The Mata v. Avianca Problem: How to Use AI in Law Without Fabricated Citations

A $5,000 fine, a public apology to six federal judges, and a name that every lawyer in America now associates with AI gone wrong. That’s the legacy of Mata v. Avianca, Inc., No. 22-cv-1461 (S.D.N.Y. 2023) — the case where attorney Steven Schwartz submitted a brief citing six cases that didn’t exist, all fabricated by ChatGPT.

But here’s what most coverage of this case gets wrong: the problem wasn’t that a lawyer used AI. The problem was that a lawyer used the wrong kind of AI for the task, then skipped verification entirely. Understanding that distinction is the difference between using AI responsibly and becoming the next cautionary tale. And as of late 2025, Damien Charlotin’s AI Hallucination Cases Database documents over 300 incidents of AI-fabricated citations in court filings — up from a handful in 2023 to two or three new cases per day.

This article breaks down what actually happened, why it keeps happening, and — most critically — why contract review AI operates on a fundamentally different risk model than research AI. If you’ve been hesitant to adopt AI tools because of Mata v. Avianca, you may be avoiding the wrong thing.

What Actually Happened in Mata v. Avianca

The facts are straightforward and worth getting right.

In 2022, Roberto Mata filed a personal injury lawsuit against Avianca Airlines, alleging a knee injury from a metal serving cart on an international flight. When Avianca moved to dismiss, Mata’s attorney Peter LoDuca filed an opposition brief. The brief was largely drafted by his colleague Steven Schwartz, who used ChatGPT to research supporting case law.

ChatGPT generated citations to six cases that sounded real — complete with docket numbers, court names, and plausible holdings. Cases like Varghese v. China Southern Airlines, Shaboon v. Egyptair, and Petersen v. Iran Air. They had the structure, cadence, and citation format of genuine case law. They were entirely fabricated.

When Avianca’s attorneys couldn’t locate the cited cases, Judge P. Kevin Castel ordered Schwartz to produce copies. Schwartz went back to ChatGPT and asked whether the cases were real. ChatGPT confirmed they were. He submitted that confirmation to the court.

On June 22, 2023, Judge Castel issued sanctions — a $5,000 fine against Schwartz, LoDuca, and their firm Levidow, Levidow & Oberman. The court also required them to send individual letters to each of the six judges falsely identified as authors of the fabricated opinions, along with copies of the sanctions order.

The case made international headlines. It became the most referenced AI-in-law case in history. And it terrified lawyers who were considering AI adoption.

Why AI Hallucination Happens — in Terms Lawyers Understand

Hallucination isn’t a bug in the software. It’s a feature of how large language models work — and understanding the mechanism matters for assessing risk.

Large language models like ChatGPT and Claude predict the most statistically likely next sequence of words given a prompt. They don’t retrieve facts from a database. They don’t look up cases in Westlaw. They generate text that sounds right based on patterns in their training data.

Legal citations are especially vulnerable because:

• Case names follow predictable patterns. A name like Petersen v. Iran Air sounds like a real aviation injury case because it matches thousands of real citation patterns the model has seen.
• Legal writing is formulaic. Holdings, procedural histories, and citation formats follow rigid conventions. AI can mimic the form perfectly while fabricating the substance.
• Lawyers are trained to trust citations. When you see a properly formatted citation — 678 F.Supp.3d 443 (S.D.N.Y. 2023) — your instinct is to trust it, not verify it. That trust is earned in normal legal practice. It’s exploited by AI hallucination.

This isn’t unique to ChatGPT. Any general-purpose language model can hallucinate. Stanford research has documented that hallucination rates for legal citations range from 6% to over 30% depending on the model, the complexity of the question, and the jurisdiction.

It Keeps Happening: Post-Mata Sanctions Cases

Mata v. Avianca wasn’t an isolated incident. It was a preview.

Noland v. Land of the Free, L.P. (2025) — A California appellate court found that “nearly all of the legal quotations in plaintiff’s opening brief, and many of the quotations in plaintiff’s reply brief, were fabricated.” The court imposed $10,000 in sanctions — double the Mata penalty.

Johnson v. Dunn (N.D. Ala., July 2025) — The court went further than fines: it disqualified the attorneys from representing their client for the remainder of the case and directed the clerk to notify bar regulators in every state where the attorneys were licensed.

