Citation validation 101: why AI hallucinates, and how to catch it

A "hallucination" in large-language-model parlance is a plausible-sounding output that is not grounded in any real source. In legal practice, hallucinations look like cases that do not exist, opinions attributed to the wrong judge, or pin cites that point to the wrong page.

Why models hallucinate

Large language models are trained to produce the next most likely token. They are not retrieval systems. When the model has high confidence in the shape of a citation but low confidence in the content, it tends to commit to a confident-sounding answer rather than say "I don't know." That is the failure mode.

Three controls that actually catch hallucinations

1. Retrieval-augmented generation (RAG). Before generating an answer, retrieve from a real database (Westlaw, Lexis, CourtListener) and constrain the model to that retrieved context.
2. Post-hoc citation validation. After the model writes, parse every citation, look it up in a real database, and surface anything that doesn't match as "unverified."
3. Surface the uncertainty. Show users which citations are verified, which are missing, and which were not checked. This is the part most tools skip.

How Draftiro does it

Draftiro uses post-hoc validation against the CourtListener database — a free, public corpus of US case law maintained by the Free Law Project. Every case citation the AI emits is parsed and validated. Verified citations get a checkmark in the chat UI. Unverified citations get a warning. Statutes, regulations, and rules are surfaced as "unchecked" so you know to confirm them in your primary authority (the United States Code, state statutes, court rules).

Validation is the floor, not the ceiling. You should still read every citation before filing. The point of validation is to catch the silent failures — the ones that look right at a glance.