A bounded task that takes minutes and has, in every reported case, been the thing nobody did.
Start a conversation with the AI Adoption Concierge, already scoped to citation verification. Pick a starting point, or describe your situation directly.
Generative models produce citations that look exactly like real ones. They have the right reporter format, plausible party names, a year that fits, and a holding that supports the proposition — because supporting the proposition is what the model was asked to do. This is not a bug that will be trained away soon; it is a consequence of how the systems generate text. The check is therefore not optional and it is not complicated: someone confirms each authority exists, and confirms it says what the draft claims. What the reported decisions show, without exception, is firms that had no point in their process where that happened.
Four distinct failure modes, in rough order of how hard they are to catch.
Cases that do not exist, formatted perfectly. The easiest to catch and still the most commonly filed.
A genuine citation attached to a proposition it does not support — harder to catch, because the case checks out.
Quoted language that appears nowhere in the cited opinion, sometimes from an otherwise real case.
Real and once good law, cited without the subsequent history the model did not weight.
Quoted provisions that do not read that way, or do not exist in that section.
An unchecked citation reused across later filings, so one failure becomes several.
What a workable check involves.
These are outcomes from reported 2026 decisions, not hypotheticals.
In a June 2026 matter in the Northern District of Mississippi, a judge sanctioned four lawyers over AI-hallucinated citations — with filings from opposing sides both affected. Verification is not a plaintiff or defence problem.
Not reliably, and treating it as though it can is a distinct failure mode. Asking a model whether its citations are real invites the same generative behaviour that produced them — it will often confirm them confidently. Retrieval-based legal research tools that cite from an actual corpus are materially better positioned than open-ended chat, because there is a real document behind the reference. Even then the proposition check remains human work: a real case can be cited for something it does not hold, and no current tool reliably catches that.
Someone other than the person who generated the draft, where the stakes justify it, for the same reason firms do not have people check their own work generally. The practical arrangement in most firms is that the drafter runs the mechanical checks and a reviewing lawyer confirms the propositions on anything going to a court or a client. What matters more than the specific allocation is that it is assigned rather than assumed — the reported failures are consistently cases where everyone believed someone else had looked.
Simply, and in a way that survives the file being reviewed later. A short note on the matter recording who verified authorities and when is enough for most purposes, and it is the difference between explaining a process to a court and having nothing to point to. Some firms use a checklist attached to the filing workflow. The specific mechanism matters less than that the record exists before anyone needs it.
Establish the facts quickly and take advice from someone qualified in professional responsibility in that jurisdiction — this is precisely the situation where general guidance is not a substitute. What the reported decisions suggest is that courts have responded far more severely where the problem was denied, minimised or discovered by the court rather than disclosed. The Ninth Circuit's 2026 sanctions order turned in part on a failure to disclose that inaccuracies came from generative AI. Candour appears to matter as much as the original error.
Describe how work moves through your firm. The Institute will help you place the step where it will actually happen.