AI legal research without fake citations: the workflow that works.
Legal research is where lawyers use AI most — around four in five legal AI users, by 2026 survey counts. It is also where AI’s most famous failure lives: filings citing cases that do not exist, which courts have sanctioned since the 2023 Mata v. Avianca decision. Both facts are true at once. The gap between them is a workflow.
The model may suggest authority. It may never be the source of it. Every citation is retrieved from a real reporter and checked before a lawyer reads the draft.
Why models invent cases
A language model writes what a plausible answer looks like. A plausible legal answer contains citations, so it produces citation-shaped text. Asking it to be careful does not fix that. Changing where citations come from does.
The workflow
1. Retrieve first. Search your research subscriptions and your firm’s own memos, and give the model only what was found.
2. Cite only what was retrieved. The drafting step can quote and cite nothing outside the retrieved set.
3. Verify every citation. Check that each exists, is quoted accurately and is still good law — before the draft is shown to anyone.
4. Show the source. Every proposition in the memo links to the passage it relies on.
5. A lawyer reads before anything is filed. No exceptions, and the system records who did.
What it saves when done properly
A research memo that takes an associate six hours typically becomes about an hour and a half including review — the time moves from finding to reading and judgment. Across forty memos a month, that is roughly 180 associate hours.
What to ask a legal AI vendor
Where do citations come from? What happens when no authority is found — does it say so? Can I click from every proposition to its source? Is our matter data used to train anything?
A worked example: a lease question
An associate asks whether a New York commercial tenant can withhold rent for a landlord’s failure to repair. The agent searches the firm’s research subscription and memo bank first, finds the controlling rule and the narrow exceptions, drafts a short memo citing only what it retrieved, verifies all twelve citations against the reporter, and flags that the client’s lease contains a clause that may change the answer. The associate spends their time on the lease clause — the judgment — rather than on finding and checking cases.
Common mistakes
Using a general chatbot for research. Without retrieval from authoritative sources and citation checks, fabricated authority is a matter of time.
Checking only that a case exists. A real case cited for the wrong proposition is still a problem. Verify the quoted passage, not just the citation.
No record of review. If something is filed, the firm should be able to show which lawyer reviewed it and when.
Our legal research agent answers a New York lease question with twelve authorities, every citation verified — and flags the lease clause that could change the answer.
Frequently asked questions
Why does AI make up legal citations?
Language models generate text that looks like a plausible answer. Without being restricted to retrieved, real sources, they can produce citations that look right but do not exist.
How can a law firm use AI for research safely?
Retrieve from authoritative sources first, let the model cite only what was retrieved, verify every citation and quote automatically, link each proposition to its source, and require lawyer review before filing.
How much time does AI legal research save?
In our modelling a six-hour research memo becomes about an hour and a half including review — roughly 180 associate hours a month for a firm writing forty memos.