A public database has logged more than 1,227 court cases in which AI fabricated legal citations, and the count climbs by several every day. The cause is not a bug. Large language models generate text one token at a time. They complete the standardized shape of a citation, producing real-looking author names, court names, and reporter numbers for cases that never existed. If you cite, quote, or rely on anything an AI hands you, the source is unverified until you open the original yourself.
How many fake AI court cases have been recorded
Damien Charlotin, a research fellow at HEC Paris, maintains an open database of legal decisions where a court found that a party relied on AI-hallucinated content. It crossed roughly 719 logged cases in January 2026 and passed 1,227 within a few months. By mid-2026 the running tally had moved past 1,450. The exact figure is a moving target by design: it is a live tracker, not a fixed report. That motion is the point. The number is not a finished count of a past problem. It is the readout of a meter that ticks up while you read this sentence.
At the 1,227 milestone, roughly 800 of the catalogued cases came from United States courts, with the rest spread across other jurisdictions. The database only counts cases where a judge explicitly addressed the hallucinated content, which means it captures the ones that got caught. The true number of filings carrying a fabricated citation is likely higher, because courts do not check every authority in every brief. The tracker is a record of detections, not of attempts.
Roughly two-thirds of the logged cases at the 1,227 mark were US filings. The database tracks only what judges caught, so treat every count as a floor, not a ceiling.
Why AI invents citations that look perfect
A language model does not store a library of cases and look them up. It predicts the next token based on patterns it absorbed during training. A legal citation has a rigid, predictable form: a party name, the letter v, another party name, a volume number, a reporter abbreviation, a page, a court, and a year. The model has seen that shape tens of thousands of times. When you ask for a supporting case, it fills in that template with the most statistically plausible pieces, the way it would finish any familiar phrase.
Here is what most explainers get backwards: the model is not bad at law. It is dangerously good at it, excellent at imitating the surface style of legal writing, with its neutral citations, naming conventions, and judicial prose. Surface fidelity is exactly what fools the reader. The training signal rewards fluent, coherent text, not accuracy, and the model has no live connection to a case reporter to check whether the case it just named exists. Once it outputs a high-probability but fictitious citation, nothing inside the model flags it as false.
This is the same mechanism behind a more general failure: a base model would rather produce a confident-sounding answer than admit it does not know. The citation case is just the highest-stakes version, because the fabricated output carries the full authority of a court record. A confident wrong answer about a restaurant menu costs you a bad dinner. A confident wrong answer in a legal brief costs a sanction and a name in a public ledger of the mistake.
Fabrication, false quotes, and misattribution are different failures
Not every hallucination is a fully invented case. The tracker sorts entries into categories, and the distinction matters for how you verify. A wholly fabricated citation points to a case that never existed. A false quote attaches words to a real case that never said them. A misrepresentation describes a real case but gets its holding backward. The most dangerous category is the false quote, because the case is real, the citation resolves, and a quick check that only confirms the case exists will wave it straight through. The first category fails an existence test. The second and third pass it, which is what makes them harder to catch than the obvious fakes.
Have lawyers been punished for AI-fabricated citations
The landmark case is Mata v. Avianca. In 2023, a lawyer in the Southern District of New York filed a brief built on cases that ChatGPT had invented, then doubled down when the court asked for copies. The judge imposed a 5,000 dollar sanction on the attorneys under Rule 11 of the Federal Rules of Civil Procedure.
The lawyer asked ChatGPT whether the cases were real. It said yes, they were in Westlaw and LexisNexis. They were not. The AI vouched for its own fabrications, and a trained attorney believed it.
Mata was the warning shot, not the exception. Judges across the United States have since sanctioned, fined, and publicly named lawyers for the same mistake, and on at least one day in early 2026 more than a dozen separate US decisions flagged suspected AI hallucinations in filings. The cost is not only money. It is credibility with the bench, a damaged client matter, and a name that now appears in a searchable public record of the error.
This is not only a lawyer problem
Courts catch fabricated citations because opposing counsel and judges check sources for a living. Most knowledge work has no such adversary. The analyst who pastes a statistic into a board deck, the journalist quoting a study, the doctor citing a guideline, the student building a bibliography, the consultant referencing a regulation: all face the same risk with none of the verification built in. A fabricated source in a sales memo or a grant application fails just as silently. It just never produces a published opinion.
