Nationwide Distribution by the Court Administration Office in March
Guidance on Prompt-Drafting Routines and More
The Court Administration Office has published an artificial intelligence (AI) user guide for judges titled "AI Guidebook for Judges (Prompt Engineering)." As the need to use commercial AI tools such as ChatGPT and Gemini in trial practice has grown, the guide organizes how judges can use AI safely. The guidebook will be distributed to judges nationwide starting in March 2026.
The guidebook presents principles that must be observed when using commercial AI in trials. It first explains how AI works and its technical limitations, and then provides a "practical checklist." This is to help check risk factors such as hallucinations, in which AI presents incorrect information in a plausible manner, data bias, and infringement of personal information. The checklist includes questions such as: Has the AI output been used only as reference material, with the final decision made directly by the judge? Has a separate verification process been conducted on the premise that AI may contain errors? Have the purpose of using AI and the scope of personal data processing been clearly defined?
The guide also presents basic principles for drafting prompts and step-by-step routines. The basic sequence is: inform the AI of the role it should perform → assign the task → set the goal → explain the background → specify the desired answer format → supplement the answer by giving additional instructions based on the first response → verify the result directly.
It introduces various techniques to improve the quality of AI responses through prompt design (for example, providing examples, making the model think step by step, and breaking down complex problems into smaller parts).
The guidebook also includes examples of AI use in trial practice. It presents 20 examples of practical use by field, including general, civil, criminal, administrative, and intellectual property, and classifies them into high, intermediate, and low levels according to the difficulty of prompt writing and the level of technique used.
These include: General - enhancing readability and revising writing style to explain the gist of a judgment (difficulty "low"); General - summarizing public materials to organize the basic facts section of a judgment (low); General - verifying whether precedents cited in the parties' briefs are fabricated (hallucinations) (intermediate); General - translating foreign legal literature (statutes, precedents, academic articles, etc.) (intermediate); Criminal - calculating detention periods using Excel table generation (intermediate); Civil - verifying errors in medical expert opinions, such as calculating loss of earning capacity rates (low); Administrative - automatically identifying and visualizing key statutes and related legal principles in administrative cases (statutory network map) (intermediate); Civil - reviewing the applicability of international conventions using the Chain of Thought technique (high), among others.
In October 2025, the Court Administration Office formed a "Research Group for Producing an AI Guidebook for Judges" and prepared the guidelines. The team was led by Kwon Changhwan, Presiding Judge at the Ansan Branch and senior member of the Judicial AI Committee (36th class of the Judicial Research and Training Institute), and consisted of eight judges and one secretary, for a total of nine members. Presiding Judge Kwon said, "To strengthen the court's AI capabilities, it is important to accumulate practical experience in improving the efficiency of trial work by using commercial AI." Around the world, AI is being used in various fields, not only in law, and for courts too the introduction of AI has become inevitable and public expectations have risen. However, courts in reality face constraints such as budget and personnel. Presiding Judge Kwon explained, "By using the guidebook to strengthen judges' internal literacy capabilities regarding AI and to accumulate diverse examples of AI use in trials, we will be able to discover a broader range of areas in which AI can be used in future trials."
Key Prompt Engineering Methodologies
Example: After presenting [Fact pattern A → Summary A], [Fact pattern B → Summary B], instruct: "Summarize Case C in the same format."
2. Thought Generation (a method of asking the AI model to explicitly reveal its reasoning process)
Example: Request that it show each step leading to the final answer: "Present the response in the order of 1. Organize the facts → 2. Identify the issues → 3. Apply the legal principles → 4. State the conclusion."
3. Decomposition (a method of breaking down a complex problem into smaller problems and solving them in parallel)
Example: "Identify the key issues in the plaintiff's claims, list the required facts applicable to each issue, determine whether each required fact is satisfied, and synthesize those determinations to derive a final conclusion."
4. Ensembling (a technique that combines multiple different prompts or responses to produce a better result)
Example: "Based on the three different example sets A, B, and C, draft language to be used in the decision and then integrate them."
5. Self-Criticism (a method of inducing the AI model to evaluate and improve its own answers)
Example: "Conduct a debate on the given facts and issues from the perspective of counsel for the plaintiff and counsel for the defendant. Each round of discussion must include a specific legal rebuttal to the opposing side's immediately preceding argument."
6. Advanced techniques for reducing hallucinations
Example: "Search a specific precedent database to find relevant current statutes and the latest case law, and reflect them in your answer."
Reporter Park Suyeon, The Legal Times
※This article is based on content supplied by Law Times.
© The Asia Business Daily(www.asiae.co.kr). All rights reserved.
