Simultaneous ChatGPT-4o outputs for type 2 diabetes pharmacotherapy: Accuracy, usefulness, and impact variability.

PURPOSE: The primary objective of the study was to determine how clinically accurate and useful are ChatGPT-4o-generated outputs when identical prompts are entered simultaneously in three independent ChatGPT sessions utilizing the same user-facing model. The secondary objective was to identify if differences in outputs affect clinical outcomes (i.e., impact variability). METHODS: Five clinical prompts were developed focusing on diabetes management and counseling. Each clinical prompt was simultaneously inputted into three separate devices using ChatGPT-4o to generate outputs. A modified Delphi technique was then utilized involving five diabetes management clinical pharmacists. Each clinical pharmacy faculty independently rated each ChatGPT-generated output based on accuracy (i.e., poor, borderline, good) usefulness (i.e., not useful, somewhat useful, very useful) and impact variability (i.e, low, moderate, high). After initial assessment, responses were collated and anonymously shared among the pharmacy faculty. The faculty members were invited to revise their evaluations based on collective feedback. Pharmacy faculty then convened in a virtual panel with moderators to discuss evaluations and work towards consensus. RESULTS: Consensus was achieved for all ChatGPT outputs. Accuracy ratings ranged from borderline to good. Two of the clinical prompts yielded outputs receiving different accuracy ratings which may have impacted variability. Usefulness ratings ranged from somewhat useful to very useful, with one clinical prompt yielding outputs that received different usefulness ratings. CONCLUSION: While outputs were generally accurate and useful, limitations and inconsistencies were noted. Users should be aware simultaneously generated outputs across multiple devices using ChatGPT can vary in their accuracy and usefulness in diabetes management.