This study evaluated the accuracy and reproducibility of ChatGPT models, specifically GPT-4 and GPT-4o, by reviewing Japanese-language clinical research protocols and informed consent forms using Japanese prompts. The integration of generative AI technologies into clinical research ethics reviews has the potential to enhance consistency, reduce human error, and decrease the manual effort required to assess complex documents. This study primarily aimed to assess and compare the ability of these models to accurately extract and summarize key elements such as research objectives, study design, and ethical considerations, which are critical for ethical review processes. We developed and optimized custom prompts to improve the performance of the models, focusing on the essential aspects of the protocol and informed consent review. The results showed that GPT-4o achieved an 80% accuracy rate in identifying research objectives and a 100% accuracy rate for research design, indicating superior consistency compared with GPT-4, which, despite being slightly less accurate, still showed significant potential for application in ethics reviews. Furthermore, a comparison between customized GPTs and standard prompts revealed that customized GPTs provided significantly higher reproducibility and accuracy, underscoring the value of fine-tuning and Retrieval-Augmented Generation techniques for enhancing AI-assisted review processes. Additionally, challenges in parsing complex PDF documents were identified, highlighting the importance of standardized document formatting to ensure accurate AI analysis. These findings demonstrate the potential of AI-driven systems to improve the efficiency, accuracy, and standardization of research ethics evaluations, potentially setting new standards for AI integration in clinical research practice.