Everybody in the profession knows the Avianca story by now, and most of us took the wrong lesson from it.
On June 22, 2023, Judge P. Kevin Castel of the Southern District of New York sanctioned two lawyers and their firm $5,000 after a brief cited six federal decisions that didn’t exist. ChatGPT made them up. Opposing counsel couldn’t find them. Neither could the court. Judge Castel went through the fake Varghese opinion and called parts of its reasoning gibberish. The lawyers had to send corrective letters to every judge whose name had been stuck on a fake opinion.
The lesson everybody took: don’t let AI write your citations.
Here’s the one I’d take instead: a fake case is the least dangerous citation failure there is, because it’s the only one that announces itself. A case that doesn’t exist gets caught the first time anybody looks it up. It’s embarrassing and it’s sanctionable, but the dumbest check in the world will find it.
The one that should keep you up at night is the one that sails through every check you’re running right now.
The bad cite that looks perfect
Picture a citation in a brief. The case is real. It’s in the reporter where the brief says it is. The pin cite lands on the right page. The parallel cite is right. The signal fits. The italics are right. A cite checker, human or software, would wave it through without blinking, because everything you can check about its form is correct.
And the proposition it’s cited for isn’t in the opinion.
Nothing about the shape of that cite is wrong. That’s exactly the problem. Every tool most firms use checks shape.
Here’s how it really breaks down, roughly from easiest to hardest to catch:
| Failure | Caught by |
|---|---|
| Case does not exist | Any lookup. Trivial. |
| Wrong reporter, volume, or page | Any lookup. Trivial. |
| Quote is not verbatim — a word changed, a tense shifted | Reading the source with the brief in hand |
| Quote is verbatim but attributed to the wrong case | Reading the source |
| Real case, cited for a proposition it does not contain | Reading the opinion and thinking |
| Dicta presented as holding | Reading the opinion and thinking |
| Holding accurate but the case was reversed, superseded, or the statute amended | Checking subsequent history — and knowing to |
| Ellipsis or bracket that quietly reverses the meaning | Comparing the quote to the surrounding paragraph |
The top two are what most people mean when they say they “checked the cites.” The bottom six are where briefs actually go wrong, and only one thing catches any of them: somebody goes and reads the source.
Why AI makes this worse
“AI makes things up” is true, and it’s too blunt to build a defense around. Look at how it actually goes wrong.
A language model writes text that sounds right in context. A citation is about as rigid and patterned as anything in legal writing, which makes it exactly the kind of thing a model spits out smoothly. The model has seen a million citations. It knows what one looks like. It’ll give you one that looks perfect.
It’s also seen a million sentences shaped like “X v. Y held that [proposition].” Writing a new one, with a real case and a believable proposition, is the same move. The model isn’t lying. It has no idea there’s a difference. It’s turning out well-formed text, which is its whole job.
So the danger in AI-assisted legal writing isn’t sloppiness. It’s fluency. The output is cleaner than a rushed associate’s and wrong in ways the polish hides. A cite that’s 90% right is far more dangerous than one that’s obvious garbage, because somebody checks the obvious garbage.
And it works against how you naturally work. When the draft looks good, you check it less.
What Rule 11 actually says
This isn’t just about quality. Here’s Minnesota Rule of Civil Procedure 11.02:
“By presenting to the court (whether by signing, filing, submitting, or later advocating) a pleading, written motion, or other document, an attorney or self-represented litigant is certifying that to the best of the person’s knowledge, information, and belief, formed after an inquiry reasonable under the circumstances” …
and, at subsection (b), that
“the claims, defenses, and other legal contentions therein are warranted by existing law or by a nonfrivolous argument for the extension, modification, or reversal of existing law or the establishment of new law.”
Rule 11.03 is the stick: if the court determines Rule 11.02 has been violated, “the court may … impose an appropriate sanction upon the attorneys, law firms, or parties that have violated Rule 11.02 or are responsible for the violation.”
Read that certification slowly. It’s on the person presenting, and it’s measured by how reasonable that person’s inquiry was. Not the tool’s. Not the associate’s. Yours. There’s no reading of this rule where “the software found the case” counts as your inquiry.
That’s the reason for everything below. If you’re going to certify an inquiry, you’d better have a defined one, and it ought to leave a paper trail.
