance.

There have been cases of reviewers misusing LLMs, leading to author complaints about improper reviewing. To maintain a journal’s reputation, it is necessary to establish a checklist and mechanisms to clarify authors’ responsibilities. When authors submit papers written with LLMs without proper validation, liability issues may arise, making a checklist crucial for clear author confirmation.

Ari Chong: ChatGPT helps prevent plagiarism through paraphrasing, and we encourage its use in this regard. However, authors should review and edit content responsibly, and a checklist can help clarify this. Reviewers should also be aware of the issues that can arise when using LLMs in their reviews and use them responsibly.

Suk Hyun Lee: LLM usage is becoming a trend, like Grammarly or Google Translate. I agree on preparing a checklist to clarify accountability, using examples from other journals like PLOS One as references.

Mijin Yun: ChatGPT is a very useful tool that can sometimes surpass human capabilities. It is good at writing specialized articles where it is hard to distinguish fact from fiction without specialized knowledge. It is interesting to see that LLMs also provide an opportunity for people to learn new information. This should not be a problem for papers, as long as the methodology discloses the use of LLMs, but I am concerned that authors might use ChatGPT and submit their papers without review.

Byeong-Cheol Ahn: We seem to be in a transitional phase where validation is important. As long as authors validate their work, the use of LLMs should not be an issue. Following the example of PLOS One, it is appropriate to mention the use of LLMs in the methodology or, if not, in the acknowledgments.

Hee-Seung Henry Bom: Thank you for active participation and invaluable suggestions. The editorial board will have to decide on a policy, but I think what we discussed today will be very helpful, and we will be doing a lot of work in the future, including organizing training sessions on how to write papers with LLMs including ChatGPT.