The useful thing about this line of cases is that it is not speculative. Judges have written down what they did and why. Read in order, the orders show a penalty that has gone up every year, and reasoning that has moved from technology to responsibility.
2023: a penalty and three letters
Mata v. Avianca is where the record starts. Judge P. Kevin Castel found that two attorneys and their firm submitted non-existent judicial opinions with fake quotes and citations produced by ChatGPT, then stood by those opinions after the court questioned them [1]. He was careful about what the case was not about:
Technological advances are commonplace and there is nothing inherently improper about using a reliable artificial intelligence tool for assistance. But existing rules impose a gatekeeping role on attorneys to ensure the accuracy of their filings. [1]
The sanction: a $5,000 penalty, jointly and severally, paid into the registry of the court — plus letters to the client and to each judge who had been named as the author of a fabricated opinion [1]. That second part is the one lawyers remember.
Six months later the Chief Justice devoted part of his year-end report to the same problem, noting that AI had made headlines for lawyers using the application to submit briefs with citations to non-existent cases, and adding, in a parenthesis, “(Always a bad idea.)” [7]
2024: the appellate courts treat it as conduct
In Park v. Kim, the Second Circuit found counsel had cited a decision that did not exist and had not read or otherwise confirmed its validity. The panel referred the attorney to its Grievance Panel and ordered her to give a copy of the decision to her client [2]. No fine. The consequence was a licensing matter and a conversation with the client.
2025: fines stop being the point
Three orders from 2025 show the shift.
In Wadsworth v. Walmart, motions in limine cited nine cases; eight did not exist. The court revoked the drafting attorney’s pro hac vice admission, removed him as counsel of record, and fined him $3,000, with $1,000 each against the two attorneys who signed [3].
In Lacey v. State Farm, a special master struck the offending briefs, denied the discovery relief they sought, and ordered the plaintiff’s two law firms to pay $31,100 to the defense, jointly and severally [4]. The money followed the wasted work.
Johnson v. Dunn went furthest. Judge Anna M. Manasco publicly reprimanded three attorneys, disqualified them from the case, directed the clerk to publish the order in the Federal Supplement, ordered each lawyer to serve the order on their clients, opposing counsel and presiding judges in every pending matter, and referred them to the Alabama State Bar [5]. Her reason for going past a fine is the sentence worth copying into an office policy:
If fines and public embarrassment were effective deterrents, there would not be so many cases to cite. [5]
How many cases there are to cite
Damien Charlotin maintains a public database of decisions in which a court addressed AI-generated hallucinated content. As of 14 September 2026 it lists 2,041 decisions worldwide, 1,396 of them in the United States [6]. Its own breakdown attributes 812 to lawyers and 1,174 to self-represented litigants, and it records 32 in which the responsible party was a judge [6]. Fabricated material is tagged in 1,690 of the decisions, misrepresented authority in 854, and false quotes in 549 [6].
Two things follow. First, courts are not treating this as a novelty any more; a tracker with four figures in it is a body of law, not a curiosity. Second, the defect that is easiest to describe — a case that does not exist — is not the only one being sanctioned. Roughly fourteen hundred of those tags are for authority that exists and was described wrongly.
What the orders actually ask of you
Nothing in these decisions turns on whether AI was used. Every one of them turns on whether the filing was verified before it was signed. Mata framed it as a gatekeeping role [1]. Park framed it as the basic obligations of counsel [2]. Johnson framed it as accountability proportionate to the harm [5].
So the practical question for an office is not “do we allow this software.” It is: what would have to be true of a tool before a filing it touched could be signed without a second, manual pass? That is a question about architecture, not about policy, and it has a testable answer — a system that can only render what it has already checked against the source, and that says so on the page when it cannot.
Apodicta is built to that standard, and it is the standard the orders above describe: render nothing you can’t ground, and name what you hold.