Study Finds Legal AI Answers Often Cite Wrong Laws

A recent investigation published on arXiv (2608.02621) indicates that large language models (LLMs) often deliver accurate legal responses while referencing incorrect or irrelevant statutes, or the opposite. The study examined 238 items from Taiwan's bar examination and discovered that, for criminal law queries, 24.0–42.4% of valid answers were correct but lacked the appropriate legal authority, whereas 15.2–21.7% were incorrect yet cited the right authority. This disconnect between the correctness of answers and their authoritative basis was evident even with standard reasoning prompts that did not request statutory references. Additionally, a separate statutory-retrieval test and a citation-abstention intervention were included, highlighting the independent nature of answer and citation behavior. The findings imply that scoring based solely on answers in legal assessments may not accurately reflect a model's capability to anchor responses in appropriate legal authority. The research was carried out by an unnamed team and can be found on arXiv.