Why AI hallucinates legal authorities — and the tell-tale signs
2026-10-13
AI models generate plausible text, not verified facts. Unless a tool is grounded in a legal database, it predicts what a citation should look like — which is why a fabricated case reads exactly like a real one. The tells are: a citation with no pin cite, a case name that is slightly off, a quote that cannot be found, and confident language ('it is well established') with nothing behind it. The fix is not a better model but a fixed rule: never accept an authority you have not opened.
It predicts, it does not look up
A general model completes text. Ask for a case and it will produce something shaped like a case — a plausible name, a plausible year, a plausible citation. It is not lying; it has no notion of true and false, only of likely and unlikely.
The four tells
- No pin cite, or a pin cite that lands on nothing.
- A case name that is close but not exact.
- A quote you cannot find in the judgment.
- Confident framing with no source: 'the leading authority confirms'.
The rule that ends it
Label every claim authority or reasoning, and open every authority before it goes in a document. A model that is told to label, and a lawyer who checks, beats any model that is told to be careful.
If you cannot open it, it is not an authority — it is a guess.