The five minutes that matter are the last five before the brief goes out. Here is what to do with them when part of the draft came out of a model, or out of a research tool built on one.
Do the steps in order. Each one answers a different question, and the second question is the one that costs people their evening.
Step 1. Make the citation resolve to a document
Take the citation as a string and try to turn it into one specific document: a case at that reporter and page, a statute at that section, an exhibit at that Bates number.
For reported cases, the Free Law Project publishes a free citation lookup that takes a citation and returns the opinion it resolves to, if any [5]. Your own research service does the same thing. This is an existence check, and it is mechanical: the string resolves or it does not. Budget fifteen seconds per cite.
If nothing resolves, do not go looking for a version of the name that does. Delete the sentence and the citation together, and write the proposition again from an authority you pulled yourself.
One thing that is not a check: asking the model. In Mata v. Avianca the lawyer questioned ChatGPT about the reliability of its own work, asking whether the lead case was “a real case” and whether the others were fake, and it answered that they were real and could be found on Westlaw, LexisNexis and in the Federal Reporter [1]. The court took that exchange as evidence of how he came to believe it, not as verification.
Step 2. Make the passage support the sentence
Now open the document at the pincite and read the passage the draft is relying on. Ask one question: does this passage hold the proposition the sentence hangs on it, in this jurisdiction, as of this date?
This is a support check, and no lookup performs it. It is also where the interesting failures live. The Stanford RegLab evaluation of the leading AI legal research tools separated responses that state the law wrongly from responses that state it correctly while citing something that does not support it, which the researchers called misgrounded, and counted both as hallucinations; the tools they tested hallucinated between 17 and 33 percent of the time [3].
A misgrounded citation survives the check a busy reader actually performs, which is to click the cite, see a real case, and move on. That is exactly why you have to read the passage rather than the caption.
Budget two to three minutes for the one or two citations that are carrying the argument, and do them first. The string cite that follows “see also” can wait; the case your motion stands on cannot.
Step 3. Check the quotation word for word
If the sentence quotes, compare the quoted words against the opinion on screen. Not the sense of them, the words.
This step exists because of what special masters and judges keep finding. In Lacey v. State Farm, a review of a ten-page brief found that roughly nine of its 27 citations were incorrect in some way, at least two of the cited authorities did not exist at all, and several quotations attributed to real opinions were, in the order’s words, phony and did not accurately represent those materials [2]. An existence check clears most of that brief. The quotations do not clear.
When the case is real but says something else
This is the common outcome, and it is not a disaster if you catch it. You have three honest moves:
- Narrow the sentence to what the passage actually holds, and keep the cite.
- Keep the sentence and go find the authority that holds it, on your own.
- Cut it. If neither of the first two works in the time you have, the claim was not ready. A brief that says less and is right beats a brief that says more and gets an order to show cause.
What you may not do is leave the sentence standing with a cite that does not carry it because the deadline is tonight.
Why this is your job and not the tool’s
ABA Formal Opinion 512 puts it under Rule 1.1: a lawyer’s reliance on, or submission of, a tool’s output without an appropriate degree of independent verification or review could violate the duty of competence [4]. The opinion also says the amount of verification depends on the tool and the task, which is the part worth taking seriously [4]. A tool that only proves a case exists has done fifteen seconds of the work and left you the three minutes that matter.
So keep the procedure, and make it boring: resolve, read the passage, compare the quote, then decide. On a normal motion it costs five minutes. On the motion where something is wrong, it is the five minutes you will be glad to have spent.
The alternative is to make the check part of the machine rather than part of your evening. That is the design Apodicta argues for: the existence question and the support question both run outside the model, before anything renders, and what fails a check is held and named instead of printed.