Not Explainable but Verifiable
Paul Burgess, Ehsan Shareghi · 2025
Abstract Large language models (LLMs) as a form of generative artificial intelligence (GenAI) systems have caused both excitement and fear in the legal community. These technologies have considerable potential to revolutionize the way that legal answers can be derived. This revolution could relate to the relative speed, efficiency, cost, accuracy, and availability of legal solutions; it could change the way in which the law is applied by lawyers or judges; or its use could impact the way in which the legal system operates. There are, however, fundamental problems in the use of these technologies that prevent this revolution from being realized. This article focuses on and provides ways to overcome two of these: the propensity of GenAI systems to hallucinate and the inability for GenAI outputs to be explained. It gives reasons that these are fundamentally problematic when answering legal questions and giving legal advice and then sketches the design of a system that can overcome both: a verifiable language agent. The article then sets out what a verifiable agent is and explains how it addresses the aforementioned shortcomings. In doing this, it not only identifies that a verifiable agent would allow the potential inherent in LLMs to be realized in answering legal questions but also identifies how such an agent could work to do so.