1,600 Court Cases of Fake AI Citations. One Cause.

There is now a public database of more than 1,600 court decisions worldwide in which someone filed AI-fabricated material and a judge responded. It is the largest documented record of AI being used wrongly in professional work that exists anywhere, and it is worth reading even if you will never see the inside of a courtroom — because the failure is identical in every single case, and it is not the one people assume. The model did not malfunction. It produced plausible text, as designed. What was missing was the step where a human opened the citation and checked that it existed.

The numbers, and how fast they moved

Legal researcher Damien Charlotin maintains the tracker that everyone in this area cites. Its trajectory is the story:

• roughly 200 cases in mid-2025
• 719 by January 2026
• 1,227 by early April 2026
• 1,598 by 9 June 2026
• 1,668 by 2 July 2026

More than a thousand of those are US cases. The curve is not flattening, and it is worth being precise about what it counts: not every instance of AI misuse in litigation, but the subset where a court noticed and put its response on the record. The real denominator is unknowable and larger.

Money followed. US courts imposed at least $145,000 in sanctions for fabricated citations in the first quarter of 2026 alone. The largest single instance to date is roughly $109,700 in combined sanctions and fees in Couvrette v. Wisnovsky in the District of Oregon.

Courts stopped issuing warnings

For most of 2025 the judicial response was an admonition and an order to show cause. Through 2026 that changed, and the reasoning behind the change is the part worth stealing.

Analyses of decisions from February to April 2026 show the range: a $2,500 sanction in Fletcher v. Experian; $15,000 in punitive sanctions per attorney in Whiting v. City of Athens; case removal and denial of all compensation in Farris; public reprimands and referrals to bar disciplinary authorities. One court wrote that judges "should begin meeting this challenge" by "moving beyond admonitions and reprimands into more punitive sanctions."

Two rulings carry the principles that generalize beyond law.

The first is the rejection of novelty as a defence. As one decision put it: modern generative AI may be a new technology, but the same rules apply. The obligation was never "do not use new tools." It was "do not certify something you have not verified," and it long predates the tool.

The second is Farris, where the attorney had delegated the checking to staff. The court found that reliance fell short of his obligations as attorney of record. Verification could not be delegated away from the person whose name was on the filing. That is the sentence to carry into any other profession: the duty attaches to whoever signs, not to whoever typed.

The discipline has reached its maximum. A federal court in Pennsylvania sanctioned an attorney $1,500 and suspended him from that court for six months in June 2026. Nebraska produced the first reported US bar discipline removing a lawyer from practice entirely over AI filings. And in a February 2026 divorce appeal, an Omaha attorney filed a brief in which 57 of 63 citations were defective — 20 hallucinated cases, three fabricated decisions, invented statutory quotes. Sixty-three citations, and no one opened one.

Why this is a workflow problem, not an AI problem

Notice what is absent from the whole record: nobody was caught because the AI output looked wrong. It looked right. Fabricated case citations have the correct reporter format, plausible party names, and a year that fits the argument. That is what a language model is for.

The detection failure is structural. A citation is exactly the kind of claim humans check by pattern-matching on format rather than by opening the source, and AI output is perfectly formatted by construction. Every profession has an equivalent: a drug interaction that sounds standard, a load calculation with the right units, a statistic attributed to a real institution, an API method that fits the library's naming convention perfectly and does not exist.

The lawyers in the database were not careless people who could not read. They were people whose verification step was implicit — something that used to happen automatically, because finding a case in the first place meant looking at it. Generative AI removed the retrieval step and, with it, the incidental verification that had been riding along inside it. Nothing replaced it, and the gap only became visible when a judge went looking.

The verification protocol

If you use AI for anything you sign your name to, the fix is a step you perform explicitly rather than one you assume happened:

1. Open every cited source. Not search for the title — open the URL or the record. A source that cannot be opened does not exist for your purposes.
2. Quote-match the claim. Confirm the specific number, holding or sentence you are relying on appears in the source you opened. AI output frequently cites a real document that does not say the thing.
3. Check the source supports the use. A real case can be real, findable, correctly quoted, and still not stand for the proposition it is cited for. This is the failure mode that survives the first two checks.
4. Log what you checked. A dated list of what you verified is what separates an honest error from the appearance of a fabricated filing after the fact. Several of the harsher sanctions in the record followed the response to the discovery, not the original error.
5. Do not delegate step 1 to anyone whose name is not on it. That is Farris, and it is the rule with the widest reach outside law.

The whole protocol costs a few minutes per claim. Against $109,700, a six-month suspension, or the removal of a career, it is the cheapest process any professional will ever adopt — and the growth curve of that database says most people still have not adopted it.