Mata v. Avianca and the ChatGPT Fake Citation Sanctions - 01
Updated: Sep 2
AI Hallucination in Legal Practice • Confabulation Risk • Verification Failure • Professional Accountability
When a generative tool is trusted as a source of record.
Quick Reference
Field | Verified description |
Matter | Mata v. Avianca, Inc., No. 1:22-cv-01461 (S.D.N.Y.); 678 F. Supp. 3d 443 (2023) |
Court / judge | U.S. District Court, Southern District of New York; P. Kevin Castel, J. |
Sanctioned parties | Steven A. Schwartz; Peter LoDuca; and Levidow, Levidow & Oberman P.C. ("the Levidow Firm") |
What happened | An affirmation in opposition to a motion to dismiss cited nonexistent judicial opinions with fabricated quotes and citations generated by ChatGPT. Respondents stood by them after their existence was questioned. |
Sanction | $5,000, jointly and severally, paid into the Registry of the Court within 14 days |
Non monetary terms | Letters, within 14 days, individually addressed to plaintiff Roberto Mata and to each judge falsely identified as the author of a fabricated opinion |
Underlying case outcome | Avianca's motion to dismiss was granted the same day. The claim was time barred under the Montreal Convention's two year limitation period |
Legal basis for sanctions | Fed. R. Civ. P. 11. The court found subjective bad faith |
Central lesson | A generative tool is not a source of record. In high trust work product, every load bearing citation must be verified against an authoritative source before filing. |
1. What Happened
This did not begin as an AI case. It began as a deadline problem.
Roberto Mata sued Avianca over an alleged personal injury on an international flight. Because the flight fell under the Montreal Convention, Avianca argued the claim was filed too late and moved to dismiss it as time barred.
The AI failure entered through the opposition filing. To answer Avianca's motion, the filing cited cases that looked real: names, citations, quotations, holdings, and procedural details. Several were not real. Avianca's counsel told the court it could not locate the decisions, and the court ordered the respondents to produce them.
THE MISSED SIGNAL
The warning arrived before the sanction. Opposing counsel could not find the cases. After the production order, the respondents still had a path back to the record. They should have verified independently, withdrawn what could not be verified, and corrected the filing. The court later weighed the decision to stand by the material in finding subjective bad faith.
The court's findings, issued after a June 8, 2023 hearing, established that Steven A. Schwartz used the generative AI tool ChatGPT to perform the research and did not verify its output against an authoritative source. Schwartz was not admitted to practice in the district. Signing and filing fell to Peter LoDuca, the attorney of record. When the fabricated authorities were questioned, the respondents did not immediately withdraw them.
On June 22, 2023, Judge Castel imposed a $5,000 penalty, jointly and severally, payable into the Registry of the Court within 14 days. He also required letters to the plaintiff and to each judge falsely named as an author. The same day, the court granted Avianca's motion to dismiss because the claim was time barred.
2. Why This Was High Risk
Generative AI does not retrieve authority. It produces text that has the shape of authority, and it does so very convincingly. The confidence is the trap.
The fabricated cases carried the outer signals lawyers are trained to trust: case names, reporter citations, quotations, procedural posture, holdings, and named judges. That is what makes confabulation dangerous in a court filing. NIST AI 600-1 describes confabulation as confidently stated content that is false or erroneous.
The second, compounding risk was self verification. Asking the same model whether its own output was real does not create evidence. It produces more generated text. A closed loop where the system both generates the false content and the assurance is not a check.
THE CRITICAL INSIGHT
A verification step that runs inside the system being verified is not a control. It is the same system answering twice. This is the transferable lesson well beyond legal work. True assurance requires verification oversight independent from the agent of creation, so the check can fail on its own merits.
The use of a tool was not the failure. Judge Castel expressly noted there is nothing inherently improper about using a reliable AI tool for assistance. The failure was abandoning the human gatekeeping role while presenting the work as reliable.
3. The Governance Failure Pattern
The evidence backed pattern is source verification failure. A load bearing court filing left the firm carrying legal authorities that no one checked against an authoritative reporter.
Two secondary patterns are supported by the record:
Oversight absence: the record does not show a firm level policy, a review gate, or an escalation rule for AI assisted legal work. The defensible inference is that no effective gate required source verification before filing. It does not show that any specific person waved it through.
Accountability gap under warning: once opposing counsel and the court could not locate the cases, the situation should have been treated as an incident. Instead the respondents stood by the material, which the court weighed in finding subjective bad faith.
4. NIST AI RMF Mapping
The NIST AI Risk Management Framework is voluntary. These are retrospective control mappings, not a statement that NIST requirements legally applied.
Each subcategory below is scored against the AI RMF Playbook's Suggested Actions and Transparency and Documentation Questions for that specific subcategory, not against the bare subcategory title.
Status caveat. "Not Compliant" reflects an affirmative finding of failure in the record. "Insufficient Record Evidence" means the record did not address this control either way. Record-silent is not the same as non-compliant, and this table does not claim otherwise.
The verdicts at a glance, then the reasoning behind each one.
