Performance of ChatGPT-4o and ChatGPT-5.2 in child psychiatry assessments: A comparative study

Abstract

Although previous studies have assessed the adequacy of ChatGPT (Generative Pre-Trained Transformer) in psychiatry, no research, to our knowledge, has focused specifically on child psychiatry. This study aimed to evaluate the performance of ChatGPT in child psychiatry by testing it with board-level questions from the question bank of a reference textbook in the field. Seven tests from the textbook were randomized using random.org, and Tests 1,2, and 5 were selected. A total of 600 questions were presented to two versions of ChatGPT: GPT-4o (July 2024) and GPT-5.2 (December 2025). Each response was recorded as correct or incorrect. The questions were further categorized into two groups based on length (short vs.long) and question type (multiple-choice questions–MCQ vs. true/false questions-TFQ). ChatGPT-4o answered 522 (87.0%) of the 600 questions correctly, while GPT-5.2 provided correct answers to 546 (91.0%). The improvement in ChatGPT-4o\'s and ChatGPT-5.2’s accuracy was statistically significant (p < 0.001). No significant differences were found between short and long questions for either version (ChatGPT-4o: p = 0.61; ChatGPT-5.2: p = 0.60). Both versions achieved higher accuracy on MCQs compared to TFQs, this difference was borderline significant for ChatGPT-4o (p < 0.047) and clearly significant for ChatGPT-5.2 (p < 0.009). Both versions of ChatGPT performed successfully, with the latest version, ChatGPT-5.2, showing significantly improved accuracy. As large language models continue to enhance, improvements in accuracy may contribute to reducing certain risks in psychiatry-related tasks.