Citation Errors, Hallucinations, and the Mata v. Avianca Lessons

Citation Errors, Hallucinations, and the Mata v. Avianca Lessons

By now most of you have heard about Mata v. Avianca, Inc., No. 22-cv-1461 (S.D.N.Y. 2023). It became the canonical cautionary tale about AI in legal practice almost overnight. But the details are worth revisiting carefully, because the court's reasoning maps directly onto Rule 11 obligations in ways that have implications well beyond one embarrassing sanctions order.

What Actually Happened

Roberto Mata sued Avianca Airlines for a personal injury claim. His attorney, Steven Schwartz of the firm Levidow, Levidow & Oberman, used ChatGPT to conduct legal research and drafted a brief citing six cases in opposition to Avianca's motion to dismiss. The problem: none of the cases existed. Not one. ChatGPT had fabricated case names, citations, and even plausible-sounding judicial opinions. When Avianca's counsel pointed out they couldn't locate the cases, Schwartz doubled down. He asked ChatGPT to confirm the cases were real, and ChatGPT obligingly confirmed they were. He then submitted an affidavit attesting to their authenticity.

Judge P. Kevin Castel was not amused. On June 22, 2023, the court issued an order to show cause why Schwartz and his colleague Peter LoDuca should not be sanctioned. The June 26 hearing transcript is genuinely painful reading. Schwartz testified that he had "never used ChatGPT before" for legal research, didn't understand that it could generate false information, and had relied on the tool's own assurances that the cases were real.

On June 22, 2023, Judge Castel sanctioned both attorneys $5,000 and ordered them to notify each judge falsely identified as an author of the fabricated decisions. The opinion explicitly stated that the attorneys had "abandoned their responsibilities" by failing to verify the AI's output through any conventional legal research tool.

The Rule 11 Framework

Rule 11 of the Federal Rules of Civil Procedure requires that every pleading, motion, or other paper presented to the court be signed by an attorney who certifies, to the best of their knowledge after "an inquiry reasonable under the circumstances," that the legal contentions are warranted by existing law or a nonfrivolous argument for its modification. The rule has teeth: sanctions can include penalties, attorney's fees, and nonmonetary directives.

The key phrase is "inquiry reasonable under the circumstances." Courts have always interpreted this as an objective standard. It doesn't matter whether you subjectively believed your citations were accurate. It matters whether a reasonable attorney, exercising due diligence, would have verified them.

Judge Castel's opinion in Mata didn't break new ground on Rule 11 doctrine. He applied the existing standard straightforwardly. The fabricated citations were not warranted by existing law. The attorneys had not conducted a reasonable inquiry. The fact that an AI tool generated the citations didn't change the analysis at all. If anything, it strengthened the case for sanctions, because relying on an unverified novel technology without cross-checking its output against Westlaw, Lexis, or even Google Scholar falls well below the objective reasonableness threshold.

Three Specific Failures the Court Identified

• No independent verification. Schwartz never checked any of the six fabricated cases against a traditional legal database. Any one search would have revealed the problem instantly.
• Reliance on the tool to verify itself. Asking ChatGPT to confirm its own output is circular. The court treated this as essentially meaningless as a verification step.
• Continued attestation after red flags. When opposing counsel flagged the citations as unfindable, the appropriate response was to immediately verify. Instead, Schwartz submitted further affirmations of authenticity.

The Ripple Effects

Since Mata, courts have moved quickly. As of early 2024, at least a dozen federal judges have issued standing orders requiring disclosure of AI use in legal filings. Judge Brantley Starr of the Northern District of Texas was among the first, requiring attorneys to certify that no portion of any filing was drafted by AI, or that any AI-drafted content was checked by a human for accuracy. The Eastern District of Texas, the District of Colorado, and several state courts have followed with their own variations.

The Fifth Circuit issued guidance in late 2023 requiring attorneys to certify either that generative AI was not used in preparing briefs or that all AI-generated content was verified by a human. Other circuits are watching.

These aren't hypothetical risks anymore. In Park v. Kim, No. 22-cv-2057 (E.D.N.Y. 2024), a court flagged potentially AI-generated citations and ordered supplemental briefing. In Kruse v. Karv Automotive Group, a Texas attorney was sanctioned after submitting a brief with fabricated case law. The pattern is recurring because the underlying technology problem hasn't changed: large language models generate plausible text, not verified facts.

What Rule 11 Actually Requires in an AI Context

The reasonable inquiry standard under Rule 11(b) has always been technology-agnostic. Whether you got a bad citation from a junior associate, a poorly maintained brief bank, or a generative AI model, the obligation to verify is yours. But AI introduces a specific wrinkle: the output looks authoritative. A hallucinated case citation from ChatGPT comes complete with a case name, a reporter citation, a year, and sometimes a convincing summary of the holding. It pattern-matches perfectly to what a real citation looks like. That makes the verification obligation more important, not less.

A few practical points on what Rule 11 compliance looks like when AI is part of the research workflow:

• Every citation must be independently verified against a primary legal database. This is non-negotiable. If it's not in Westlaw or Lexis, it doesn't go in the brief.
• AI disclosure obligations vary by jurisdiction, but the trend is clearly toward mandatory disclosure. Track the standing orders in every jurisdiction where you practice.
• Document your verification process. If you're ever called to account for a citation, you want a record showing when and how you confirmed it. This is basic risk management.
• Firm-level policies need to exist. Individual attorney discipline is fine, but the real exposure is at the firm and organizational level. Malpractice carriers are already asking about AI use policies.

The ABA issued Formal Opinion 512 in July 2024, confirming that lawyers who use AI tools must ensure competence, diligence, and communication consistent with Model Rules 1.1, 1.3, and 1.4. The opinion explicitly addresses generative AI and states that lawyers must understand the technology's limitations sufficiently to use it competently. This isn't a suggestion; it's an ethical obligation.

The Deeper Problem with General-Purpose AI for Legal Work

The root cause of the Mata debacle is that ChatGPT is a general-purpose language model. It was not designed for legal research. It has no concept of whether a case exists. It optimizes for plausible-sounding text, and in the legal domain, plausible-sounding text that is factually wrong can end careers and harm clients.

This is why the distinction between general-purpose AI and purpose-built, compliance-aware AI matters so much for regulated industries. A tool that doesn't know the difference between a real case and a fabricated one is fundamentally unsuitable for legal research without extensive human oversight. And "extensive human oversight" often means doing the work twice, which defeats much of the efficiency argument for AI adoption in the first place.

How FirmAdapt Addresses This

FirmAdapt's architecture is built around the principle that AI outputs in regulated environments must be verifiable and traceable. Rather than generating free-form text and hoping a human catches errors, FirmAdapt constrains AI outputs to verified data sources and maintains audit trails that document how conclusions were reached. For legal workflows, this means citations are grounded in actual source material, not probabilistically generated text that resembles a citation.

The platform also supports the kind of firm-level policy enforcement that Rule 11 compliance increasingly demands. Automated verification steps, jurisdiction-specific disclosure tracking, and human-in-the-loop review workflows are built into the system rather than bolted on as afterthoughts. For organizations where a single hallucinated citation can trigger sanctions, malpractice exposure, and reputational damage, that structural approach to compliance is the relevant differentiator.