How to cite case law with AI without fake cases

Getting something wrong and making something up are different things. A language model that miscites a decision usually isn't confused. It is generating the most probable text, and a case citation has a very predictable shape.

Why a model produces cases that don't exist

A case number is a pattern: some digits, a year, a court, a chamber. A model trained on large volumes of legal text learns that pattern very precisely. What it doesn't learn is the mapping between a specific identifier and a specific case.

When you ask for a decision that supports a position, the system doesn't search a repository. It produces a sequence that looks like the ones it has seen: a well-formed number, a date that fits the period, a plausible judge, a holding aligned with what you asked for. Everything fits. None of it means the case exists.

Lawyers in several countries have filed AI-generated citations that didn't match any real decision. The pattern repeats: nobody checked because nothing seemed to need checking.

What makes a case citation verifiable

A citation you can check has four elements, and all four must hold up against a source:

1. The court and chamber

   Which body issued the decision. It sets the weight of the precedent and it is the first thing to confirm.

2. The identifier and date

   Case or docket number, with its date. This is what lets you find the decision in the official repository.

3. The literal passage

   The text that supports the claim, not a paraphrase. A paraphrase can be faithful or it can shift the meaning, and from the outside you can't tell which.

4. The holding, separated from dicta

   Which part of the decision decides the case and which part is commentary. Citing dicta as if it were the holding is a substantive error, even when the citation is authentic.

The first two confirm that the decision exists. The last two confirm that it says what you claim. These are different problems and you check them separately.

How to check before you cite

Don't ask the system whether its citation is real. Asking the model that produced it generates another prediction, with the same confidence and the same lack of grounding.

Instead:

1. Find the decision at the source. The court's repository, the case file, your research database. If it isn't there, drop the citation. Don't "adjust" it.
2. Read the passage in context. A sentence on its own can reverse its meaning two paragraphs later.
3. Confirm it is still good law. A real decision may have been overruled. Existing and applying are different things.
4. Ask for the opposite. Which decisions support the opposing position? A system that only confirms what you want to hear isn't researching.

The fourth step is the one most people skip. It is also the fastest way to see whether the system is reading something or just agreeing with you.

How the problem changes when the system reads your documents

It helps to change the question. Instead of "how do I stop the AI from inventing?", ask where the answer is coming from.

A system that answers from training memory can't give you a verifiable source, because there is no source, only a probability distribution. A system that answers only from documents you uploaded, such as the case file, the decisions you already have or the contract in dispute, can point to the exact document and page, because the passage existed before you asked.

Verification stops being a search and becomes a check: open the page cited and confirm the text is there.

That doesn't remove legal judgment. Nothing does. It moves the work from "confirm this exists" to "assess whether this applies", which is where judgment actually adds value.