The absence of an adversary makes the danger worse, not better. In litigation, a fabricated case meets an opponent paid to find it and a judge with the power to punish it, so most fakes get exposed and counted. In a board deck or a research memo, no one is paid to attack your sources. A fabricated statistic can travel through a slide, into a decision, and out to a customer without a single person checking whether the cited study exists. The legal world is not uniquely exposed. It is simply the one field where the checking is loud enough to leave a record.
The danger scales with how authoritative the output looks. The more a model's answer resembles a formatted citation, statistic, or quotation, the more it deserves a manual check, because format is exactly what the model is best at faking.
A verification protocol for any AI source
The fix is a habit, not a tool. Treat every citation, quote, statistic, and named fact from an AI as a claim to be confirmed against a primary source before it leaves your hands. Four steps cover almost every case.
- Make the model show its work. Ask it to quote the exact passage it is relying on and name the source precisely. A model that fabricated the claim will often stumble, contradict itself, or quietly revise when pushed for specifics.
- Open the primary source yourself. Do not trust the model's summary of a case, paper, or report. Find the original document in an authoritative database or on the publisher's site and read the relevant passage.
- Confirm the citation resolves to the right thing. Check that the case number, volume, page, or DOI actually points to the document the model named, not merely to something real.
- Verify the claim, not just the existence. Read enough of the source to confirm it actually says what the model claimed. Existence is necessary but not sufficient; the holding or finding has to match.
Builders can reduce the rate of fabrication with grounding and retrieval, feeding the model real documents and asking it to cite only from those, but reduction is not elimination. Even a grounded system can misquote a source it was handed. The human check at the end is the only step that fully closes the gap.
| Question | Looks real but fabricated | Genuinely verified |
|---|---|---|
| Does the source exist? | Often no, or it is a real source with a fake quote attached | Confirmed in a primary database or on the publisher site |
| Where did you read it? | Only inside the AI's answer | In the original document, opened directly |
| Does the citation resolve? | Number or DOI points nowhere or to a different work | Citation maps to the exact document named |
| Does the claim match? | Holding or statistic is reversed or invented | Source text confirms the specific claim |
Reading from a source beats reconstructing from memory
Fabrication thrives when a model has to reconstruct a fact from training patterns instead of reading it from a source you actually own. That is the gap MemX is built to close. MemX is a consumer AI memory app, an external memory layer over your own documents, notes, and photos across Android, iOS, and WhatsApp, so an assistant can ground answers in material you provided rather than guessing at it. It is private by architecture, with per-user keys, encryption at rest, and an on-device first pass.
Grounding in your own files narrows the room for invention, but it does not replace the final human check, and it should not. The point of a memory layer is to make verification fast, by keeping the real source one click away instead of buried in a model's weights. The discipline stays the same: read the source before you rely on it.
01why does AI make up court cases and citations
Language models predict text token by token and complete the familiar shape of a citation rather than retrieving a real record. They have seen the format thousands of times, so they fill the template with plausible names and numbers for cases that do not exist.
02how many fake AI court cases have been recorded
A public database run by Damien Charlotin logged over 1,227 cases of AI-fabricated content submitted to courts by mid-2026, up from about 719 in January 2026. The total keeps rising daily and has since passed 1,450.
03have lawyers been punished for citing fake AI cases
Yes. In Mata v. Avianca (2023), a New York federal judge fined the attorneys 5,000 dollars under Rule 11 for filing ChatGPT-invented cases. Many US judges have since sanctioned or publicly named lawyers for the same error.
04is an AI citation safe if the case is real
Not necessarily. A real case can carry a fabricated quote or a reversed holding the model attached to it. The citation resolves and the case exists, so a quick check waves it through. Confirm the source actually says what the model claimed, not just that it exists.
05does this only affect lawyers
No. Lawyers get caught because courts check sources, but analysts, journalists, students, and clinicians face the same risk with no built-in verification. A fabricated source fails just as silently in a report or bibliography as in a legal brief.
The 1,227 figure is striking because it represents only the fabrications that a court bothered to document. Treat it as a smoke detector for every field that cites sources. The mechanism that produces a fake case also produces a fake statistic, a fake quote, and a fake reference, and the only reliable defense is to open the source before you stand behind it.