Arizona Social Security Case (August 2025) — A judge found that 12 of 19 cited cases were fabricated, misleading, or unsupported, sanctioning the attorney whose brief was “replete with citation-related deficiencies consistent with artificial intelligence generated hallucinations.”

And in a notable 2025 development, courts began sanctioning lawyers for failing to detect their opponent’s AI-fabricated citations — establishing that the verification duty runs both ways.

The pattern across every sanctions case is identical: a lawyer used a general-purpose AI tool for legal research, submitted the output without verification, and fabricated citations ended up before a judge.

The Critical Distinction: Research AI vs. Review AI

This is the argument that most Mata coverage misses entirely, and it’s the one that should change how you think about AI risk.

Research AI (High Hallucination Risk)

General-purpose AI tools like ChatGPT, Claude, and Gemini are generative — they create text from scratch. When you ask them to find supporting case law, they don’t search a legal database. They generate text that looks like case law. Sometimes they get it right (because the case appeared in training data). Often they don’t.

The hallucination risk profile:

• Generates citations, case summaries, and legal analysis from scratch
• No built-in source verification
• Designed to produce plausible-sounding content
• Outputs fabricated cases, misquoted holdings, and invented statutes
• Confidence level of the output has no correlation with accuracy

Contract Review AI (Fundamentally Different Risk)

Purpose-built contract review tools operate on a completely different model. They don’t generate legal citations or case law. They analyze a specific document you provide as input.

When a contract review AI examines your NDA, it:

• Identifies clauses that exist in the document you uploaded
• Categorizes those clauses by type (indemnification, termination, IP assignment)
• Scores risk based on what’s present — and flags what’s missing
• Generates structured output (risk scores, clause categories) not freeform legal analysis
• Never cites case law, statutes, or legal authority it might fabricate

There’s nothing to hallucinate when the task is “read this paragraph and tell me whether it contains a unilateral termination right.” Either the language is there or it isn’t. The AI is classifying existing text, not inventing new text.

This doesn’t mean contract review AI is infallible — it can miscategorize a clause, miss a bespoke provision, or score risk differently than you would. But those are accuracy issues, not hallucination issues. And they’re the same types of errors a junior associate or paralegal might make, which is why human review remains non-negotiable.

Building a Hallucination-Proof AI Workflow

Whether you’re using AI for research, review, or drafting, these practices protect you.

Before You Use Any AI Tool

1. Match the tool to the task. If a purpose-built tool exists for what you need — contract review, document comparison, legal research with verified citations — use it instead of general-purpose AI. This is the single most effective risk reduction strategy.

2. Understand the tool’s architecture. Does it generate text from scratch (high hallucination risk) or analyze documents you provide (lower risk)? Does it cite sources it retrieves from a database (CoCounsel, Lexis+ AI) or generate citations from training data (ChatGPT)? ABA Formal Opinion 512, issued July 2024, requires lawyers to understand how their AI tools work before relying on them.

3. Test on known documents first. Before using any AI tool on client work, run it against a contract you’ve already reviewed manually. Compare the AI’s output to your own analysis. Where does it agree? Where does it diverge? Where is it wrong?

During AI-Assisted Work

4. Never submit AI output without line-by-line review. For legal research: verify every citation in Westlaw, Lexis, or Google Scholar. Read the actual opinion — don’t trust the AI’s summary of the holding. For contract review: check every flagged clause against the actual contract language. Verify that “missing clause” findings are actually absent from the document.

5. Be skeptical of confidence. AI doesn’t express uncertainty the way humans do. A fabricated citation reads with the same confidence as a real one. Treat all AI output as a first draft requiring verification, regardless of how polished it appears.

6. Document your review process. Keep a record of what tool you used, what it produced, and how you verified the output. This protects you against malpractice claims and bar complaints. It also satisfies the supervisory requirements under ABA Model Rule 5.3.

After AI Review

7. Apply professional judgment to every recommendation. AI doesn’t know your client’s business objectives, risk tolerance, negotiation leverage, or the relationship dynamics with the counterparty. These factors determine whether a flagged “risk” actually matters. Your judgment is what clients pay for — AI just gives you a faster starting point.

8. Sign off on the final work product as your own. If it has your name on it, you own it. Period. AI-assisted work carries the same professional responsibility as any other work product, as Judge Castel emphasized in the Mata sanctions order.

What Courts and Bar Associations Now Require

The regulatory response to AI hallucination is accelerating. As of early 2026, over 300 federal judges have issued standing orders, local rules, or pretrial orders addressing AI use in court filings.

