AI Citation Sanctions: What 49 Real Court Cases Show About How Judges Are Punishing Fabricated Case Law

Published 2026-07-16 · By Andy Gaber, founder of Citation Safe

>

Why a “sanctions database” and not just a list of scandals

Every AI-hallucinated-citation story gets told the same way in the trade press: one lawyer, one bad brief, one embarrassing headline. That framing undersells what’s actually happening. As of this writing, Citation Safe’s public database tracks 49 separate court decisions where a judge took some sanctioning action over AI-fabricated citations — not 49 headlines, 49 rulings, each with a docket number, a court, a decision date, and (where the record supports it) a description of exactly what was fabricated and what the court did about it (citationsafe.com/sanctions-database).

The database is not original research. It is compiled from Damien Charlotin’s AI Hallucination Cases Database, a CC0-licensed academic project that tracks legal decisions — court orders, opinions, and rulings — where a court or tribunal addresses AI use, alleged or established, in more than a passing reference (damiencharlotin.com/hallucinations). Citation Safe’s 49-case set is a filtered view of that larger record: specifically, the cases where the outcome was a warning, an order to show cause, a monetary sanction, a bar referral, or some combination of those, rather than every decision that merely mentions AI. Each entry in Citation Safe’s set links back to the underlying court order or contemporaneous reporting, and the individual case teardowns published alongside the database follow a fixed template built directly against that source record — no model writes the summary (citationsafe.com/blog/sanctions).

That distinction matters for anyone using this data to make a risk argument. The 49-case number is not “how many lawyers have gotten in trouble for AI hallucinations, ever.” It is the current size of one curated, continuously-updated subset. Charlotin’s broader database — the one Citation Safe draws from — is larger and growing faster: contemporaneous reporting put it at roughly 1,598 logged decisions by June 9, 2026, up from about 200 a year prior, an eightfold increase in twelve months (Scientific American, June 2026). Some of those 1,598 decisions involve allegations that didn’t result in a sanction, cases still pending, or foreign tribunals whose procedural posture doesn’t map cleanly onto “sanction” in the U.S. sense. The 49-case set is deliberately the narrower, harder-edged slice: sanction actually imposed, or a show-cause order actually issued.

What actually gets fabricated

Reading through even a handful of entries in the database shows the fabrication problem isn’t one failure mode — it’s at least three, and they require different detection methods.

Fabricated case law. The citation refers to a case that does not exist at all. In Patterson v. Nuvision Credit Union (Cal. Ct. App., 4th Dist., decided 2026-07-02), the opening brief cited a 1961 California appellate decision, a Seventh Circuit case, a Fifth Circuit unpublished opinion, and multiple bankruptcy decisions, none of which the court could locate as cited — five separate fabricated citations in a single filing (Citation Safe teardown, sourced to the underlying court PDF).

False quotes attributed to real cases. This is the harder failure mode to catch by skimming, because the underlying case is real — only the quoted language isn’t in it. The same Patterson brief attributed quotations to Romero, Tweel, Strong v. County of Santa Cruz, and Lake v. Reed that the court found do not appear in those opinions. In Cartagena v. Dixon, Blackburn, and T.A. Blackburn Law, the Southern District of New York found at least 17 instances of quoted language that did not match the cited sources — the attorney characterized the practice as paraphrasing; the court disagreed and struck the filing (citationsafe.com/sanctions-database/cartagena-v-dixon-blackburn-and-t-a-blackburn-law-2-20260710).

Misrepresented holdings. The case is real, correctly quoted, and cited for a proposition it doesn’t actually support. The Patterson court flagged exactly this: a criminal case (Romero) cited for the proposition that UCC provisions are binding — a real opinion, put to a use the court found inapposite.

This three-part taxonomy isn’t incidental to how Citation Safe frames the problem — it maps directly onto the three-layer verification approach the product runs against every citation before a brief is filed: does the citation exist, does a quoted passage actually appear in the source, and does the cited case support the proposition it’s cited for (citationsafe.com — teardown template, e.g. the Parnell case). Each layer catches a different failure mode; a tool that only checks whether a case exists (Layer 1) will miss both false quotes and misrepresented holdings.

