Introduction

Recently, large-scale language models (LLMs) such as Generative Pre-trained Transformer (GPT) have been attracting much attention due to their ability to achieve human-like performance in tasks such as translation, text editing, and article generation. However, there are mixed opinions on the use of this technology in scientific writing. This is an Editorial comment on the development of LLM and its application to the writing of scientific papers, its problems, and whether it should be mentioned as the author.

Development of LLM

The current rapid development of LLMs began with a neural network architecture called Transformer, published in 2017, which can capture relationships between words in a sentence by learning based on a scoring system called “Attention” [1]. GPT is a language model that extends Transformer and is pre-trained on large text data sets and is capable of generating more human-like text as the scale increases from GPT-1, GPT-2, to GPT-3. ChatGPT is a chat application combining GPT-3 with reinforcement learning that was released by OpenAI in November 2022 and has now become such a hot topic that it has been featured in various mainstream news programs.

LLM's ability to write scientific papers and the problems it faces

Recently, ChatGPT was used to generate abstracts for research articles in leading medical journals and examined whether researchers could distinguish them, and found that 32% of the abstracts generated by ChatGPT were real abstracts, whereas 14% of real abstracts were mistakenly identified as ChatGPT-generated abstracts [2]. At this point, ChatGPT's paper writing ability seems to have reached a level close to that of humans. However, on the other hand, LLMs such as ChatGPT learn from a large amount of data on the Internet including information that has not undergone scrutiny for accuracy, so there is a high possibility that it learns scientifically incorrect information. Therefore, there is no guarantee that the correct content will be output by such LLMs. The following five issues have been identified as priorities for future research using such LLMs; maintaining human verification, creating rules for accountability, investing in truly open LLMs, reaping the benefits of LLMs for the scientific paper writing, and broadening the discussion [3]. Specifically, there is an urgent need to establish rules for clarification of the use of LLM, evaluation of the text quality written by LLM, and establishment of copyright and ethical rules; however, these remain largely unfulfilled at this time.

Description of LLM use

In spite of the above problems, some papers have appeared in which ChatGPT is listed as an author, and it is said that the establishment of rules for the use of LLMs is urgently needed [4]. Currently, many argue that papers authored by LLMs do not meet authorship requirements because they cannot be held accountable as authors and LLMs cannot agree to terms of use or content distribution rights for software tools [5], and we also support this.

At this time, no firm conclusion has been reached as to the extent to which the use of LLMs is acceptable. The Science plans to change its editorial policies to prohibit the submission of English text produced by ChatGPT [6]. On the other hand, the Nature does not completely prohibit the use of LLMs, but strongly recommends that researchers using LLM tools should document this use in the methods or acknowledgements sections [7]. We follow the Nature's policy and recommend that if authors use the text as it is output by a large language model such as ChatGPT, it should be carefully noted in the acknowledgements session. For examples, “ChatGPT was used for the English translation, which was validated in detail by the authors. Or “ChatGPT was used to survey previous reports and to summarize them, the validity of which has been verified by the authors through careful research of each original article.”

Conclusion

This paper has described the Japanese Journal of Radiology editors' current opinions on writing papers using LLMs. Please note that this field is in an evolving state of development and that guidelines and legislation have not yet caught up with the area, so these opinions may change significantly in the future.