Generative AI: Potential and Pitfalls in Academic Publishing

Abstract

G enerative artificial intelligence (AI) has attracted enormous attention since the release of Chat-GPT (1) in late 2022. Other generative AI chatbots used to generate text (including Bard, Claude, and CoPilot (2-4)) and images (including Dall-E, Midjourney, and Stable Diffusion (5-7)) have likewise seen a remarkable explosion in development and uptake. While ChatGPT has been promoted for its potential to assist users in efficiently and easily creating text to serve a wide range of purposes, this editorial focuses on the journal's opinion on the use of generative AI technology in creating articles submitted to this journal.ChatGPT utilizes a technology known as a generative pretrained transformer large language model (8). This is a form of machine learning in which very large language-based data sets are used to train computers to comprehend natural language. While natural language processing is not new, ChatGPT is currently unique in its ability to not only understand queries and information to which it has access, but also to generate new, comprehensive, and fluent language-based content. Models, such as ChatGPT, which generates language, and Dall-E, Midjourney, or Stable Diffusion, which all generate images, are collectively referred to as "generative AI", and can be seen as marking a dramatic shift in both the capabilities and widespread access to AI technology.Since its release on November 30, 2022, ChatGPT achieved the most rapid adoption of a consumer software application in history by amassing over 100 million users by January 2023 (9). Users have leveraged ChatGPT to write software, song lyrics, stories, poems, and letters. With its ability to recall previous prompts within a conversation, users can fine-tune its responses to modify the content or tone of the content it generates. The underlying model used to generate content is continuously improved through feedback from users. While users easily identified flaws in the responses provided by the earlier GPT-3 version, the model has iterated quickly. GPT-4, OpenAI's latest effort in scaling up deep learning, has been described by its creators to exhibit "human-level performance on various professional and academic benchmarks" (10). It has demonstrated impressive abilities in passing standardized examinations including the Law School Admission Test, Scholastic Aptitude Test, a unified bar exam, and even the United States Medical Licensing Exam (10).ChatGPT, and other large language model-based applications, use large data sets consisting of text available on the web. Sources may include articles, books, web-based advertising, and social media posts. ChatGPT offers several exciting and desirable potential benefits, including its potential to make completing written articles more quickly as well as completing literature review summaries. With its ability to help users write in English fluently, it has been touted as a way to help make academic publishing more equitable and diverse (11). The time saved in summarizing data and generating articles could help r...