Common requirements include:

• Disclosure of AI use. Many courts require attorneys to identify which AI tools were used and which portions of a filing were AI-assisted.
• Certification of accuracy. Several judges, including Judge Baylson in the Eastern District of Pennsylvania, require attorneys to certify that every citation has been verified for accuracy.
• Identification of the specific tool. Some orders require naming the AI tool used, not just disclosing AI assistance generally.

At the state level, bar associations across the country are issuing guidance. California emphasizes understanding LLM risks before use. Florida’s Opinion 24-1 mandates disclosure when AI impacts billing. Texas Opinion 705 requires human oversight of all AI-generated work product.

The trend is clear: use AI, but verify and disclose. And the ABA’s checklist for responsible AI use published in early 2026 consolidates these requirements into a practical framework.

The Lesson — and the Opportunity

Mata v. Avianca wasn’t a failure of artificial intelligence. It was a failure of verification. Steven Schwartz didn’t get sanctioned for using ChatGPT. He got sanctioned for submitting fabricated citations without checking whether they were real. Every subsequent sanctions case follows the same pattern.

The lawyers who will thrive aren’t the ones avoiding AI — they’re the ones using it with the right tools and the right workflow. For contract review specifically, purpose-built AI tools that analyze documents rather than generate citations eliminate the hallucination risk that caused Mata. The risk isn’t zero — miscategorization and accuracy issues exist — but it’s a fundamentally different category of risk, one that standard attorney review practices are designed to catch.

If you’ve been avoiding AI because of Mata v. Avianca, you may be solving the wrong problem. The question isn’t whether to use AI — it’s whether you’re using the right AI for the right task, with the right verification process in place.

For lawyers ready to start with contract review AI that’s designed for verification rather than hallucination, Clause Labs’s free analyzer lets you upload any contract and see a structured risk analysis in under 60 seconds — no citations to fabricate, no case law to verify, just clause-by-clause analysis of the document you provide. Try it on your next contract and see how purpose-built AI differs from the general-purpose tools that created the Mata problem.

For a deeper look at the ethical framework governing AI use in legal practice, read our guides on whether AI contract review is ethical and how ABA Rule 1.1 applies to technology competence.

Frequently Asked Questions

Can contract review AI hallucinate?

Contract review AI can make errors — miscategorizing a clause, missing a bespoke provision, or misjudging risk severity. But it doesn’t hallucinate in the Mata sense because it doesn’t generate citations, case law, or legal authority. It analyzes the specific document you provide and produces structured output (risk scores, clause identification, missing provisions) rather than freeform legal text. The risk profile is accuracy, not fabrication.

How is Clause Labs different from ChatGPT for legal work?

ChatGPT is a general-purpose language model that generates text from scratch — including citations that may not exist. Clause Labs is a purpose-built contract review tool that analyzes the specific document you upload. It identifies clauses, scores risk, flags missing provisions, and suggests edits based on what’s actually in your contract. It never generates case citations or legal authority. The architecture eliminates the hallucination vector that caused Mata v. Avianca.

What should I do if I suspect AI output contains hallucinated content?

Stop, verify, and document. Check every citation against a verified legal database (Westlaw, Lexis, Google Scholar). If you find fabricated content, do not submit it. Remove it from your work product. If you’ve already submitted a filing containing unverified AI content, consider notifying the court proactively — courts have shown more leniency toward attorneys who self-report than those who are caught.

Do I need to disclose that I used AI to review a contract?

This varies by jurisdiction and context. For court filings, over 300 federal judges now require AI disclosure. For transactional work (contract review and negotiation), disclosure requirements are less defined, but ABA Formal Opinion 512 recommends transparency with clients about AI use, particularly regarding confidentiality and billing. Check your state bar’s guidance — several states now have specific AI disclosure requirements.

Has any lawyer been sanctioned specifically for using AI contract review tools?

As of early 2026, no. Every documented sanctions case involves general-purpose AI (primarily ChatGPT) used for legal research — specifically, the submission of fabricated citations. No sanctions case has involved a purpose-built contract review tool used within its designed parameters. The risk pattern is clear: sanctions arise from unverified AI-generated citations, not from AI-assisted document analysis.

---

This article is for informational purposes only and does not constitute legal advice. Consult a qualified attorney for advice specific to your situation.