Effects of temperature settings on information quality of ChatGPT-3.5 responses: A prospective, single-blind, observational cohort study

Abstract

Abstract Objective The effect of temperature settings on the quality of ChatGPT version 3.5 (OpenAI) responses related to drug information remains unclear. We investigated ChatGPT-3.5’s response quality on apixaban information with and without the temperature being set to 0. Methods On 6 September 2023, 37 questions regarding apixaban, derived from the frequently asked questions on the Bristol–Myers Squibb’s website, were entered into ChatGPT in Japanese. The primary endpoint was the effect of temperature settings on ChatGPT-3.5’s responses to apixaban-related questions. The response accuracy, clarity, detail, and adequacy were rated on a 5-point Likert scale by 10 pharmacists, with higher scores indicating higher response quality. Cumulative score means were analyzed using the Mann–Whitney U test. In the subgroup analysis, evaluators were limited to pharmacists at university hospitals. Welch’s t-test was employed in sensitivity analysis to validate primary endpoint findings. Results The mean scores for ChatGPT-3.5’s apixaban-related responses with (13.08) and without (14.40) the temperature being set to 0 were not significantly different (p = 0.064). Accuracy differed significantly (3.15 vs. 3.54, p = 0.045), whereas clarity, detail, and appropriateness were similar. Subgroup analysis (13.30 vs. 14.21, p = 0.394) and sensitivity analysis confirmed similar results (13.45 vs. 14.52, p = 0.105). Conclusions ChatGPT-3.5 temperature setting does not significantly affect overall responses to apixaban-related inquiries. However, the variance in accuracy suggests that ChatGPT-3.5 is unable to consistently provide precise responses. Hence, it is more suitable as a supplementary tool rather than a primary medical resource.

Ähnliche Arbeiten