In March 2023, a New York attorney submitted a legal brief citing six cases that didn't exist. They had been fabricated by ChatGPT. The attorney was sanctioned. The incident made international news and set back the adoption of AI in legal practice by years.

When we built Mr.Lawyer, we treated citation hallucination as an existential risk — not a product quality issue. Here's the verification architecture we built to eliminate it.

Why LLMs Hallucinate Legal Citations

Large language models are trained to produce plausible text. In legal contexts, a plausible citation looks like a real case name, a realistic court, a believable year, and a reasonable-sounding holding. The model has seen thousands of real citations during training and can generate convincing-looking fake ones with no additional effort.

The problem is compounded by the nature of legal research: users often don't know what cases exist, so they can't easily spot a fabricated one. A hallucinated citation from a lower court in an obscure jurisdiction is nearly impossible to catch without verification.

The Mr.Lawyer Verification Architecture

Layer 1: Retrieval-Only Citation Generation

Mr.Lawyer never asks the LLM to generate citations from memory. Every citation in a Mr.Lawyer response must come from a verified source in our legal database. We use a retrieval-augmented generation (RAG) architecture where the LLM is explicitly instructed to cite only from the retrieved context — and is penalised in our evaluation framework for any citation that doesn't appear in the retrieved documents.

Layer 2: Citation Existence Verification

Before any citation reaches the user, it is checked against our legal database. We maintain a comprehensive index of Indian case law, statutes, and regulations. If a citation doesn't resolve to a document in our database, it is flagged and removed from the response.

Layer 3: Holding Verification

Existence is necessary but not sufficient. A real case can be cited for a holding it doesn't actually support. Our third verification layer checks that the holding attributed to a case in the LLM's response is consistent with the actual text of the judgment. We use a separate verification model that reads the case and the attributed holding and produces a consistency score.

Layer 4: Source Linking

Every citation in a Mr.Lawyer response includes a direct link to the source document. Users can verify every citation themselves in one click. This isn't just a UX feature — it's a forcing function that keeps our verification pipeline honest.

The Result

Since launch, Mr.Lawyer has produced zero fabricated citations in production. This isn't because the LLM never tries to hallucinate — it does. It's because the verification architecture catches and removes hallucinations before they reach the user.

The verification pipeline adds latency — typically 2–4 seconds per response. We've found that legal professionals consider this an acceptable tradeoff for guaranteed citation accuracy. In legal practice, a fast wrong answer is worse than a slightly slower right one.