Japan’s Legal Community with Delayed Progress in Utilizing AI Could Be Seen as That of a Third-Class Country

From Logic Programming to Deep Learning

The development of AI (Artificial Intelligence) has been underway for decades. In the 1980s, the development was mainly based on logic programming.

It is a system intended to automate the answering process by putting rules of experts’ logic and tips on solving problems into a computer.

Reformulation experts’ logic and tips to rule-format from Civil Code, a lot of work by experts in the field is required. It was called an expert system at that time. It was expected that the system enabled to automatically utilize experts’ logic and wisdom.

It was especially expected to be used in the legal world and in medical diagnosis. For example, it is possible to diagnose a disease by putting the knowledge of doctors into logical rules, such as “if the patient has this symptom, then the disease is this or that.”

In the legal world, legal norms make legal inferences with a clear logical structure, which is “a legal requirement gives legal effect.” For example, the provision for compensation for loss or damage in Civil Code Article 709 provides that “a person who has intentionally or negligently infringed any right of others, or legally protected the interest of others” (legal requirement) “shall be liable to compensate for any damages resulting in a consequence” (legal effect).

Therefore, it was thought that if the Civil Code and other laws and judicial precedents were made into such logical rules and put into a computer, an automatic judgement could be made in civil lawsuits and other cases.

However, it required a huge amount of work, such as manually reformulating Civil Code provisions and judicial precedents into logical rule format. Also, the necessary natural language processing did not progress easily nor quickly enough.

Meanwhile, in the world of computing, approaches such as machine learning and deep learning made significant progress in the 2000s. A major catalyst was that the backpropagation algorithm for efficient learning on brain-like neural networks was developed in the late 1980s. The significant improvement in the computing power of computers advanced practical applications in the 2000s.

Just as the brain thinks and learns by creating new synaptic connections between brain cells and adjusting the strength of the connections, the algorithm adjusts the connections and strength between multilayered nodes that correspond to brain cells to provide a mapping between problems and answers.

The computer suggests possible answers to an example problem and adjusts the connections and strength between nodes to match the correct answers so that the answers are closer to the correct ones. This is “learning,” and if the computer learns more sets of example problems and answers, it will get smarter. There are many layers of nodes, so this is called deep learning. It requires a great amount of data and a huge amount of computation.

Since the late 1990s, improvements in the performance of computers have made these learning methods more effective and faster, and deep learning through neural networks has become the central method of AI development. In this way, deep learning does not require rulemaking or linguistics, which makes it possible to recognize images such as identifying human faces.

“AlphaZero,” Which Grows Without Teaching Signals, and the Law

In computational deep learning, data regarding example problems and their correct answers are called teaching signals. The more teaching signals given, the more AI learns and becomes smarter.

For example, the well-known “AlphaGo” grew to be a strong AI Go player by automatically learning a large number of game records of professional Go players as teaching signals.

However, from 2016 to 2017, in online Go games, there appeared a player who called itself a master and defeated the players considered to be the world’s top players one after another. This player was “AlphaZero,” an evolutional version of “AlphaGo.”

In fact, unlike “AlphaGo,” teaching signals are not given to “AlphaZero.” Only the rules of Go were provided, and it learned the rest by continuously playing against opponents created in a virtual space.

As a result, it built the optimal neural network for winning the game of Go by itself in a short period of time. The rules of Go clearly determine winning or losing the game, so it created teaching signals on its own, and learned from them. It is an autodidactic system where it studies by itself.

It was very shocking that people considered to be the world’s top Go players could not compete with “AlphaZero,” but it is short-sighted to believe that AI surpassed humankind’s intelligence.

It only means that in games with clear rules such as Go, AI can quickly establish a way to surpass humans by automatic learning for the correct answers to win. AI has exceeded humans only in terms of the amount and the speed of computation (actually, it has exceeded humans for decades).

In short, there are two ways for automatic learning in AI. One is to give teaching signals. In this case, the more teaching signal data provided, the more AI learns and improves on its performance. In other words, it is getting cleverer.

The other is to teach only the rules that determine the correct answers to problems. In this case, the rule is clear, so the determination result is also clear. AI does not know the process in the middle, for example, how to play the correct move in the game of Go. AI creates various processes (moves) from the problem to the correct answer on its own, and learns the process with a better result (win or lose) through trial and error. If the rules are clear like this, AI can grow without teaching signals.

AI can also learn beyond the teaching signals that people have experienced or expect and exceed human performance.

However, the real world is not based on the rules as simple and clear as a game. The laws and judicial precedents, which are the rules of society, are written in natural language, and ambiguity remains to a certain degree. Moreover, it is clear that the court’s decision is not always correct, as there are cases where the decision is overturned by the appeal or divided among the courts, and a precedent is changed later. If we want to create AI that can be used in various situations in the real world, for now, giving a lot of teaching signals will be the only way available to lead to automatic learning in AI.

The Legal Communities in the World Are Moving Toward the Introduction of AI

AI has begun to be applied around us in various ways. Of course, the risk of abuse and improper use has become a reality.

Examples of applications include junk mail filters. The data in junk mail are provided to AI for learning. It seems that a variety of applied learning algorithms, such as deep learning, Bayesian networks using the Bayesian theory to derive a posterior probability by modifying the prior probability based on data, and combinations of these, are used according to each system. As a result, accuracy has been dramatically improved. Bayesian networks and deep learning have also been applied practically to marketing by the use of POS system data (called AI marketing).

