Why legal AI hallucinates — and what stops it

A hallucinated citation in a legal brief isn't a minor bug. It's a lawyer's credibility and a client's case, staked on a case name a model invented because it sounded plausible.

Where the hallucination actually comes from

Most legal AI tools are a general-purpose language model with a search box bolted on. The model still generates its answer from what it learned during training — which means it can still generate a citation that was never in the documents it retrieved. Retrieval reduces hallucination. It doesn't remove it.

What a verification layer has to do

Stopping this requires treating citation generation as a claim that has to be checked, not text that has to sound right:

• Every citation the model produces gets matched against the retrieved corpus.
• Anything outside that corpus is stripped before the answer ever reaches a screen.
• Anything that can't be confirmed is marked as unconfirmed, not silently dropped.

That last point matters as much as the first two. A system that only shows verified claims but hides its uncertainty is still asking for blind trust — it's just better at hiding when it's wrong.

Retrieval has to match how law actually works

Legal research isn't a single kind of question. Sometimes you need the exact phrase from a section. Sometimes you need everything conceptually related to an argument, regardless of wording. A single retrieval strategy will always be wrong for some fraction of queries.

Running dense, sparse, graph, and keyword retrieval in parallel — then fusing and reranking by authority — covers more of that range than any one method alone.

The standard, not the pitch

None of this is a differentiator you can market with an adjective. It's the minimum bar for a tool a lawyer can actually rely on. That's the standard Juriva is built to.