It is easy to treat “AI hallucinates case law” as an abstract warning until you see specific, documented examples of what that actually looked like inside a real filing. The following are drawn from the public record — court orders, published opinions, and Damien Charlotin’s AI Hallucination Cases database, which had documented more than 1,750 cases as of mid-2026 — not from hypothetical scenarios.
Case citations invented wholesale
In Mata v. Avianca, Inc., No. 22-cv-1461 (S.D.N.Y. 2023), attorney Steven Schwartz used ChatGPT to research airline-liability precedent and submitted a brief citing several cases, including one styled Varghese v. China Southern Airlines, that did not exist. ChatGPT had produced not just a case name but a full citation, procedural history, and quoted excerpts for a case that had never been decided. Judge P. Kevin Castel imposed a $5,000 monetary sanction on the attorneys and their firm after finding the conduct involved bad faith once the fabrication was challenged.
Generic-sounding names that should have raised suspicion
In Kruse v. Karlen (Mo. Ct. App. E.D. 2024), a self-represented appellant filed a brief containing 24 case citations, of which the court found only two were genuine and even those did not stand for what the brief claimed. Several of the fabricated citations used generic, almost template-like names — the kind of placeholder-sounding case names (a pattern seen repeatedly in AI-hallucination cases generally) that are themselves a red flag worth training yourself to notice. The appellant, who had outsourced the brief to an unlicensed “legal consultant” who used AI without disclosing it, was sanctioned $10,000.
A high volume of defective citations in a single filing
In Coomer v. Lindell (D. Colo.), a defamation case involving MyPillow CEO Mike Lindell, Judge Nina Y. Wang found a filing from defense attorneys Christopher Kachouroff and Jennifer DeMaster contained nearly thirty defective citations — a mix of misquotes, incorrect attributions, and citations to cases that did not exist. This is a useful example because it shows fabrication is not always an isolated one-off error; it can appear at volume across an entire brief when no verification step was run at all.
An internal firm AI tool generating the fabrications, not a public chatbot
Wadsworth v. Walmart Inc. (D. Wyo. 2025) is a useful corrective to the assumption that this only happens with consumer tools like ChatGPT. Attorney Rudwin Ayala used his firm’s in-house AI platform (MX2.law) to add case law to a motion in limine, prompting it to “add more case law regarding motions in limine.” Eight of the nine cases the tool returned did not exist. Judge Kelly Rankin revoked Ayala’s pro hac vice status and fined him $3,000; the two attorneys who signed the brief without independently verifying it were fined $1,000 each.
A real case, correctly identified, still cited for the wrong thing
Not every documented example involves a fully invented case. Park v. Kim, 91 F.4th 610 (2d Cir. 2024), involved a citation attorney Lee admitted she generated using ChatGPT after struggling to find a relevant case through traditional research tools. The Second Circuit referred her to its Grievance Panel, noting that presenting an unverified citation — regardless of the tool used to find it — falls below counsel’s basic obligation to make a reasonable inquiry into the law before filing.
Outside the United States
The pattern is not US-specific. In Ayinde v. London Borough of Haringey and the related Al-Haroun v. Qatar National Bank matter, heard together by the England and Wales High Court in June 2025 under its “Hamid” jurisdiction, one filing cited five cases that could not be located, and a second cited 45 cases of which 18 did not exist. The court ordered costs sanctions and set out a public warning that misuse of AI-generated citations can, in more serious cases, lead to contempt findings or referral for criminal investigation.
What these examples have in common
Across every example above, the underlying failure was not the AI tool — it was the absence of a verification step before filing. Different tools (ChatGPT, an internal firm platform, Google Bard) and different jurisdictions produced the same failure pattern, which is precisely why the fix is process-based rather than tool-specific: verify existence, quotes, and proposition support for every citation, every time, regardless of which AI generated the draft.
A pattern within the pattern: fabricated quotes attached to real cases
Beyond wholesale invention, a subtler and arguably more dangerous category appears repeatedly in the documented record: a real, correctly-cited case with a fabricated or altered quotation attached to it. This is harder to catch because an existence check alone passes cleanly — the case is real, the reporter and volume are correct, and only a word-for-word comparison against the actual opinion text reveals the quoted language does not appear anywhere in it. Several entries in Damien Charlotin's database fall into this specific sub-category, and it is worth treating as a distinct failure mode from full fabrication, requiring its own dedicated check rather than being caught incidentally by an existence-only verification process.
Fabricated procedural history and docket numbers
Some documented examples go beyond fabricating a case name to inventing an entire, internally consistent procedural history: a docket number, a filing date, a panel of judges, and a summary of holdings, all presented with the same confident tone as accurate information. This level of detail is part of what makes these fabrications convincing on a normal read — a docket number looks like exactly the kind of specific, verifiable-seeming detail nobody expects to be invented, which is precisely why it sometimes is.
What these examples suggest about training your own eye
Reviewing enough documented examples builds a useful intuition over time: watch for citations that feel unusually perfectly on-point for your exact argument, generic or placeholder-sounding party names, and citations you do not specifically remember encountering in your own prior research on a topic you know reasonably well. None of these signals is proof of fabrication by itself, and none of them replaces the actual existence and quote checks — but they are useful triggers for extra scrutiny before you even run the formal verification step.
A common question
Are these examples rare outliers, or is this happening at real scale?
The Charlotin database’s growth — from roughly 1,000 documented US cases in early 2026 to over 1,750 globally documented cases by mid-2026 — suggests this is a steadily growing pattern, not a handful of isolated incidents. That said, the database counts cases where a court explicitly addressed suspected AI hallucination, which is likely a fraction of actual occurrences; many fabricated citations are presumably caught and quietly corrected before ever reaching a judge, or never caught at all.