AI legal research tools invent cases for an architectural reason, not an accidental one. A general-purpose language model has no library card. It predicts the next plausible token from patterns in its training data, and when you ask for a citation, out comes a string that looks like a citation because citation-shaped strings are what it has seen.
The reporter volume is plausible, the judge's name is plausible, and the holding sounds like something a court might actually say. None of it has been checked against a real opinion, because no real opinion is in the loop.
Legal AI vendors have spent the last three years engineering around that problem. The fixes are real, they are uneven, and every associate using these tools should understand what the guardrails do before trusting a summary in a brief.
The Fabrication Problem Is a Retrieval Problem
Hallucination in a legal setting isn't the model being creative. It's the model answering a question without access to the source material the answer requires. The National Center for State Courts describes the failure mode plainly in its practitioner guide, noting that generative systems can produce confident, well-formatted output that isn't tied to any real authority. Reported failure modes include fabricated case names, real cases cited for holdings they don't contain, statutes quoted with language that was never enacted, and pinpoint cites that go nowhere.
A sanctioned filing rarely reads as obviously wrong on the page, and that's the whole trouble. The brief looks like a brief. The block quote looks like a block quote, the Bluebook formatting is clean, and opposing counsel is the one who notices. By then the court is already reading.
Prompt Engineering Alone Doesn't Fix It
The intuitive fix, and the one many firms tried first, was to write better instructions. Tell the model not to invent cases. Tell it to say "I don't know," tell it to only cite real authority, and add a system prompt that reminds it every time.
It doesn't work, at least not reliably. A model with no mechanism to verify a citation cannot follow an instruction to only produce verified ones. What it can do is follow the instruction to sound like it is following the instruction, which is worse. Some outputs get more cautious; others append a confident disclaimer to the same fabricated cite. The guardrail is theatrical because the underlying system still has no connection to a legal database.
The other tempting fix is a bigger model. Bigger models do hallucinate less on trivia. On legal citations they still hallucinate, because the problem isn't that the model forgot the case. The case may never have been in the model in the first place, or it was in there as a fragment mixed with a hundred similar fragments. Scale doesn't create ground truth.
Retrieval-Augmented Generation Changes the Architecture
The approach that meaningfully reduces fabrication is retrieval-augmented generation, usually shortened to RAG. The model is paired with a search system that pulls real documents from a legal corpus, and the model is instructed to answer only from those documents. The citation is generated alongside the passage it came from, and the passage is a real one because it was retrieved rather than invented.
This is the design pattern behind the current generation of purpose-built legal research tools, and it's the architecture behind newer vendor offerings as well. When Law.co added AI case summaries and statute research added AI case summaries and statute research to its platform, the pitch was specifically that digests and statute analyzes come back with citations tied to the source material rather than generated freehand. That's the shape of a serious tool now. If a vendor cannot explain where the underlying text comes from and how it's retrieved, treat the output as unverified.
RAG isn't a cure. The Stanford "Hallucination-Free?" study, available as an arXiv preprint, documented meaningful error rates even in retrieval-based commercial legal tools, including citations to real cases that didn't stand for the propositions asserted. What RAG does is change the failure mode from "invented from nothing" to "grounded but sometimes misread," which is a much easier problem for a trained lawyer to catch.
How Firms Should Evaluate a Tool Before Rollout
Buying decisions in this category are still being made on demos. That's a mistake. A demo is optimized. The questions worth asking are duller and more useful.
The Associate Is Still the Last Guardrail
The best-engineered retrieval system in the category still produces output that has to be read by a lawyer before it goes anywhere near a filing. That means pulling the cited case, reading past the quoted passage, checking subsequent history, and confirming the statute is the current version. It's much the same verification work associates have long done, applied to a first draft that arrived faster.
The firms getting this right are the ones treating AI output as a research lead rather than a research product. The tool narrows the pile. A person reads what's left. That division of labor is what makes the speed gains real, and it's what keeps the firm's name out of the next sanctions order.