Current Korean court ruling disclosure systems inadequately balance public access and privacy regarding Artificial Intelligence (AI) owing to low disclosure rates and inefficient manual anonymization. This results in hindering AI development as well as the public’s right to information. This study proposes and evaluates a generative AI-based automatic anonymization system using fine-tuned Small Language Models (sLLMs, 7-8B scale), designed for secure operations within courts. Using actual court data, we compared the performance of fine-tuned sLLMs with that of few-shot Large Language Models (LLMs, e.g., GPT-4o). The results demonstrate that fine-tuned sLLMs achieve high accuracy and recall (F1 > 98%), compared to LLMs, proving their feasibility for safe and efficient internal automation without data leakage risks. Furthermore, we analyzed the legal liability challenges concerning potential AI anonymization errors. Reviewing the applicability of existing doctrines revealed legal uncertainties surrounding AI system failures. Fine-tuned sLLMs offer a realistic technological solution for harmonizing privacy protection and access rights. This study underscores the need for supportive legal and institutional frameworks to enhance judicial transparency and to advance the legal AI sector.