The use of AI image recognition has also improved the performance of individual identification and facial recognition. Image and video recognition technology is also used in the medical field. Based on various test images, it has become possible to establish a system to detect subtle abnormalities that doctors may have overlooked or not noticed, and to quickly produce diagnostic results.

As in the case of healthcare, practical applications have been making progress in the legal community, in which the use of AI has been expected from the early stage of AI development.

For example, in the United States, a civil trial requires a huge amount of documentary evidence from both sides under the system of discovery (the disclosure of evidence). In the past, lawyers read it in parts, so in some cases, hundreds or even thousands of lawyers needed to be mobilized.

However, most documents, both corporate and private, are created electronically, and paper documents can also be scanned into an electronic file using OCR, so many documents are submitted in electronic files. This is called eDiscovery. Even if the number of electronic files is huge, there are important keywords and statements in a lawsuit, so AI can sift documents by parsing the files and identifying them. In other words, AI’s support has significantly reduced the time and labor of lawyers.

In addition, European nations, the United States, China, and Japan have advanced research to introduce AI support into factual finding and legal decision-making. The Bayesian networks described earlier are being applied to fact-finding where AI determines whether the facts and allegations are true or false based on evidence and testimony. There are also attempts to use AI to make legal decision-making of applying laws and judicial precedents by deploying deep learning and natural language processing. The attempt to use AI for legal inference, which guides the court’s conclusion by inferring legal reasoning between multiple relevant legal norms, by applying logic programming, described earlier, began 40 years ago, and has now reached a level that can be used in the legal education.

In order to make AI learn to the point where these attempts have reached practical levels, the data of judgements (the main text and reason) and case records (evidence such as photographs and testimonies, complaints, answers, and briefs) are essential as teaching signals. The more data provided, the more AI learns.

However, in Japan, it is said that less than 10% of judgements are released in law reports, and no case record is included.

Although anyone can have access to case records by specifying the name of the party or the case number, those who are not involved in the case do not know the name of the party and the case number in the first place, and even if they can have access to it, those who are not a party or a stakeholder are not allowed to zerox it. This makes it very difficult to advance the learning in AI.

On the other hand, in the United States, federal court records are released on the website PACER (Public Access to Court Electronic Records), which is operated by courts, under the philosophy of expanding the right of Americans to access court information.

The data has reached one billion cases in both criminal and civil cases, and anyone can view it if they register on the internet and pay a certain fee. Of course, it can also be utilized as teaching signals for learning in AI.

It can be said that this is due to the differences in American and Japanese attitudes to privacy and protection of personal information (by the way, law reports in the United States use real names, but those in Japan keep names anonymous).

However, considering the world’s situation in which various ICT technologies, including AI, have become widely spread in society and will continue to spread even wider, it is important to balance the protection of personal information with systems that make effective use of data. Without this in mind, Japan could be left behind in the world, and the Japanese people and Japanese society would not be able to fully enjoy the benefits of the development of AI and ICT.

In fact, the adoption of IT in the legal community has been promoted not only in the United States but also around the world. In Spain and elsewhere, all case records, including the use of court footage, are electronic and completely paperless. Clearly, it will lead to the introduction of AI to support courts. In France, in most civil cases it has become mandatory to use ADR (Alternative Dispute Resolution), including ODR (Online Dispute Resolution) in which AI and IT are used, before filing to the court. Singapore and South Korea are also far ahead of Japan in IT adoption in the legal community.

On the other hand, Japanese courts are also moving toward the adoption of IT, but the immediate goal is to use e-mail. Even now, paper documents and faxes are dominant. Possible applications of IT in civil cases are e-Filing, e-Court, and e-Case Management. Progress has been slow, but telework has been expanded owing to the COVID-19 pandemic, and there may be unexpected sudden developments.

To Prevent “AI Defeat” in the Japanese Legal Community

In the Carlos Ghosn case, people around the world pointed out that Japan’s judicial system was out of date, and if things continue as they are, Japan’s legal community could be regarded as second class or third class in a globalized world.

I think the declining birthrate and aging population will continue. The population has started to fall. In this situation, it is essential for Japanese society to introduce IT, AI, and robots. If the courts alone continue to use paper and faxes, and report only a fraction of cases in law reports, it is no wonder it is criticized that they have not improved since the Middle Ages.

The legal community in Japan also needs to change, and I think the use of AI will be the major pillar.

In the midst of the wave of DX adoption, I think everyone, including those working in the private sector, has already been experiencing the already-started corporate transformation for survival, but it is important for each to continue learning for the future. This is not about matter of taste. It is about survival.

AI has made rapid progress in the last few years. We need to make good use of those technologies. This is not a question of whether we can use them. It is a question of how to compete in the world by making full use of them. We need to keep our eyes on the future.

* The information contained herein is current as of September 2021.
* The contents of articles on Meiji.net are based on the personal ideas and opinions of the author and do not indicate the official opinion of Meiji University.
* I work to achieve SDGs related to the educational and research themes that I am currently engaged in.

Information noted in the articles and videos, such as positions and affiliations, are current at the time of production.