The ability of large language models to reason logically

Abstract

The rise of large language models (LLMs) has sparked a new wave of debate in legal scholarship about the extent to which these systems are capable of replacing human thinking, particularly logical reasoning. This article examines whether LLMs can consistently apply the basic principles of formal logic in the interpretation and subsumption of legal norms. Methodologically, the research is based on a series of experiments that use five types of typical logical errors in legal reasoning. Several models (ChatGPT-5, Gemini 2.5 Pro, Copilot, Grok 4, and Perplexity) were tested at two levels of input: lay and expert prompts. The results show recurring logical inconsistencies across all models, with more convincing outputs only achieved for explicitly formulated expert tasks. The findings underscore the fundamental difference between statistical text generation and actual legal reasoning, which requires logical deduction and transparency. In conclusion, while LLMs can be useful tools to support lawyers, they are not yet capable of fully replacing their work.