Knowing the Form, Not the Function: Measuring Wild Institutional Hallucination in LLMs

This preprint introduces Answer–Justification Decoupling (AJD), a diagnostic phenomenon in which legal language models may produce correct answers while failing to identify or generate the correct legal authority, or conversely may invoke grounded legal authorities without reaching the correct answer. Using Taiwan’s bar examination as a measurement instrument rather than a legal benchmark, the study separates answer correctness, citation grounding, grounding state, and epistemic commitment across four large language models and 236 questions. The paper identifies a naturally occurring but objectively verifiable form of institutional hallucination: statutory authority markers that are not explicitly requested, arise inside expert-register legal reasoning, and can be checked against an official legal database. It reports high rates of unrequested statutory citation, Answer–Justification Decoupling, the Admission Paradox, and a citation-abstention intervention showing that citation behavior can be suppressed independently of answer accuracy. This is a timestamped preprint version corresponding to the short-paper version submitted to ARR/EMNLP. A longer version with additional authority-discrimination experiments is in preparation.