The outcome spectrum: from a warning to bar referral

Not every case in the 49-case set ends the same way, and the spread is instructive for anyone trying to estimate real downside risk.

At the mild end, several entries resolve as a bare warning — the court identifies the problem, admonishes counsel, and moves on without a formal sanction. Julia Rose v. Arts Bonita, Inc. (M.D. Fla., 2026-07-12) and Jordan Slach v. City of Battle Ground (W.D. Wash., 2026-07-02) both resolved this way. A step up, admonishment appears as a distinct, slightly harder-edged category from a plain warning in several entries, including John Hurt v. Ampcus, Inc. (E.D. Tex., 2026-07-08).

Order to show cause is the procedural stage where a court requires the attorney to justify why sanctions shouldn’t follow — it is not itself a punishment, but it is the point where the paper trail becomes serious. Michael L. Ruiz v. Magellan Financial & Insurance Services (D. Ariz., 2026-07-08) and Pedro Paulo Mansur Pagano Sampaio v. Wells Fargo Bank (C.D. Cal., 2026-07-02) both sit at this stage.

Monetary sanctions are directly comparable across cases in a way qualitative outcomes aren’t. Del Biaggio v. Bansen (Cal. Ct. App., 4th Dist., 2026-07-10) carried a $1,500 penalty; Patterson v. Nuvision Credit Union carried $500 plus a bar referral for the unauthorized practice of law. These are modest by the standard of headline-grabbing 2023-era sanctions (the Mata v. Avianca case that started this entire genre carried a $5,000 penalty), but they are not the largest penalties on record — reporting elsewhere describes a 2026 Oregon case with roughly $109,700 in combined sanctions, fines, and opposing-party fees across two attorneys, described as the largest aggregate penalty of the current wave (Forbes, “Judge Boots Attorneys After Both Sides Submit AI-Hallucinated Citations,” June 2026).

Bar referral is the outcome with consequences that extend past the case at hand — a referral to a state or federal grievance committee, which can trigger a separate disciplinary proceeding independent of whatever happens in the underlying litigation. Marion Parnell, Jr. v. Florida Department of Corrections (11th Cir., 2026-07-10) combined a partly-struck brief, an adverse costs order, and a bar referral in a single decision. Cartagena is notable for a second reason: the court’s order explicitly referenced a prior $5,000 sanction against the same attorney in Jakes v. Youngblood before granting the motion to strike and referring him to the Grievance Committee — a documented repeat offense, not a first mistake (citationsafe.com/sanctions-database/cartagena-v-dixon-blackburn-and-t-a-blackburn-law-2-20260710).

At the far end, consequences can reach beyond the sanctioned attorney’s own filing. In Pooja Ramesh Singh v. Jammu and Kashmir Bank Ltd. & Anr. (Supreme Court, India, decided 2026-07-02), the outcome recorded is that the underlying NCLT and NCLAT tribunal judgments were set aside — meaning the fabricated-citation problem didn’t just cost a sanction, it unwound the substantive rulings the fabricated authority had been used to support.

Jurisdiction spread: mostly U.S., not exclusively

The tracked 49 cases are heavily concentrated in U.S. federal and state courts, but the record is not U.S.-only. The current set includes decisions from Canada (Adeleke v. Minister of Citizenship, Federal Court, 2026-07-06), the United Kingdom (Tobosaru v Romania; Tofan v Romania, High Court, 2026-07-08), Australia (multiple entries, including Jovanovic v Hobart City Council before the Tasmanian Supreme Court and Ba v Sterling Parts Australia Pty Ltd before the Family Court), and India (the Supreme Court case above) (citationsafe.com/sanctions-database). Within the U.S. entries, the spread runs across federal district courts, circuit courts of appeals, state appellate courts, and at least one state supreme court and one Delaware Chancery Court matter — this is not a problem isolated to any single tier of the judiciary.

