Hallucination and Citation Errors in Legal AI: What You Must Verify

Here's the most dangerous thing about using AI for legal research: it sounds confident when it's completely wrong. You ask it to cite a case that supports your argument, it gives you a case name, holding, and year—and none of it exists. This is called "hallucination," and it happens regularly, even with good AI systems.

Hallucination isn't a glitch or rare accident. It's how these systems work. AI generates text based on patterns in its training data, not by accessing live case databases. Sometimes those patterns produce plausible-sounding but entirely invented citations. To an untrained reader, they look completely real.

Why This Happens

Large language models (the AI systems you use) don't "know" facts the way you do. They recognize patterns and predict what comes next statistically. A real case name like "Smith v. Jones, 123 F.3d 456 (9th Cir. 2015)" teaches them the *pattern* of how citations look. But they don't distinguish between real cases and made-up ones that fit that pattern.

When you ask AI to cite a case on a specific topic, it generates a citation that *sounds* right based on patterns. If it matches no real case, you don't know until you check.

What Gets Hallucinated Most

Case names and holdings are high-risk. Dates are often wrong. Court jurisdictions get confused. Statute numbers sometimes don't exist. Direct quotes are frequently paraphrased or invented. Any specific legal cite should make you suspicious until verified.

Interestingly, AI is better at general legal principles ("non-competes must be reasonable") than specific citations. If you ask for principles, you'll catch errors more easily. If you ask for "find me a case that says X," you're in hallucination territory.

How to Protect Yourself

Never cite anything from AI without verifying it independently. Use Google Scholar, your jurisdiction's legal database, or a legal research platform. Type the case name as AI provided it. If it doesn't exist, search for variations or similar cases on the topic.

Ask AI to *explain* legal principles in your own words, not to cite specific authorities. Use it for brainstorming what issues might matter, not for the citations themselves.

If you *must* use AI-suggested citations, treat them as starting points for your own research, not as final answers. "AI suggested this case exists—now let me find it and verify the holding." That's the right workflow.

Real Example

You ask ChatGPT: "What case established that non-competes over 3 years are unreasonable?" It confidently responds: "In *Johnson v. Miller, 287 F.3d 112 (3rd Cir. 2003)*, the court held..." You Google it. The case doesn't exist. You've just wasted time and almost cited fiction in your brief.

Try this: Ask an AI system to cite a case on a niche legal topic you know something about. Then search Google Scholar for that exact citation. See if it exists. Most people are shocked at how often it doesn't. This inoculates you against trusting AI citations blindly going forward.