County counsel worries a court-clerk summarizer may invent case citations that were never filed. Which LLM limitation should security emphasize for this legal-adjacent workflow?

LLM factual reliability is limited: a model can produce fluent citations, docket numbers, and case names that look professionally correct but were never filed. In a court-clerk summarizer, invented citations can create integrity and due-process risks if legal-adjacent staff or the public treat generated references as real. Emphasize that fluency is not verification; legal citation workflows need grounding, human review, and systems of record, not confidence based on tone. The claim that LLMs cannot emit Unicode characters used in docket numbers is false; models can generate mixed alphanumeric and punctuation sequences. The claim that LLMs always refuse prompts mentioning a court case is also false: refusals depend on policy, and many models will discuss cases unless constrained. Dedicated hardware does not make citations accurate; where the clerk's office runs servers has no bearing on whether a generated citation is real. Exam caveat: grounded summarization of provided filings differs from open-ended citation generation. Rephrasing text present in source documents lowers the risk, while asking the model to find supporting cases raises hallucination risk; distinguish this from copyright or training-data concerns. Practical check: require retrieval from approved legal databases and human verification, flag identifiers absent from inputs, retain audit logs, and treat summaries as drafts until cite-checked. Define which tasks are assistive drafting versus authoritative legal research.