The pattern is always the same, and once you know it you see it everywhere. A lawyer, under deadline pressure, asks a large language model to help draft a brief. The model returns clean, confident prose studded with case citations that look exactly like real law. Correct format, plausible names, a court, a year, a pin cite. There is only one problem. Some of those cases were never decided, some of those quotes were never written, and the model produced them with the same fluent certainty it uses for facts that happen to be true. Nobody checks every citation against a real database. The brief goes out. And then opposing counsel, or worse, the judge, tries to pull the case and finds nothing there.
What has changed in 2026 is not the mistake. It is the response. The courts have run out of patience for the excuse that the technology is new, and they have started treating a fabricated citation as exactly what it is: a false statement made to a court, filed under a lawyer's signature, which the lawyer is professionally obligated to have verified. The consequences have escalated accordingly, and the numbers are no longer symbolic.
When Both Sides Get Caught At Once
In June a federal judge did something that would have been unthinkable a couple of years ago. Rather than singling out one careless attorney, the judge punished four lawyers after discovering that both sides of a lawsuit had leaned on AI to prepare their filings. Read that again. This was not one desperate solo practitioner cutting a corner. It was opposing teams, adversaries who are supposed to check each other's work as a matter of professional survival, both submitting material shaped by a tool that invents authority. When both sides of a case are compromised by the same failure mode, the adversarial system that is supposed to catch errors stops functioning as a safety net. The judge had to become the fact checker of last resort, and the lawyers paid for putting the court in that position.
The Bill Hits Six Figures
If the four-lawyer case shows how wide the problem is, an Oregon dispute shows how expensive it has become. In a lawsuit involving a vineyard, AI hallucinated case citations helped run up sanctions in the neighborhood of one hundred and ten thousand dollars. That is not a slap on the wrist or a mandatory ethics seminar. That is a number large enough to end a small practice, and it lands specifically because the fabricated authority was not a harmless typo. It was presented to a court as real law, it wasted the time of the judge and the opposing party, and it undermined the integrity of the proceeding. Courts measure that harm in money, and the meter is running higher every time this happens.
The vineyard case matters because it puts a concrete price on a risk that many lawyers still treat as theoretical. The pitch for using these tools is efficiency, the promise that a machine can draft in minutes what used to take hours. But a sanction near six figures erases an enormous amount of that saved time in a single order, and it does so while attaching the lawyer's name to a public finding of misconduct. The math on AI-assisted drafting looks very different once you weigh the downside honestly.
Even A Warning Now Carries Weight
Not every case ends in a five-figure penalty, but the warnings themselves have sharpened. In June a judge issued a formal warning after a lawyer representing student protesters cited AI-hallucinated cases in a lawsuit against Columbia and Barnard. On its face a warning sounds mild. In context it is a marker. Judges do not put these admonitions in the record casually, and once a court has flagged a lawyer for citing fictional authority, the tolerance for a repeat performance drops to zero. The warning is the courtesy that comes before the sanction, and the fact that these warnings are now routine tells you how common the underlying failure has become.
The Columbia and Barnard matter also illustrates who gets hurt when this goes wrong. The lawyer was representing student protesters, people relying on counsel to make their strongest possible case. When a brief is polluted with invented citations, it is not only the attorney who takes the reputational hit. The client's actual arguments get dragged down with the fabrications, because a court that has caught you inventing one case starts to distrust everything else you filed. Hallucinated authority does not just embarrass the lawyer. It can quietly sink the client.
The Liability Is Spreading Past The Lawyers
Here is the development that should worry the companies building these systems, not just the professionals using them. In June a German court held Google liable for an AI hallucination. For years the industry position has been that the output is the user's responsibility, that a chatbot is a tool and any false statement it produces is the fault of whoever relied on it. A court assigning liability to the company behind the model punctures that defense. It suggests the machine that manufactures confident falsehoods may itself be a source of legal exposure, not merely a neutral instrument that innocent users misuse.
Sanctioning a lawyer treats hallucination as human error. Holding the company liable treats it as a product defect. Those are two very different theories of blame, and in 2026 the courts started testing both at the same time.
This is the throughline that connects the courtroom stories to everything else documented here. The reason these tools fabricate cases is not a bug that a patch will fix. Large language models generate text by predicting what is statistically likely to come next, not by retrieving verified facts from a checked record. A citation that looks right and a citation that is right are, to the model, nearly indistinguishable, because it is optimizing for plausibility, not truth. That is why the fabrication is always so convincing and always delivered without hesitation. The confidence is not a signal that the model knows something. It is the default setting, applied identically to fact and fiction. We break that mechanism down in detail in our explainer on why AI hallucinations happen and in the growing documentation of AI hallucination lawsuits reaching the courts.
Why This Keeps Happening To Smart People
It is tempting to write these lawyers off as lazy or incompetent, but that misreads the failure. The people getting sanctioned are not fools. They are professionals fooled by a system engineered to sound authoritative. The output arrives formatted like real legal writing, in the register lawyers trust, with none of the tells that usually signal a bad source. There is no broken link, no obvious typo, no amateur phrasing to trigger suspicion. The fabrication wears the exact costume of legitimate research. That is precisely what makes it dangerous, and it is why the only safe assumption is that every AI-generated citation is false until independently confirmed against a real database.
The legal profession is now learning, at real cost, a lesson this site has documented across every domain these tools touch. The danger was never that AI would obviously fail in ways anyone could spot. The danger is that it fails invisibly, wrapped in fluent confidence, in exactly the situations where a wrong answer does the most damage. A courtroom is one of the least forgiving places to discover that a machine will lie to you without knowing it is lying. Anyone weighing how much to trust these systems for serious work can start with our running record of how ChatGPT's confidence diverges from its accuracy and the wider timeline of AI failures that keeps growing.
The Verdict
Fabricated case citations have crossed from punchline to precedent. Judges are sanctioning multiple lawyers at once, penalties are climbing into six figures, and courts have begun assigning liability to the companies behind the models. The confident invention of cases that never existed is not a glitch that will be patched. It is how these systems work, and the legal system is now billing for it.