Computational Models of Legal Reasoning: A Research Study of the Theoretical Foundations and Practical Limits of Machine Decision-Making

Abstract

Computational legal reasoning has developed from early rule-based expert systems into a
diverse field encompassing defeasible logic, case-based reasoning, formal argumentation,
value-sensitive models, statistical prediction, machine learning, and large language models.
These approaches seek to represent, reconstruct, predict, or assist legal reasoning by
translating selected features of legal judgment into computationally processable structures.
Their development has generated an important jurisprudential question: which parts of legal
reasoning can be formalised, and which depend upon interpretation, institutional authority,
contextual judgment, contested values, or forms of practical reasoning that resist complete
computational representation? This study examines the principal theoretical foundations of
computational legal reasoning and evaluates the practical limits of machine decision-making
in legal contexts. It analyses rule-based and defeasible reasoning, precedent-based models,
argumentation frameworks, value-sensitive approaches, data-driven prediction, and
contemporary language-model systems. Particular attention is given to open-textured legal
concepts, analogical reasoning, conflicts among values, uncertainty in facts, source hierarchy,
model hallucination, benchmark limitations, and the institutional legitimacy of automated
legal decisions. The study argues that computational models are most effective when they
represent bounded aspects of legal reasoning rather than when they are treated as substitutes
for the entire judicial process. Formal systems can improve consistency, retrieval, argument
mapping, and structured comparison, while statistical systems can detect patterns and assist
prediction. However, neither approach eliminates the need for legal interpretation,
evidentiary judgment, normative evaluation, and authoritative human responsibility. The
article concludes that the most defensible future lies in hybrid systems in which computation
supports legal reasoning while final normative responsibility remains embedded in legally
accountable institutions.