How to Catch a Hallucinated Case Citation
Short answer: The only reliable way to catch a fabricated citation is to verify it at the source. Three stages are enough: does the decision exist, are the reference details internally consistent, and does the decision actually say what is claimed. The third stage is the most often skipped and the one where most errors hide.
Language models fill gaps when information is missing. In law this produces citations that look formally perfect but do not exist: a correctly formatted docket number, a plausible date, a convincing summary. The problem is that none of this is detectable at a glance.
The checklist below defines the three stages a citation must pass before it reaches a filing.
Stage 1: Does the decision exist
Search the citation in the official source. If a search by number returns nothing, there are two possibilities: the decision was not published, or it does not exist. Checking neighbouring numbers for the same chamber and year separates the two.
Warning sign: If the system cites a decision but cannot produce a link, that citation is unverified. A sentence asserting that a court has ruled a certain way means nothing without a number and a link.
Stage 2: Are the details consistent
In fabricated citations the most common flaw is that the components do not agree with each other. Check these four.
- Chamber and subject match. A criminal chamber number attached to an employment dispute means the citation is fabricated.
- Filing year and decision year. A decision cannot predate its filing, and the gap should correspond to a plausible proceeding length.
- Number format. The format should match the known convention for that court and chamber.
- Did the chamber exist then. Court structures change over time; a recent decision attributed to a dissolved chamber is not possible.
Stage 3: Does it support the claim
This is the stage most often skipped and where most errors are found. The decision may genuinely exist, its details may be consistent, and it may still not contain the asserted holding. A system can link to a real decision and claim something that decision does not say.
Open the full text and locate the sentence that directly supports the claim. If you cannot find it, the citation cannot be used. Reading the summary is not sufficient; the summary may itself have been generated.
| Error type | How it looks | How to catch it |
|---|---|---|
| Entirely fabricated | Correctly formatted number, absent from the source | Number search in the official source |
| Real decision, wrong reference | Decision exists but the number belongs to another | Cross-check the reference components |
| Real decision, unsupported claim | Link works, holding differs | Find the supporting sentence in the full text |
| Superseded decision | Genuine but later changed | Check the decision's current status |
| Quote out of context | Sentence exists but relates to different facts | Read the facts of the case |
Choosing a system that lowers the risk
These checks are necessary in every system, but some systems reduce the risk structurally. Look for the following during evaluation.
- Every claim linked to a clickable source. Without it, stage three cannot be performed at all.
- Access to full text. A system that shows only summaries makes the claim unverifiable.
- Out-of-scope behaviour. A system that can say a question is beyond its coverage is safer than one that answers everything.
- Status indicators. Flagging decisions that were later changed.
- A published coverage list. Knowing which courts are indexed makes an out-of-scope citation immediately suspect.
How to test it: Ask about a regulation that does not exist, or about a very narrow topic. A system that states the question is out of scope is a good sign. One that produces a detailed, confident answer has a high tendency to fabricate.
A firm-level rule
Individual vigilance does not scale. In a corporate legal function these checks need to be written into a rule. A minimum set has three clauses.
- No citation from AI output enters a filing without its full text having been read.
- The person verifying a citation may differ from the person drafting, but the verification is recorded.
- An unverifiable citation is removed, however supportive it appears.
Frequently asked questions
Why does AI invent case citations?
Language models fill gaps statistically when information is missing. In law this produces citations that are formally convincing but do not exist: a correctly formatted docket number, a plausible date and a persuasive summary. The formal polish of the output says nothing about the truth of its content.
How do I catch a fabricated citation?
Apply three stages: verify the decision exists by searching its number in the official source, check the internal consistency of the reference details (chamber and subject, filing and decision years, number format), and find the sentence in the full text that directly supports the claim. The third stage is the most often skipped and where most errors hide.
Is a citation with a working link trustworthy?
No. A system can link to a real decision and assert something that decision does not say. That is the difference between the presence of a citation and its accuracy. Open the link and locate the supporting sentence in the full text before relying on it; reading the summary is not enough, because the summary may itself have been generated.
Who is accountable for a wrong citation?
The lawyer is accountable for a citation in a filing submitted to a court. AI vendor contracts generally leave responsibility with the user. That is why the system showing sources and the output being checkable matters more than the liability clause in the contract.
Which system features reduce the risk?
Every legal claim linked to a clickable source, access to the full text of decisions, the ability to state that a question is out of scope rather than generating an answer, status flags on decisions that were later changed, and a published list of indexed sources. These make verification possible; they do not remove the need for it.
What rule should a firm set for AI use?
Three clauses at minimum: no citation from AI output enters a filing without its full text having been read, the verification is recorded, and an unverifiable citation is removed however supportive it appears. These need to be written down because individual vigilance does not scale.