Why 36% of AI-Generated Legal Citations Are Wrong

The Problem Nobody Wants to Talk About

A Stanford and Yale study found that 36% of citations generated by GPT-4 are completely fabricated. They look real — plausible case names, proper citation formats, convincing holdings — but the cases don't exist.

This isn't a theoretical risk. As of 2026, there are now 729+ documented court cases where attorneys submitted AI-generated citations that turned out to be fake.

Real Consequences

The most famous example is Mata v. Avianca (2023), where a New York attorney submitted a brief containing six fabricated cases generated by ChatGPT. The attorney was sanctioned, publicly embarrassed, and the case became a cautionary tale covered by every legal publication.

But it keeps happening. Why?

Why Large Language Models Hallucinate

ChatGPT, Claude, Gemini — they're all language models, not databases. When you ask for a citation, they don't look it up in a case law database. They generate text that looks like a citation based on patterns they learned during training.

The model predicts: "A case about landlord-tenant disputes in Utah would probably be cited as something like Smith v. Johnson, 2019 UT App 47." It sounds real. It follows the right format. But it's fiction.

The Architecture Gap

The fundamental problem is that general-purpose AI tools skip the retrieval step. They go straight from question to answer, generating everything — including citations — from patterns.

Retrieval-Augmented Generation (RAG) solves this by adding a retrieval step before generation:

1. Your question is converted to an embedding (a mathematical representation)
2. That embedding is matched against a real database of case law and statutes
3. The AI generates its answer using only the real sources it found

Every citation in the response links to a document retrieved from the database. That removes the main driver of hallucination — the model is citing documents it actually retrieved, not generating citations from memory — though retrieval reduces the risk rather than eliminating it, which is why anything that cannot be grounded in a retrieved source gets flagged instead of asserted.

Dual-RAG: Searching Two Databases Simultaneously

CaseRead.ai takes RAG a step further with Dual-RAG — simultaneously searching:

• Public case law and statutes (federal and state databases)
• Your firm's private documents (briefs, memos, precedents in your secure vault)

This means a single research query can cite both binding precedent and your firm's winning arguments from a similar case three years ago.

What This Means for Your Practice

If you're using ChatGPT or any general-purpose AI for legal research, every citation needs manual verification. That defeats the purpose of using AI to save time. (We break the trade-off down in detail in CaseRead vs. ChatGPT for legal research.) And if an AI-drafted brief is already sitting on your desk, you can check the citations in it for free — up to 25 per run — before it goes anywhere near a filing.

With a RAG-based system like CaseRead, citations are grounded at the architecture level: answers cite only sources the system actually retrieved, and anything it cannot verify is flagged rather than asserted. You still exercise professional judgment about which cases to cite — and the final check before filing is still yours. If you're weighing what a purpose-built tool costs against the incumbents, our guide to Westlaw alternatives for small firms compares the options on exactly this citation-integrity test.

Try It Yourself

Search any legal question on CaseRead.ai. Click any citation — it links to the source the system actually retrieved, and anything the system couldn't verify arrives flagged.

Start free — no credit card required.