In twenty-five years of litigation — from IP trials at national firms to regulatory work in emerging technology — I have seen every kind of citation error. Opposing counsel cites the wrong standard. A paralegal transposes a page number. A Westlaw search pulls the district court case when you needed the appellate holding.
Legal AI introduces a new category: citations that were never looked up at all. The model generates a case name, a reporter citation, a court, a year, and a holding that reads like a judge wrote it. Sometimes the case exists but the quote is wrong. Sometimes the case does not exist.
Why it happens
Large language models are pattern completers. They have seen millions of legal documents in training and can produce text that matches the shape of legal writing. That is not research. Research requires retrieving a specific source, reading it, and grounding every assertion in what that source actually says.
What to demand from any legal AI tool
- A live trace showing every query run and every source retrieved or excluded
- Verbatim quotes from retrieved text, not paraphrase-only citations
- Explicit documentation when authority was not found
- A verification pass that flags citations the system cannot match to retrieved sources
- Per-section confidence scoring — not a single “trust me” summary
If a tool cannot show you the retrieval path, treat every citation as unverified. Lexora was built because that standard should be the default, not the exception.