The four-pass gate
In our practice, nothing with a citation in it goes to a court, to opposing counsel, or to a client as final until it clears four passes. Each pass asks one question, and each pass reads the source. None of them gets to lean on memory, on the drafter’s summary, or on the pass before it.
Pass 1 — Existence. Does this authority exist, at this cite, with these parties, in this court, on this date? Pull it. Not “confirm it sounds familiar” — pull the actual text. If you can’t pull it, it isn’t a citation. It’s a guess.
Pass 2 — Quotation. Does every quoted string match the source character for character? This one’s mechanical, so do it mechanically: a string comparison catches the changed word, the quietly cleaned-up punctuation, the dropped “not.” Watch the ellipses and brackets. An ellipsis that drops a qualifier can flip a holding and still be, technically, an accurate quote.
Pass 3 — Proposition. This is the pass that matters, and it’s the one nobody runs. Does the source really stand for what the brief says it does? To answer that you read the discussion around it, not the headnote. Headnotes aren’t the opinion, and they’re often where a misattribution starts. Separate holding from dicta. Make sure the procedural posture supports how you’re using the case: an opinion reversing a grant of summary judgment doesn’t establish what the facts were. Make sure the jurisdiction actually governs the point.
Pass 4 — Currency. Is it still good law? Reversed, vacated, superseded by statute, cut back by later cases, or (the sneaky one) a rule that’s been renumbered or abrogated since the last time you used it. This pass goes stale in a way the others don’t. A cite you verified two years ago isn’t verified today.
What comes out is a manifest: every authority in the document, what got checked, what got pulled, and what happened to it. Verified, corrected, or removed.
What makes it actually work
Writing the four passes down doesn’t give you a gate. Four rules do.
1. The checker isn’t the drafter. Whoever built the argument is the worst person alive to check whether the authority supports it, because they already believe it does. In our chain the jobs are split for real: the drafting specialist hands off to a citation specialist whose whole job is to disbelieve the draft. If you’re a solo, the stand-in is a separate session with a fresh context that hasn’t seen your argument. Give it the document and the sources, not your reasoning.
2. It’s a step, not a habit. Habits fail on the day you’re in a hurry, which is the day the brief is due. Make verification a required stop between drafting and formatting, so skipping it means you notice you skipped it.
3. No citing from memory. Ever. Our drafting instruction says it flat out: the drafter may not produce a citation from recollection. Either the authority was pulled and read in this session, or it goes in as a placeholder marked for research. That wipes out the whole fake-case category before verification even starts, because there’s nothing to make one up from.
4. If you can’t verify it, it comes out. Not flagged. Not highlighted in yellow for later. Out. Flags get lost between the draft and the filing. An argument that’s lost its support should look like it’s missing its support, so somebody has to deal with it.
What it costs you
It’s slower. There’s no version of this that isn’t slower, and anybody selling you a cite checker that’s instant and also does Pass 3 is selling you Pass 1 with a nicer screen.
But the math isn’t close. A brief with thirty citations takes maybe ninety minutes to run through all four passes, less if you enforce the drafting rule above, because most of Pass 1 got done while drafting. On the other side of the ledger: cite a case for something it doesn’t say and you get a court that stops trusting what you tell it, an opponent who opens the reply with your mistake, a client whose motion rides on an argument that’s quietly lost its footing, and, on a bad day, a Rule 11 proceeding.
Ninety minutes against a loss with no ceiling. It’s cheap insurance, and it’s the single best process control I know of for a practice that uses AI at all.
And there’s a bonus I didn’t see coming. Pass 3 makes the brief better. Make somebody spell out, for every authority, exactly what it stands for and why that gets you to this proposition, and you find the arguments that were weaker than they looked. More than once the verification pass has kicked a section back for a rewrite, not because a cite was wrong but because the honest answer to “what does this case actually give us” just turned out to be “less than we implied.” Better to hear that from your own reviewer than from the bench.
Sources
- Mata v. Avianca, Inc., 678 F. Supp. 3d 443 (S.D.N.Y. 2023) (sanctions order entered June 22, 2023)
- Minn. R. Civ. P. 11.02 and 11.03
Rules quoted above were verified against the Revisor of Statutes rather than cited from memory. General commentary on practice management; not legal advice, not ethics advice. No client matter is described in this article.