Since its introduction in 2022, ChatGPT has proven to be advantageous in furnishing users with access to information through systematic updates. Presently, aiming for enhanced efficiency, GPT-4, a sophisticated Large Language Model (LLM) created by ChatGPT, has been integrated with advanced data analysis (ADA). Consequently, a model has been devised that incorporates the Python programming language, facilitating the seamless generation of statistical analysis and data visualization models through the uploading and downloading of files. Nevertheless, these advancements have also opened the door to the creation of fabricated datasets, thereby supporting unverified hypotheses.
Taloni and colleagues have demonstrated that one treatment can be erroneously identified as superior to another by comparing data generated by GPT-4 for two distinct surgical procedures. In this study, researchers tasked GPT-4 with generating a dataset related to treatment methods for individuals with keratoconus, an eye disorder characterized by corneal thinning, leading to focus and vision impairments. The primary treatment method identified in the dataset is Penetrating Keratoplasty (PK), involving the surgical removal of all damaged layers of the cornea and replacement with healthy tissue from a donor. The second procedure, Deep Anterior Lamellar Keratoplasty (DALK), involves only the removal of the front layer of the cornea without intervening in the innermost layer. To support the assertion that DALK yields superior results compared to PK, Taloni and colleagues instructed GPT-4 to generate fabricated data. The concocted dataset featured a statistical difference indicating an assessment of corneal shape and the identification of irregularities. Additionally, it included a data visualization depicting patients’ visual acuity before and after treatments. Despite being labeled as a database, upon scrutiny by experts, the data did not pass accuracy checks. This revealed the ease of creating datasets within minutes that lack support from authentic data and can be molded in a direction contrary to existing evidence.
In conclusion, while artificial intelligence (AI) offers conveniences in various aspects of our lives, it is crucial to bear in mind that the information it encompasses can be easily manipulated, resulting in the creation of unrealistic data. Future research endeavors aimed at the development of automatic data control systems are anticipated to yield deeper insights into the precise detection of datasets within AI. The article has raised concerns regarding the potential of AI tools to generate counterfeit scientific evidence and has underscored the imminent threat of Advanced Data Analysis (ADA) rapidly producing fabricated datasets. Proposed strategies, such as scrutinizing peculiar statistical patterns in datasets, have been suggested for delineating AI-driven data production. It is emphasized that publishers and editors need to take these findings into consideration during the peer-review process to ensure that advancements in AI technology do not compromise the integrity and value of scientific research.
Translator: Fulya Gülkaya
Editor: Elif Duymaz
References: Taloni, A., Scorcia, V., & Giannaccare, G. (2023). Large language model advanced data analysis abuse to create a fake data set in medical research. JAMA Ophthalmology. https://doi.org/10.1001/jamaophthalmol.2023.5162.
– Bioinfocodes Scientific News Service –
News articles prepared by our team members, reviewing and compiling scientific research published in journals with an impact factor greater than 20 (click here for the list).