Should We Acknowledge ChatGPT as an Author?

The editorial examines whether ChatGPT, an AI language model, should be acknowledged as an author on scientific papers. In the context of rapid advances in large language models (e.g., GPT-3 and ChatGPT) and their growing influence on publishing, education, and science, the authors consider authorship criteria (ICMJE) and practical implications, aiming to clarify appropriate use and attribution of AI in scholarly work.

The piece references existing guidance and discourse: the ICMJE authorship criteria stating that tools like ChatGPT cannot approve manuscripts or take responsibility; a letter to the editor in the Journal of Epidemiology concluding that ChatGPT is not an author; and publisher policies (e.g., Science family banning AI use without permission; Elsevier and Cambridge allowing AI as a tool with disclosure). It situates the debate within ongoing policy developments and concerns about AI-generated inaccuracies.

The editors conducted a simple survey of the Journal of Epidemiology Editorial Board to assess views on ChatGPT’s role and author responsibilities. They also posed sample queries to ChatGPT to illustrate reliability issues: (1) asking for citations on coffee intake and liver cancer risk in Japan (Feb 20, 2023) and (2) asking the top causes of death in Japan in 2020 (answered with 2019 figures) (Mar 20, 2023). No detailed survey design, sample size, or instruments were reported.

• None of the Editorial Board respondents thought ChatGPT could be an author.
• 74% thought ChatGPT could be used as a tool; among these, 63% thought its use should be disclosed during submission.
• ChatGPT produced a fabricated/mismatched citation for coffee and liver cancer (title did not exist; DOI linked to an unrelated paper).
• When asked about top causes of death, ChatGPT incorrectly listed pneumonia as third for 2019; the actual top three were malignant neoplasms (27.3%), heart diseases (15.0%), and senility (8.8%).

Applying ICMJE criteria, ChatGPT cannot be an author because it cannot approve the final manuscript or take responsibility for content. The survey indicates editorial consensus that AI may serve as a tool, with transparency about its use. Given AI’s tendency to generate plausible but incorrect statements and unverifiable citations, authors must verify AI-generated content, and the scientific community must monitor its use. While some publishers restrict AI without permission and others allow tool use with disclosure, AI systems like ChatGPT or Perplexity AI still do not meet authorship requirements. Technological advances (e.g., GPT-4) may improve reliability but do not change accountability constraints.

The editorial concludes that ChatGPT should not be acknowledged as an author of scientific papers. It may be used as a tool with appropriate disclosure and with authors bearing responsibility for verification and integrity. Recognizing AI’s limitations while leveraging its strengths can help accelerate dissemination of scientific evidence as the technology advances (e.g., GPT-4).

The survey was simple and limited to Editorial Board members, with no reported details on sampling, instrument, or response rate. The AI reliability examples are anecdotal and time-specific rather than results of a systematic evaluation. No formal empirical study design is presented.