The Citation Trap

A Verification Checklist for AI-Assisted Legal Research. Why LLMs fabricate citations, a five-step verification protocol, safe research prompts, and what to do when you find a hallucinated case.

A Verification Checklist for AI-Assisted Legal Research


Why Citations Are the Highest-Risk Output

An LLM generates case names, neutral citations, paragraph numbers, and reported law reports the same way it generates everything else: by predicting what a plausible citation looks like. It has no connection to a legal database. It does not check whether the case exists.

The result is citations that look exactly right — correct format, plausible names, realistic year ranges — but may be partially or entirely fabricated.

This is not a rare edge case. In documented incidents across multiple jurisdictions — including the US, UK, Australia, and India — lawyers have filed briefs citing cases that do not exist.

In several cases, sanctions followed. In all cases, professional embarrassment was significant.


Why It Happens

  • The model is optimised to produce fluent, coherent, contextually appropriate text — not to produce only verified facts.
  • Legal citation formats are highly regular, making them easy to fabricate plausibly.
  • The model cannot distinguish between what it ‘knows’ and what it is constructing.
  • Confident tone is not a reliability signal. The model is equally fluent when wrong.

The Five-Step Verification Protocol

Apply this to every piece of AI-generated legal research before use.

StepActionTool
1Identify every case name and citation in the AI outputRead carefully — do not skim
2Verify existence on a primary legal databaseSCC Online / Manupatra / BAILII / Westlaw / LexisNexis
3Confirm the legal proposition attributed to the caseRead the judgment, not just the headnote
4Check the case has not been overruled or distinguishedUse the citator / treatment history function
5Sign off: record that verification was completedFile note or matter management system entry

Verification Checklist

For each AI-generated citation, confirm every item before use:

  • Case name confirmed on primary database
  • Neutral citation / report citation confirmed
  • Judgment text accessed and ratio verified
  • Case not overruled, reversed, or distinguished on the relevant point
  • Paragraph or page reference confirmed in the judgment
  • Statute section confirmed in current consolidated version
  • Verification recorded in the file

Safe Research Prompts

These prompts reduce hallucination risk by not asking the model to generate citations.

Instead of: “Find cases on X” Try: “What legal issues arise from X? I will find the cases myself.”


Instead of: “Cite the leading authority on Y” Try: “Summarise the legal framework for Y. Do not include citations — I will verify them separately.”


Instead of: “Is this clause enforceable under Indian law?” Try: “What are the typical enforceability arguments for and against this type of clause? Flag which areas I should research.”


What To Do If You Find a Fabricated Citation

  1. Do not use the output. Stop the research workflow immediately.
  2. Check downstream documents. If any letter, brief, or advice note has already used the citation, correct it before it leaves the firm.
  3. Report internally if your firm has an AI governance log.
  4. Regenerate the research using verified databases only.

Part of the AI Foundations for Lawyers series.