The Specification Problem of Legal Expert Systems

Automated legal decision-making relies on computer programs called legal expert systems, that are executed on machines not capable of legal reasoning by themselves. Rather, it is up to the programmer to ensure that the behavior of the computer program faithfully captures the letter and intent of the law. This situation is merely an instance of the more general specification problem of computer science. Indeed, the way programs are written and executed requires the programmer to express her intention in a particular form of logic or statistical model imposed by the programming language or framework. On the other hand, the intended behavior of the program or specification, here communicated through the law, is usually described using natural language or domain-specific insights. Hence, every software endeavor begins with a requirement analysis, which consistsin extracting from the specification corpus a set of requirements that the computer system must obey.In the case of automated legal decision-making and legal expert systems, the members of this set of requirements are the possible legal reasoningbits that the computer program is expected to perform. Viewing the problem through this lens immediately allows for identifying the key questions forassessing the safety and correctness of legal expert systems. First, when and how is it possible to express legal reasoning as a set of requirements for a computer system? Second, how to check that these requirements are correctly translated to computer code? Third, can we ensure that the computer code does not introduced unwanted, unlawful behavior? In this article, we take a tour of the general computer science answers to these three questions and assess their efficiency in the particular situation of legal expert systems. To do so, we introduce the distinction between result-constrained and process-constrained legal specifications. From this distinction naturally stem different software solutions, ranging from machine-learning-based to algorithm-based. Finally, we conclude by a discussion about the critical software qualification for legal expert systems, and what this qualification could entail in terms of technical and organizational change