Background: Science, Technology, Engineering, and Mathematics (STEM) education in Asian universities struggles to integrate Knowledge, Skills, and Attitudes (KSA) due to large classes and student reluctance. While ChatGPT offers solutions, its conventional use may hinder independent critical thinking. Objectives: This study introduces PA-GPT, using ChatGPT as a "virtual peer" in peer assessments to promote active learning and enhance knowledge, higher-order thinking skills (HOTS), and attitudes--the core of KSA in STEM. Methods: A randomised controlled trial involved 61 first-year engineering students (43 males, 18 females) from a university in Southern Taiwan enrolled in "Network Embedded Systems and Applications." Participants, all with prior ChatGPT experience but no programming background, were purposively sampled. They were randomly assigned to the experimental group (n = 31) using PA-GPT or the control group (n = 30) using traditional ChatGPT. Over 8 weeks, data were collected using pre- and post-tests: a knowledge construction test (20 items, [alpha] = 0.85); a HOTS scale ([alpha] = 0.78-0.83) measuring critical thinking, problem-solving, and creativity; and the S-STEM questionnaire ([alpha] >0.80) assessing attitudes towards STEM subjects and 21st-century learning. ANCOVA analysed the data, controlling for pre-test scores, and Levene's test checked homogeneity of variances. Results and Conclusions: ANCOVA results showed that PA-GPT significantly outperformed traditional ChatGPT in enhancing knowledge construction (F = 9.89, p = 0.002), critical thinking (F = 37.00, p < 0.001), problem-solving (F = 9.40, p = 0.003), creativity (F = 7.22, p = 0.009), and attitudes towards mathematics (F = 25.52, p < 0.001), engineering/technology (F = 16.06, p < 0.001), and 21st-century learning (F = 26.38, p < 0.001). These findings demonstrate that PA-GPT effectively addresses challenges in student engagement and HOTS development in STEM education by simulating peer interactions. Peer Assessment with ChatGPT (PA-GPT) promotes active learning and self-reflection, potentially revolutionising AI-assisted education in large class settings. This study provides pioneering evidence for the effectiveness of AI-driven peer assessment in enhancing comprehensive STEM competencies, offering a promising direction for future educational technology integration.