Legal text classification in Korean sexual offense cases: from traditional machine learning to large language models with XAI insights

Bigger AI models aren't always better for specialized legal work: Fine-tuned smaller models (KLUE-BERT) achieved 99.3% accuracy in classifying Korean sexual offense cases, outperforming larger general-purpose models like GPT-4.0, suggesting that domain-specific training matters more than model size for legal applications.

AI struggles with implicit legal context: Even high-performing models had difficulty interpreting subtle contextual cues and implicit meanings in legal documents, indicating that AI tools cannot yet fully replace human legal judgment in nuanced cases.

Explainability is essential for responsible legal AI: The study used explainable AI (XAI) techniques to identify which linguistic features influenced model decisions and reveal misclassification patterns, demonstrating that transparency about how AI reaches conclusions is critical before deploying these tools in the justice system.

Real-world legal data exposes AI limitations: When tested on data resembling actual case records (KICS data), the model's performance declined, highlighting the gap between controlled testing environments and practical deployment—an important consideration for responsible AI implementation.

AI should support, not replace, legal professionals: The research frames AI-assisted tools as supporting legal professionals rather than making independent decisions, emphasizing that human expertise remains essential in legal classification tasks where accuracy and fairness have serious consequences.