IMDA cuts Singapore court judgment summarization from days to minutes with GPT-Legal on AWS

Legal research in Singapore traditionally required manually reviewing lengthy court judgments, some exceeding 200 pages, taking legal professionals 5 to 10 hours per case. Existing general-purpose LLMs could not handle Singapore's specific legal terminology and mixed-language content, making a customized solution necessary for the 75% of legal practitioners who rely on the LawNet research platform.

IMDA's BizTech Group partnered with the Singapore Academy of Law (SAL) and the AWS Generative AI Innovation Center to build GPT-Legal, a domain-specific fine-tuned LLM deployed on Amazon SageMaker. The team used SageMaker's data lineage capabilities to track public and proprietary training data for compliance, and used Amazon S3 and EC2 for testing and evaluation. Validation tooling was integrated to detect and flag potential hallucinations in generated summaries so researchers can cross-check output.

Launched in September 2024, GPT-Legal now generates catchwords, facts, and holdings for unreported judgments, cutting summarization time from days to minutes. The tool has processed and summarized over 15,000 Singapore court judgments, improving access to case information for legal professionals via LawNet AI.

• Domain-specific fine-tuning (legal terminology, mixed-language content) was necessary because general-purpose LLMs underperformed on Singapore legal text.
• Built-in hallucination-detection/validation tooling was treated as essential to earning trust for AI-generated legal summaries.
• Success with a narrow legal-research use case is driving exploration of extending the same summarization approach to contracts, prospectuses, and other sectors like healthcare.