logo
Volume 24, Issue 9 (12-2025)                   JRUMS 2025, 24(9): 773-774 | Back to browse issues page


XML Persian Abstract Print


Download citation:
BibTeX | RIS | EndNote | Medlars | ProCite | Reference Manager | RefWorks
Send citation to:

Rezaeian M. Limitations of Using Generative Artificial Intelligence in Writing Scientific Articles: Use of Fake Citations and Tendency Towards Western Texts. , JRUMS. 2025; 24 (9) :773-774
URL: http://journal.rums.ac.ir/article-1-7906-en.html
Abstract:   (1087 Views)
The editorial in the September issue of this year was dedicated to the journal’s approach regarding the use of generative artificial intelligence (AI) tools in writing manuscripts. We noted at that time that while these tools offer positive attributes, they also possess numerous limitations. Indeed, it now seems that the limitations of using AI in writing medical articles are so significant that one must employ these tools with the utmost caution.
In the mentioned editorial, we addressed one of these limitations: if one asks a generative AI tool to write sections of a paper, such as the introduction, it may generate sources that do not exist (1). These “fake” citations are considered a byproduct of AI “hallucinations” (2).
Hua and colleagues requested both older and newer versions of a generative AI tool to produce scientific abstracts, along with ten references, for clinical research questions across seven ophthalmology subspecialties. They demonstrated that 33% of the references generated by the initial version and 29% of those generated by the updated version were fake sources created by the AI (3). Since the use of fake sources causes irreparable damage to the foundations of science, researchers—especially early-career researchers—must use these tools with a full awareness of these limitations (1).
Furthermore, Cohen and Moher pointed to another limitation of generative AI tools in scientific writing. They noted that generative AI tools are designed to produce text by utilizing Large Language Models (LLMs). For this purpose, and to operate optimally, LLMs are trained on vast datasets consisting of diverse textual content and billions of parameters (4).
Although the precise training data sources for most LLMs are either unreported or unknown, there is a strong probability that the majority are disproportionately in English and produced by institutions in high-income Western countries. Historically, institutions in Eastern and low-income countries have either not produced such data or have produced it with lower quality. Consequently, the texts currently generated by AI tools largely reflect the conditions and requirements of high-income Western countries, a factor that can further exacerbate existing disparities in this field (4).
 
Full-Text [PDF 129 kb]   (489 Downloads) |   |   Full-Text (HTML)  (358 Views)  
Type of Study: Research | Subject: آموزش پزشكي
Received: 2025/12/7 | Accepted: 2025/12/19 | Published: 2025/12/19

References
1. Rezaeian M. Updating Publication Ethics Checklist. J Rafsanjan Univ Med Sci 2025; 24 (6): 495-6.
2. Chelli M, Descamps J, Lavoue V, Trojani C, Azar M, Deckert M, et al. Hallucination rates and reference accuracy of ChatGPT and Bard for systematic reviews: comparative analysis. J Med Internet Res 2024; 26: e53164.
3. Hua HU, Kaakour AH, Rachitskaya A, Srivastava S, Sharma S, Mammo DA. Evaluation and comparison of ophthalmic scientific abstracts and references by current artificial intelligence chatbots. JAMA Ophthalmol 2023; 141(9): 819e24.
4. Cohen JF, Moher D. Generative artificial intelligence and academic writing: friend or foe? J Clin Epidemiol 2025; 179: 111646.

Add your comments about this article : Your username or Email:
CAPTCHA

Send email to the article author


Rights and permissions
Creative Commons License This work is licensed under a Creative Commons Attribution-NonCommercial 4.0 International License.