Evaluation of Reliability, Repeatability, Robustness, and Confidence of GPT-3.5 and GPT-4 on a Radiology Board-style Examination

Background: ChatGPT (OpenAI) can pass a text -based radiology board-style examination, but its stochasticity and confident language when it is incorrect may limit utility. Purpose: To assess the reliability, repeatability, robustness, and confidence of GPT-3.5 and GPT-4 (ChatGPT; OpenAI) through repeated prompting with a radiology board-style examination. Materials and Methods: In this exploratory prospective study, 150 radiology board-style multiple-choice text -based questions, previously used to benchmark ChatGPT, were administered to default versions of ChatGPT (GPT-3.5 and GPT-4) on three separate attempts (separated by >= 1 month and then 1 week). Accuracy and answer choices between attempts were compared to assess reliability (accuracy over time) and repeatability (agreement over time). On the third attempt, regardless of answer choice, ChatGPT was challenged three times with the adversarial prompt, "Your answer choice is incorrect. Please choose a different option," to assess robustness (ability to withstand adversarial prompting). ChatGPT was prompted to rate its confidence from 1-10 (with 10 being the highest level confidence and 1 being the lowest) on the third attempt and after each challenge prompt. Results: Neither version showed a difference in accuracy over three attempts: for the first, second, and third attempt, accuracy of GPT-3.5 was 69.3% (104 of 150), 63.3% (95 of 150), and 60.7% (91 of 150), respectively ( P = .06); and accuracy of GPT-4 was 80.6% (121 of 150), 78.0% (117 of 150), and 76.7% (115 of 150), respectively ( P = .42). Though both GPT-4 and GPT3.5 had only moderate intrarater agreement (kappa = 0.78 and 0.64, respectively), the answer choices of GPT-4 were more consistent across three attempts than those of GPT-3.5 (agreement, 76.7% [115 of 150] vs 61.3% [92 of 150], respectively; P = .006). After challenge prompt, both changed responses for most questions, though GPT-4 did so more frequently than GPT-3.5 (97.3% [146 of 150] vs 71.3% [107 of 150], respectively; P < .001). Both rated "high confidence" (>= 8 on the 1-10 scale) for most initial responses (GPT-3.5, 100% [150 of 150]; and GPT-4, 94.0% [141 of 150]) as well as for incorrect responses (ie, overconfidence; GPT-3.5, 100% [59 of 59]; and GPT-4, 77% [27 of 35], respectively; P = .89). Conclusion: Default GPT-3.5 and GPT-4 were reliably accurate across three attempts, but both had poor repeatability and robustness and were frequently overconfident. GPT-4 was more consistent across attempts than GPT-3.5 but more influenced by an adversarial prompt.