DocketX

DocketX / Why AI invents case citations, and what stops it

Why AI invents case citations, and what stops it

A fabricated citation is the single most damaging thing a language model can put in a legal document, and it is also the most preventable. The mechanism is not mysterious.

Why a model invents a citation

A language model produces text that is statistically likely, not text it has looked up. A reporter citation is a short, highly patterned string: a volume, a reporter abbreviation, a page. That pattern is trivially easy to produce and impossible to tell from a real one by looking at it. The model is not lying; it has no mechanism for knowing whether the case exists, because nothing in its answer involved checking.

Why prompting does not fix it

Instructions like "do not hallucinate" or "only cite real cases" change the model tone and its willingness to hedge. They do not give it a lookup. In adversarial testing we asked models for eight cases at once with pinpoint pages, invented doctrines by name, and told the system that citation checking had been disabled by the client. Models comply with that framing readily. Confidence is not correlated with correctness.

What actually catches it

The honest limit

Verification proves a citation exists in the library and that the name matches. It does not prove the case still states good law, and it does not prove the case supports the proposition it is cited for. Those are separate questions, and a tool that blurs them is selling false comfort. Read the opinion.

See how the checking works →

Questions

Has this actually happened to lawyers?

Yes. Courts in several jurisdictions have sanctioned lawyers for filings containing citations that did not exist, after the citations were produced by a language model and filed without being checked.

Does using a better model fix it?

It reduces the rate; it does not eliminate it. In our testing every model produced fabricated citations on adversarial prompts when it was not grounded in a real library and checked afterward.

What is the difference between grounding and verification?

Grounding means the model is given real law to read before it answers, which makes fabrication less likely. Verification means every citation in the finished answer is looked up, which catches what grounding missed. You want both.

Related

Keep reading

DocketX products are research and explanation tools, not legal advice and not a law firm. They show you the real rule or opinion behind a statement so you can read it yourself. If you can get a lawyer or free legal aid, do.