Verification as a Craft: Existence Is Not Enough

Existence is the easy question

Whether a case exists is the easiest question a citation has to answer. Type the volume, reporter, and page into a research database, and in a few seconds you know whether a court decided anything there. It is also the question machines now answer best. In a 2026 study, Princeton researchers planted errors in excerpts from real federal appellate briefs and set AI agents loose on them. The best agent caught every citation to a case that did not exist (opens in a new tab), and it did worst on wrong page numbers.

Yet the sanctions cases are full of failures that come after existence. In Noland v. Land of the Free, L.P., 336 Cal. Rptr. 3d 897 (Ct. App. 2025) (opens in a new tab), a California appeal, most of the cases the lawyer cited were real, but 21 of the 23 quotations in the opening brief were fabricated. So what does verifying a citation actually take, and which parts of it can you hand to a machine? This reading takes the five questions a citation has to answer in the order a careful checker asks them. The activity then has you grade three machines on those questions and answer them yourself, for five citations from a real brief.

Does the case exist?

A case citation packs several facts into one line: the case name, the volume, the reporter, the first page, and the court and year. Most add a pinpoint, the page where the point actually appears. Existence means all of those facts describe one real decision. A database that returns something for the volume and page has not answered the question yet, because the something may be a different case.

That is the first trap. In Russell v. Mells, 426 So. 3d 913 (Fla. 2d DCA 2025) (opens in a new tab), a lawyer’s answer brief quoted a case it called Cade v. Roberts, 403 So. 2d 516 (Fla. 5th DCA 1981). The Second District found no reported case by that name anywhere in Florida. Page 516 of that volume of the Southern Reporter belongs to Girardeau v. State, a First District opinion about a criminal contempt proceeding against a state legislator who refused to answer a grand jury’s questions. The citation worked in the sense that it led somewhere. It led to the wrong place. The court referred the lawyer to The Florida Bar, and it said that not meaning to mislead the court was no excuse.

The second trap runs the other way. A real case can be hard to find, because every database has gaps: an unpublished order, a recent decision, a case known mostly by its Westlaw number. A search that comes back empty makes the citation unverified, not fake. In the Princeton team’s paper, the weaker AI models often flagged real cases as fake because a free database lacked them (opens in a new tab). An automated citation check can make that mistake in either direction, confirming a citation that leads to the wrong case or condemning a real one it failed to find. So a careful checker searches three ways before deciding: by the citation, by the case name, and by any words the brief quotes from it.

Did the court decide that, on that page?

Once you have the right case, the questions stop being lookups and start being reading. Holding accuracy asks whether the court decided what the brief says it decided. In Noland, many of the real cases did not support the propositions they were cited for, and some discussed other matters entirely. The brief said one case explained why the Legislature wrote a summary-judgment statute. The court found that the case never mentions summary judgment or the statute at all.

The check starts near the end of the opinion. The disposition, the line that says affirmed, reversed, or dismissed, tells you who won and what the court did about it. The reasoning before it tells you why. Read both, because a court can repeat every word of an argument on its way to rejecting it, and a brief can quote the repetition as if it were the holding.

Then comes the page. Pinpoint support asks whether the page the citation names holds the words or reasoning behind the sentence. A pinpoint tells the judge exactly where to look. A wrong one sends the judge to a page that says nothing about the point, or says the opposite. Google Scholar and Westlaw both mark the reporter’s page numbers inside the text of most opinions, so you can land on the cited page and read what is actually there.

Page numbers were the hardest check for the Princeton agents, and the researchers traced the failure to access rather than ability. The official page numbers come from the commercial publishers’ reporters, and free databases often leave them out. A machine that cannot see the page cannot tell you what is on it.

Are those the court’s words?

Quotation fidelity asks the plainest question of the five: does every word inside the quotation marks appear in the cited opinion, in that order? It is also the easiest to check. Open the opinion, search it for the exact words, and read what surrounds them.

Fake quotations come in two kinds. Some are invented outright, like most of the quotations in Noland. The others are real words in the wrong case. The two cases that did exist in the Russell answer brief were both misquoted, and the Second District explained how: the brief “attributed quoted text from one case that was actually found in another.” Every word existed somewhere in Florida law. They were in the wrong opinion, so the citation vouched for something its case never said.

That second kind survives a quick check, because the words sound like the law, and in a sense they are. A phrase common in the case law may appear in dozens of opinions. The question is whether it appears in this one, at this page. The Sixth Circuit sorted the problems in an appellant’s briefs the same way in its sanctions opinion in Whiting v. City of Athens, Nos. 24-5918, 24-5919, 25-5424 (6th Cir. Mar. 13, 2026) (opens in a new tab). Some cited cases did not exist, some did not contain the language quoted from them, and some did not discuss the proposition at all.

Is it still good law?

A case can pass all four checks so far and still fail the last one. Subsequent history asks what happened to the case after it was decided: whether an appellate court reversed it, a higher court overruled it, or a statute or rule displaced it.

For forty years, Chevron U.S.A. Inc. v. Natural Resources Defense Council, Inc., 467 U.S. 837 (1984), told federal courts to defer to an agency’s reasonable reading of an ambiguous statute. In June 2024, the Supreme Court overruled it in Loper Bright Enterprises v. Raimondo, 603 U.S. 369 (2024) (opens in a new tab). A brief that quotes Chevron’s test today to demand that deference cites a real case, quotes it exactly, and gives the right page, and it still states law that no longer governs. Even here the answer depends on what the case is cited for. The Court said it was not calling into question the specific holdings of earlier cases decided under Chevron, which remain subject to statutory stare decisis.

A citator does this work for you. KeyCite in Westlaw puts a flag beside a case that a later court has reversed, overruled, or questioned, and lists the decisions that did it. Without a citator, the fallback is the list of later opinions that cite the case, read newest first. Both can lag behind a decision from last week, so a case decided recently deserves a fresh search of its own name.

The price of the check

Each of these checks takes minutes, and the minutes add up. A brief with forty citations and five questions each holds two hundred answers. That arithmetic is verification cost, and it is the honest reply to a supervising lawyer who wants a cite-check done in the ten minutes before a deadline.

Skipping the check does not make the cost go away. It moves it to someone else. In Whiting, the Sixth Circuit said the fake citations and misstatements forced the court to verify every citation in the briefs itself. It ordered the two lawyers to pay the other side’s attorney’s fees and double costs on appeal, plus $15,000 each to the court. Noland shows the other side has a stake too. The court declined to order the sanctions paid to opposing counsel, because the opposing side never alerted the court to the fake citations and seemed to learn of them only from the court’s own order.

The check does not end at filing, either. In the brief you are about to read, the defendant’s reply pointed out the plaintiffs’ citations to nonexistent cases, and the court later faulted the plaintiffs for never asking to correct them. A correction is a citation too, and it needs the same five questions. In Akerlund v. Atlas Air, Inc., No. 24-11033 (11th Cir. July 10, 2026) (opens in a new tab), an appeal from the Southern District of Florida, counsel’s opening brief relied on at least eight nonexistent cases. After the other side flagged the problem, counsel asked to withdraw eight listed cases as “erroneous or unverifiable.” None matched the fake cases in the opening brief, and all eight were fake as well.

The activity puts you on the checking side of that brief. You take five of its citations and run them past three machines, and you score each report on the five questions before you verify anything yourself. Then you answer the questions by hand with a clock running, one cited page at a time. Some of the five are sound. The scores you give the machines after that will tell you whether the report you trusted first deserved it.