The tool-attribution problem: courts mostly don’t say

One pattern is consistent enough across the 49-case set to be worth calling out on its own: courts almost never name a specific AI product in the order. Citation Safe’s own teardowns record tool attribution as “Implied” (the record indicates AI use was involved but the type of use is inferred rather than stated) or “Unidentified” (a court order, brief, or contemporaneous reporting doesn’t name a specific product) far more often than it records a named tool — and the site is explicit that it does not fill this gap with a guess: “Charlotin’s public database records tool attribution only where a court order, brief, or reporting on the matter states it explicitly... we do not guess” (citationsafe.com/blog/sanctions/marion-parnell-jr-v-florida-department-of-corrections-20260710).

This has a practical implication for anyone trying to argue that one AI legal-research product is safer than another based on sanctions data alone: the sanctions record mostly can’t answer that question, because most sanction orders don’t say which tool was used. What can be measured directly is hallucination rate under controlled testing — which is where the Stanford RegLab/HAI study’s figures (Westlaw AI-Assisted Research ~33%, Lexis+ AI ~17%) are more useful than the sanctions database for tool-to-tool comparison, precisely because that study tested named products against known queries rather than relying on what a court order happened to disclose (Magesh, Surani, Dahl et al., 2024, cited via citationsafe.com/scorecard).

Why lawyers keep doing this after very public warnings

The obvious question — why would anyone still file an AI-hallucinated citation in mid-2026, three years after Mata v. Avianca made international news — has a partial answer in the record itself: some of them are repeat offenders. The attorney in Cartagena had already been sanctioned $5,000 in a prior matter for the same underlying conduct before the court struck his filing and referred him a second time. That is not a story about a lawyer who didn’t know the risk. It’s a story about a workflow that didn’t change even after the first sanction — which is a stronger argument for a mechanical, pre-filing verification step than a warning ever will be, because warnings clearly aren’t sufficient on their own.

What this means for a filing workflow

• A hallucinated citation surviving to a court order means it survived every human check in the drafting process. Associate review, partner review, and opposing counsel’s read all failed to catch it in most of these 49 cases — which argues for a deterministic, non-LLM verification step as a final gate, not a replacement for human review.
• Quote-level and proposition-level fabrication is at least as common as invented case names, and it’s the harder failure mode to catch by eye, because the underlying case is real. A verification process that only confirms a case exists (an “existence check”) will still miss a large share of what’s actually landing lawyers in front of a grievance committee.
• Sanctions severity does not track neatly with jurisdiction or case type. A federal circuit court, a state appellate court, and a foreign high court all appear in the same 49-case set, at outcomes ranging from a warning to a bar referral to a tribunal judgment being set aside.
• Repeat offenses are already in the public record. At least one attorney in the tracked set had been sanctioned before for the same underlying conduct — a documented case for treating this as a process failure, not an isolated lapse.

Sources

• Citation Safe, Public AI Legal Sanction Case Database — 49 tracked cases, accessed 2026-07-16.
• Citation Safe, individual sanction teardowns: Marion Parnell, Jr. v. Florida Department of Corrections; Cartagena v. Dixon, Blackburn, and T.A. Blackburn Law; Patterson v. Nuvision Credit Union.
• Damien Charlotin, AI Hallucination Cases Database (CC0-licensed, source database for the above).
• Scientific American, “Why lawyers keep citing fake cases invented by AI”, June 2026 (1,598-case figure).
• Forbes, “Judge Boots Attorneys After Both Sides Submit AI-Hallucinated Citations”, June 2026 (Oregon $109,700 aggregate penalty).
• Citation Safe, Public FVR Scorecard, citing Magesh, Surani, Dahl et al., 2024, “Hallucination-Free? Assessing the Reliability of Leading AI Legal Research Tools” (Stanford RegLab/HAI).

Citation Safe is a verification workflow tool, not legal advice. No attorney-client relationship is created. Human review is always required. Case summaries above are compiled from public court records and contemporaneous reporting; always confirm case details against the primary source before relying on them.

—

Check a brief before you file it →