Sycophants in the Courtroom: Are LLMs Fragile to Juridical Authority and Evolving Legal Standards?

Recent research from arXiv highlights a critical distinction in the application of Large Language Models (LLMs) to medicine versus law. While LLMs have demonstrated success in medical licensing examinations, their performance in legal contexts faces challenges due to the fundamentally different nature of 'truth' and authority in law. Legal reasoning relies less on stable empirical evidence and more on the dynamic, context-dependent application of hierarchical and often contradictory authoritative sources, a paradigm distinct from medicine's empirical basis.

Why it matters

This analysis is strategically important as it underscores fundamental limitations of current LLM architectures when applied to domains requiring nuanced understanding of evolving, hierarchical, and context-dependent authority structures. It warns against uncritically transferring LLM success from one domain to another and highlights the need for tailored development and evaluation methodologies for legal applications.

Key insights

• The success of LLMs in medical licensing examinations has created an expectation of similar competence in legal domains.
• Legal 'truth' is contingent, defined by jurisdiction, temporal validity, and the hierarchy of authoritative sources, unlike the stable empirical evidence in medicine.
• Legal performance often requires determining the applicability, validity, and non-contradiction of external authority, rather than mere inference.
• A new comparative diagnostic framework is introduced to evaluate legal reasoning against medical baselines across axes like knowledge recall.

Source

arXiv — Computers and Society — https://arxiv.org/abs/2608.21409