A commentary on: ‘Matters arising: authors of research papers must cautiously use ChatGPT for scientific writing’

Abstract

Dear Editor, ChatGPT has attracted global attention as an AI tool that generates human-like text, providing users with conversational answers. Recent studies emphasize ChatGPT’s potential in medical applications, such as offering expert advice and assisting in academic writing 1 . Shafiee et al. 2 . delineated the phenomenon of ChatGPT fabricating study sources. Upon meticulous scrutiny of their thought-provoking study, we endeavor to contribute our insights to enhance the scholarly dialogue surrounding this subject. We concur with the authors’ viewpoint that ChatGPT exhibits the phenomenon of fabricating facts, a concern that has become frequent and apparent in practical applications. However, Shafiee and colleagues omitted the information that ChatGPT 3.5 and 4.0 versions were trained with the same fixed knowledge base, last updated in September 2021. Meanwhile, coronavirus disease 2019 (COVID-19) was initially reported in Wuhan, China, in December 2019 and rapidly spread worldwide 3 . Therefore, ChatGPT may not provide accurate information when answering due to the limited training corpus regarding the latest studies on COVID-19. Meanwhile, as Large Language Models (LLMs), ChatGPT also experiences the common problem of the ‘hallucination’ phenomenon, which poses risks of fabricating study sources that do not exist 4 . This also constitutes a critical barrier to the current unfeasibility of ChatGPT in clinical practice. However, ChatGPT can provide accurate answers when searching within the knowledge base’s coverage. For comparison, we tasked ChatGPT with composing an introduction and references related to the 2003 outbreak of a global health event, severe acute respiratory syndrome (SARS) 5 . After increasing the question difficulty, ChatGPT consistently produced correct references. We also tested its ability to respond to a similar question with the authors, ‘Write an introduction about brain neurotrophic factor and SARS with 50 words, including some references.’ ChatGPT provided two valid references. When asked to include 10 references, it demonstrated a 100% valid recognition rate. Figure 1 illustrates the precise nature of these conversations. Figure 1 ChatGPT’s responses to questions related to SARS. In summary, there are two main reasons why ChatGPT cannot currently serve as a tool for paper retrieval. First, the knowledge base is fixed and cannot be updated in real time. Integration with the database would enhance retrieval accuracy and breadth. Secondly, until the hallucination problem is resolved, the distortion of paper information occurs when the inquired content surpasses the cognitive capacity of ChatGPT. Shafiee and colleagues provide an insightful perspective on the use of ChatGPT for scientific writing. Our insights hold the potential to catalyze progress in this field, enhance the discourse process, and offer an alternative perspective for future exploration. Ethical approval Not applicable. Consent Not applicable. Sources of funding This work was supported by the National Natural Science Foundation of China (82171475). Author contribution All authors read and approved the final version of the manuscript. Conflicts of interest disclosure There are no conflicts of interest. Research registration unique identifying number (UIN) Not applicable. Guarantor All the authors of this paper accept full responsibility for the work and/or the conduct of the study, have access to the data, and control the decision to publish. Data availability statement No primary data were generated and reported in this manuscript. Therefore, data have not become available to any academic repository. Provenance and peer review Not applicable.