A Clark County judge sanctioned a Las Vegas defense attorney this week after finding that artificial intelligence had been “misused” in a brief filed on behalf of a man accused of bombing a restaurant.
The penalty was $500, payable to the county library. That number is the story, because it is far less than the lawyer saved by not writing the brief himself.
What Happened in Department 11
District Judge Monica Trujillo struck the filing and ordered the sanction after prosecutors flagged something specific: the reply brief cited Evans v. State, 123 Nev. 117, 159 P.3d 438 (2007). That case does not exist. The citation is formatted correctly, the reporter volumes are plausible, the year fits the numbering. It is simply not real.
Trujillo’s reasoning, as reported by the Las Vegas Review-Journal, did not rest on the fake case alone. She noted the reply was substantially longer than the attorney’s prior filing, written in a different style, with differently formatted citations and an incorrect reference to a previous case. In other words, the tell was not one hallucination. It was a document that did not read like the person who signed it.
Attorney Dean Kajioka initially denied using AI and attributed the citation to a typo when prosecutors first sought sanctions, later acknowledging that AI had “apparently” assisted. His client, Robert Schwieger, 53, faces arson and explosives counts in the bombing of Piero’s Italian Cuisine.
The Economics Point the Wrong Way
Here is the uncomfortable part. A $500 sanction and a struck brief is a bad afternoon. Drafting a substantive reply brief from scratch is somewhere between six and twenty hours of attorney time. At Nevada criminal-defense rates, the fine is roughly one to two billable hours.
Run that as an expected-value problem the way any rational actor would. The probability of detection is well under one, because catching a fabricated citation requires opposing counsel to pull every case in a brief, which is itself hours of unbilled work. The penalty when caught is a few hundred dollars and a filing you have to redo. The savings when not caught are most of a day.
That math does not deter anyone. It prices the behavior.
Compare it to what the profession does about other filing failures. Missing a statute of limitations ends careers. Commingling client funds ends careers. Filing a document with invented law, which is a false statement to a tribunal, currently costs about what a deposition transcript does.
Some courts have started moving. Penalties have climbed into five figures at the federal appellate level, and there have been bar suspensions tied to AI filings. But the modal outcome is still a small fine and an admonishment, and a modal outcome is what practitioners actually respond to.
This Is Not a Handful of Cases Anymore
The scale here is the thing most coverage of individual sanctions misses. Damien Charlotin, a research fellow at HEC Paris, maintains a public database of court decisions involving AI-hallucinated material, and by early July it had logged more than 1,600 such decisions worldwide, with over a thousand in the United States alone. The inclusion bar is strict: a court has to have found, or clearly implied, that a party relied on fabricated material. These are not suspicions. They are findings.
The rate is what should worry the courts. Eugene Volokh flagged that on a single day in March, 17 separate U.S. court decisions noted suspected AI hallucinations in filings. Seventeen, in one day, and that counts only the ones a judge wrote down.
And the responsible party is frequently a licensed attorney rather than someone representing themselves. The early framing of this problem, that it was mostly pro se litigants who did not know better, has not survived contact with the data.
Why Detection Does Not Scale
The Seventh Circuit ran into the structural problem directly in Dec v. Mullin this spring. An attorney cited two nonexistent cases and a false quotation in an appellate brief, denied using AI at oral argument, then filed a letter admitting the citations had been copied from another brief he could no longer locate. The court admonished him but declined further sanctions, reasoning the errors were unintentional, counsel was contrite, and the fabricated cases supported a legal standard nobody disputed anyway.
That last clause is the crux. The hallucinations did not change the outcome, so the court let it go. Which is defensible case by case and disastrous as a policy, because it means the sanction depends on whether the fake law happened to matter rather than on whether the lawyer verified anything.
The deeper issue is that verification is asymmetric. Generating a plausible citation costs a second. Confirming one does not exist costs a search, and confirming an entire brief costs an afternoon. Every incentive in litigation pushes the checking burden onto the party least able to bear it, which is usually the underfunded side, and in criminal practice that is frequently the defense.
Judges have become the backstop by default. Trujillo caught this one by reading closely enough to notice that a brief did not sound like the man who signed it, which is careful judging and is not a system. It is one judge, on one filing, in one department.
What Actually Changes This
Not fines at this level. A penalty has to exceed the value of the shortcut, and $500 does not come close.
Three things would. Mandatory certification that every citation in a filing was independently verified, with the certification itself sanctionable, which converts a fuzzy duty into a specific false statement. Automatic referral to the state bar rather than a discretionary one, so the cost is licensure risk instead of library money. And courts running citation-checking software on filings, which is a solvable engineering problem that several vendors have already solved.
None of that requires anyone to take a position on whether lawyers should use AI. They are going to, and for research synthesis and first drafts they probably should. The question is narrower: who bears the cost when the output is wrong. Right now it is the tribunal, the opposing party, and occasionally a defendant sitting in custody while his lawyer’s brief gets struck.
For what it is worth, the general-purpose models most of these filings come from are not marketed as legal research tools, and their reliability varies considerably depending on which one you reach for. That distinction matters less than it should, because a citation that looks right looks right regardless of which model produced it.
Schwieger’s case continues. His attorney is out $500 and one brief. The next filing in Clark County that contains a case nobody can find will probably cost about the same, and there is no particular reason to think there will not be one.