Legal AI Based on Similarity

Research aiming to introduce AI into the legal field began in the 1970s. After a long period of experimental and theoretical studies, implementation in legal practice has begun in recent years under the names of LegalTech and RegTech. When referring to such fields in this paper, I will follow convention and use the term "legal AI." I have been conducting research in this area since the 1980s. In this paper, I present my research on legal AI with particular focus on similarity in the legal domain. I have also discussed legal AI from the perspective of similarity in my earlier works, and published a report in the Journal of the Japanese Society for Artificial Intelligence in 2002.^[1]

In the judicial field, the approach to rendering judgments differs between countries with common law systems, such as the United Kingdom and the United States, and countries with codified law systems, such as Germany, France, and Japan. I will begin by discussing the common law approach of the Anglo-American system. The basic principle of common law is to render the same judgment in similar cases. Of course, codified statutes are also applied; however, in the process of judicial legal reasoning, prior decisions that serve as similar precedents provide the basis for the current judgment. According to a specialist in the history of legal thought, British courts inductively derive the law by accumulating judgments on similar cases. This process is based on an empirical approach that underlies British thought.

When engaging in legal reasoning based on this common law approach, the decisive point becomes what constitutes similarity between precedents and the current cases. Using the term "similarity" in reference to cases means that they share common characteristics and possess the same qualities. Accordingly, depending on the characteristics being examined, there may be aspects that are common and aspects that differ. Ultimately, the pertinent question is whether or not the characteristics under consideration are legally essential to the case. Making this determination is an extremely difficult and complicated process that often involves various values, interpretations, context, and the delineation of said context. Nevertheless, judgment can be facilitated through visualization and enumeration of such characteristics. Consequently, mechanical simulation becomes feasible to a certain extent. This is precisely exemplified by the HYPO legal dispute system, which is a pioneering study in the field of legal AI developed by K. Ashley.^[2] This system analyzed and enumerated essential characteristics as case data, accumulated the results as internal data, and used the results as a basis for determining the similarity between a case and precedents, as well as the relative strength of opposing arguments.

Machine learning, including learning based on neural networks, has become mainstream in AI in recent years. Machine learning ultimately extracts similarities from a large amount of learning data and leads to certain conclusions. As such, it can be considered a form of inductive reasoning. This gives it a high degree of affinity with the common law approach. However, a significant issue with AI-based learning is its low level of explainability. In particular, explanation and justification are essential in the legal field. Consequently, a neuro-symbolic approach has recently emerged in the field of legal AI research. This combines current AI techniques with symbolic processing and knowledge-based inference methods, which were dominant during the second AI boom and particularly strong in driving explanations and proofs.

It should be noted that precedents are often used for reference even in codified law countries such as Japan. In such cases, the final judgment is based on statutes or other rules, so a precedent itself is not directly applied as law; however, it can be said to have an indirect effect. Furthermore, in codified law countries, it may appear that statutes are applied without reference to similarity with prior cases. However, those statutes themselves can be seen as having formalized the legally significant common characteristics described above as requirements. Therefore, this can be seen as placing focus on similarity.

The field of legislation encompasses various activities and areas of study; for example, political activities and considerations of national principles. However, in this paper, I will focus specifically on the aspect of drafting legal texts such as statutes and ordinances.

In practice, when speaking with staff from local governments and central ministries, they expressed a strong desire to follow precedents as much as possible and avoid originality when drafting statutes or ordinances. Moreover, at least in the legal affairs of local governments, methods such as benchmarking are well established. Benchmarking involves referring to several ordinances from other municipalities, creating a comparison table (benchmark table), and using the table to draft one's own municipal ordinances. This approach is also recommended by researchers specializing in municipal legal affairs.^[3] Therefore, even in the legislative field, if similar policy objectives exist, there is a natural tendency to create provisions with similar wording and structure when actually drafting legal texts. Once again, this highlights the importance of similarity.

From among my research, I will now briefly introduce examples of studies in the judicial and legislative fields focusing on AI centered on the concept of similarity as discussed above.

(1) Legal analogy simulation (judiciary)

The analogical application of legal provisions is often used in civil litigation. Here, using actual provisions from the Civil Code would be overly complex, so let us consider a simpler rule as an example: "Vehicles are prohibited from entering this park." When thinking of vehicles, people generally imagine automobiles or trains. Furthermore, most people commonly recognize motorcycles and bicycles as vehicles. But what about baby strollers or skateboards? In such cases, if the purpose of the rule is to avoid hazards that vehicles may pose within the park, a baby stroller carrying an infant might be considered acceptable, whereas a skateboard could be regarded as dangerous. In legal analogy, the assessment of similarity often takes into account whether the similarity is supported by the purpose or rationale of the rule, and analogical reasoning is then applied accordingly. I have conducted research using mathematical logic to formalize this process of legal analogy, thereby developing an AI capable of legal analogical reasoning.^[4]

(2) Automatic synthesis of eLen ordinance templates (legislation)

The eLen Ordinance Database is an online database system which stores ordinances and regulations from municipalities throughout Japan (https://elen.ls.kagoshima-u.ac.jp/). I developed eLen as a research outcome in 2012, and it is currently operated at Kagoshima University. The database covers approximately 98% of Japan's 1,790 local government bodies and is updated three to four times per year. In addition to the standard database functions for searching ordinance data, the system provides a function to automatically generate the benchmarking tables which I introduced above. These tables are used as a reference by municipal staff when drafting ordinances. Furthermore, since 2022, the eLen Ordinance Database has included a function that automatically synthesizes and presents ordinance templates as draft proposals for writing ordinances from groups (clusters) of similar ordinances, and the function can be used by anyone.^[5] This function is another outcome of my research focusing on similarity in legislative work. In 2025, I presented a demonstration of the database at the International Conference on Artificial Intelligence and Law (ICAIL).^[6]

In this paper, I briefly discussed how similarities can be used to take a bird's-eye view when introducing AI into the legal field. I also briefly presented some of my research on legal AI from this perspective. I hope that more people will develop an interest in this field of study.

---

[1] Kakuta, T. and Haraguchi, M. "Similarities in Legal Reasoning: From a Perspective on Dialectics and Arguments," Journal of the Japanese Society for Artificial Intelligence, Vol. 17, No. 1, pp. 14 to 21, 2002.
[2] Ashley, Kevin. "Modelling Legal Argument: Reasoning with cases and hypotheticals," The MIT Press, 1990.
[3] Tanaka, T. "Challenges in Drafting Ordinances: Utilizing Benchmarking Methods," Shinzansha Publishing, 2002.
[4] Kakuta, T. and Haraguchi, M. "A Demonstration of a Legal Reasoning System Based on Teleological Analogies," Proceedings of 7th ICAIL, pp. 196 to 205, 1999.
[5] Kakuta, T., Shima, A., Saito, D., and Otani, T. "Development of eLen Database for Regulations of the Local Governments and a Quantitative Survey of the Regulations," Information Network Law Review, Vol. 13, No. 1, pp. 14 to 33, 2014.
[6] Kakuta, T., et al. "Ordinance Template Composition System-Supporting legislative drafting in the eLen database system," Proceedings of 20th ICAIL, pp. 471 to 472, 2025.