Legal

Legal research.
Cited, or silent.

Lawyers have been sanctioned for filing submissions containing cases that did not exist. That is not an argument against AI in legal research — it is an argument against a specific architecture, and the correct one makes a non-resolving citation structurally unable to reach a user.

Last reviewed 4 September 2026 · Engineering guidance, not legal advice.

WHY FABRICATED CITATIONS HAPPEN

A language model asked a legal question with no retrieval layer generates text that resembles a correct answer. Case citations have a strong, learnable format, so the model produces well-formed citations — to cases that do not exist. This is not a bug in the model. It is the model doing exactly what it was built to do, applied to a task it was never suited to.

The fix is architectural rather than a matter of better prompting or a better model. The system retrieves from an actual corpus of authority, generates only from retrieved passages, and validates before returning that every citation in the output corresponds to a document that was actually retrieved. A citation that does not resolve does not leave the system.

Built that way, the failure mode changes from a confident fabrication to an admission that nothing relevant was found — which is a safe failure a lawyer can work with, and the only acceptable one in this setting.

THE ARCHITECTURE

GROUNDED RESEARCH PATH

Question

from a lawyer
Retrieve

from your authority corpus
Rerank

precision over recall
Generate

from retrieved passages only
Validate citations

every one must resolve
Nothing found

say so, do not invent

Solid: the grounded path Dashed: the safe failure

The validation step is not a quality check. It is the control that stops a citation which does not resolve from reaching a user. It does not guarantee the case says what the summary claims, which is why every proposition is still read at source.

WHERE IT IS GENUINELY USEFUL

Your own precedent, which nobody has indexed. Most firms and in-house teams hold years of advice, memoranda and prior submissions that are effectively unsearchable. A retrieval system over that corpus answers the question lawyers actually ask most often — have we looked at this before, and what did we say?

First-pass framing on an unfamiliar question. Not a substitute for research, but a fast orientation: what the relevant authorities are, how the question is usually framed, where the disagreement sits. It saves the first hour, which is worth having.

Comparison across jurisdictions. Where a client operates in several and the question is how positions differ. Time-consuming by hand and well suited to structured extraction, provided each jurisdiction's corpus is properly separated.

What it is not is a research substitute. Every proposition a lawyer relies on gets checked at source, exactly as it would if a trainee had produced it. The system makes finding the source faster; it does not remove the obligation to read it.

QUESTIONS

QUESTIONS — 3

A citation to a case that does not exist is structurally prevented in the output path: generation happens only over retrieved passages, and a validation step rejects any output containing a citation that does not resolve to a retrieved document. That is a structural control rather than a probabilistic improvement. It does not stop a real case being cited for a proposition it does not support, which is the more common failure now and why the span is shown for the lawyer to read at source.

No. It needs a corpus, and for published authority that generally means a licensed source, indexed subject to the licence terms of that source. What it adds is your internal material, which no external provider has, and a retrieval interface over both.

The corpus carries currency metadata and the system surfaces it, but the treatment status of a case is a question for a proper citator. We would rather integrate one than approximate it, because being subtly wrong about whether a case is still good law is worse than not answering.

Keep reading

30-minute technical call · no deck