Fabricated citations in the age of AI: A wake-up call for editors, reviewers, and authors

As an academic editor, editorial board member, and reviewer for several medical journals, I handle dozens of submissions each year. Recently, while serving as an academic editor, I encountered a particularly serious issue. The manuscript in question contained major methodological flaws, and before making a rejection decision, I discovered that nearly 80 % of the cited references were fabricated or nonexistent. What was even more concerning was that the manuscript had already passed the journal's initial technical screening and undergone three rounds of peer review by three different reviewers, yet neither the editorial staff nor the reviewers noticed that most of the references were false. This was the first time I had encountered such a situation in my editorial career, and it was both shocking and thought-provoking. It led me to reflect on how editors and journals can better fulfill their gatekeeping responsibilities in an era when artificial intelligence is increasingly influencing academic publishing.

Large language models (LLMs), such as Generative Pre-trained Transformer (GPT), have become useful tools for improving the efficiency of manuscript drafting and revision, particularly for researchers who are non-native English speakers.^{1, 2, 3} However, their use also introduces new risks, such as fabricated or plagiarized content, false citations, and factual inaccuracies.^{1, 2, 3} Several recent studies have shown that LLMs frequently generate non-existent references.⁴ Across six independent investigations, approximately 51 % of the 732 citations analyzed were fabricated. Although newer versions such as GPT-4 have reduced this rate (from 55 % fabricated citations with GPT-3.5 to about 18 % with GPT-4), citation errors remain common. Even among genuine references, 43 % of GPT-3.5 and 24 % of GPT-4 citations contained substantive errors. While OpenAI recently released GPT-5, claiming further reductions in false citations, there is still insufficient independent evidence to confirm this improvement.⁵

Given these realities, responsible and transparent use of LLMs is essential. Authors may appropriately employ such tools for language editing, translation, grammar correction, or formatting assistance, but must carefully review all generated content and never cite AI-generated references. Any use of LLMs should be disclosed in accordance with the journal's policy. Editors should enhance technical screening procedures to detect fabrication, plagiarism, and false citations before manuscripts enter peer review. Reviewers, in turn, must evaluate the scientific quality, originality, rigor, and ethical compliance of manuscripts, while remaining vigilant for signs of AI-generated content, fabricated data, or false references. Only through these combined efforts can we uphold the integrity and credibility of academic publishing in this rapidly evolving era.

Declaration of competing interests

The authors have no conflicts of interest relevant to this article to declare.