GPT wars: Creating and detecting fake ophthalmology data sets with artificial intelligence

Abstract

Aims/Purpose: Recently, our group demonstrated how the Large Language Model GPT‐4 can fabricate fake ophthalmology data sets, designed to support false scientific evidence. The objectives of this study were: (1) to identify peculiar statistical patterns of AI‐generated data sets, (2) to attempt at enhancing the quality of fabricated data for eliminating any discernible mark of non‐authenticity, and (3) to develop a custom GPT trained to detect signs of fabrication. Methods: Three fictional studies were designed to compare treatment outcomes for a specific ocular disease. First, prompts were submitted to the custom GPT “Data Analyst” to produce 12 “unrefined” fake data sets for the three studies. Then, the custom GPT “Fake Data Creator” was developed to generate 12 “refined” fake data sets, specifically designed to evade authenticity checks. Forensic analysis was performed on all data sets using IBM SPSS. Finally, the custom GPT “Fake Data Detector” (FDD) was developed to replicate the forensic analysis and assign 1 point for each sign of fabrication. Results: Forensic analysis of the unrefined fake data sets revealed numerous flaws, most notably: name/gender mismatch ( n = 12), non‐uniform distribution ( n = 11) and repetitive patterns of last digits ( n = 7), absence of correlation between study variables ( n = 12), and distribution shape anomalies ( n = 11). In refined fake data sets, the forensic analysis did not reveal statistical flaws, except for distribution shape anomalies ( n = 7). Overall, 5 refined fake data sets (41.7%) passed forensic analysis as authentic. The results of the statistical analyses performed by GPT FDD were identical to those obtained with IBM SPSS, with a mean of 9.67 ± 2.13 (95% IC 8.46‐10.87) and 1.00 ± 1.00 (95% IC 0.43‐1.57) points respectively allocated to unrefined and refined data sets. Conclusions: Sufficiently sophisticated custom GPTs can perform complex statistical tasks and may be abused to fabricate seemingly authentic ophthalmology data sets, passing forensic analysis. References Taloni A, Scorcia V, Giannaccare G. Large Language Model Advanced Data Analysis Abuse to Create a Fake Data Set in Medical Research. JAMA Ophthalmol . 2023; 141(12): 1174–1175. doi: 10.1001/jamaophthalmol.2023.5162

Ähnliche Arbeiten