Subcategory | Status |
GOVERN 1.1 | Not Compliant |
GOVERN 1.2 | Insufficient Record Evidence |
GOVERN 1.3 | Insufficient Record Evidence |
GOVERN 2.1 | Not Compliant |
MAP 1.1 | Insufficient Record Evidence |
MAP 2.1 | Insufficient Record Evidence |
MAP 2.2 | Insufficient Record Evidence |
MAP 5.1 | Insufficient Record Evidence |
MEASURE 2.1 | Not Compliant |
MEASURE 2.3 | Not Compliant |
MANAGE 4.3 | Not Compliant |
NIST AI 600-1 §2.2 | Not scored |
GOVERN 1.1 · Not Compliant
Legal and regulatory requirements involving AI are understood, managed, and documented. The firm had no process that tracked how Rule 11's existing certification duty applied to AI-assisted research before the tool was used.
Evidence: GOVERN 1.1 turns on legal requirement review. The court answered with a Rule 11 violation and a subjective bad faith finding. The firm's own legal requirements tracking process, if any, is not on the public record.
GOVERN 1.2 · Insufficient Record Evidence
The characteristics of trustworthy AI were never built into a policy for generative AI use in legal work: no defined approved uses, prohibited uses, or review steps existed for treating a chatbot as a citator.
Evidence: Policy development is the control question. A sanctions opinion documents individual conduct and the penalty imposed. It does not state whether the Levidow Firm had an acceptable-use or review policy.
GOVERN 1.3 · Insufficient Record Evidence
No process tiered "AI assisted research feeding a federal court filing" as needing a higher level of risk management than routine drafting. The record shows no elevated scrutiny was applied before the affirmation was filed.
Evidence: Here, the Playbook asks what policies set risk tolerance and assessment scales. Nothing in the opinion describes the firm's risk tiering process, so there is no basis to score it either way.
GOVERN 2.1 · Not Compliant
A defensible gate would have assigned clear responsibility for verifying AI assisted work before filing.
Evidence: In the record, research ran through an attorney not admitted in the district and the filing went out under another attorney's signature, with no verification step between them. The court held both attorneys and the firm jointly and severally responsible. The firm's internal role documentation is not on the record.
MAP 1.1 · Insufficient Record Evidence
A federal court filing is a high stakes context. Documenting that would raise the verification bar.
Evidence: Deployment setting is clear from the filing. The record says nothing about whether the firm characterized or documented that context before the tool was used. That is what the subcategory asks.
MAP 2.1 · Insufficient Record Evidence
No one documented the task in operational terms: legal citation research. No one documented the method either. ChatGPT generates text, not authority. That distinction never reached the filing.
Evidence: Task performance is visible in the record. The Playbook asks whether the entity defined and documented that task and its limitations. No internal task definition was located either way.
MAP 2.2 · Insufficient Record Evidence
No one wrote down what the tool could not do, or who was meant to check it. That is the gap confabulation exploits.
Evidence: Output oversight did not appear in practice. The Playbook asks whether knowledge limits and the human oversight role were documented. The absence is inferred from the outcome, not from a located document.
MAP 5.1 · Insufficient Record Evidence
Estimating the magnitude of harm from false authorities to the client, court, and opposing counsel would flag the filing as high consequence.
Evidence: MAP 5.1 asks for documented likelihood and magnitude estimates. The harm was foreseeable, but the record contains no impact assessment to score against.
MEASURE 2.1 · Not Compliant
"Measurement" here means checking each citation against an authoritative reporter before filing.
Evidence: MEASURE 2.1 puts the check question directly. Fabricated citations reached a federal filing without being checked against an authoritative reporter. The separate documentation half of this subcategory, the TEVV artifacts themselves, is record-silent.
MEASURE 2.3 · Not Compliant
Output should be validated for the conditions of use. A court filing demands authoritative source confirmation, not model self report.
Evidence: No testing appears anywhere in the record. The Playbook asks what testing identified errors and limitations. The output went into its deployment setting without ever being demonstrated against the conditions of use.
MANAGE 4.3 · Not Compliant
Incidents and errors should be tracked, corrected, and communicated. The warning signs should have triggered correction and disclosure.
Evidence: MANAGE 4.3 asks whether incident tracking, response, and communication processes were followed and documented. After Avianca's counsel could not locate the cases and the court ordered production, the respondents did not withdraw them. The court weighed that decision in finding subjective bad faith.
NIST AI 600-1 §2.2 · Not scored
Confabulation means confidently false generated content. It is the technical mechanism at the center of the case.
Evidence: §2.2 defines a risk, not a control with an assessment question behind it. It is a diagnostic lens, and 600-1 is guidance rather than a legal standard applied here.
Mapping conclusion. Eleven NIST AI RMF subcategories were scored: five Not Compliant, six Insufficient Record Evidence, none Compliant or Partial. The Not Compliant findings cluster exactly where the court made findings of its own. MEASURE 2.1 and 2.3 carry the verification gap. MANAGE 4.3 carries the failure to treat the warning as an incident. GOVERN 1.1 carries the untracked legal requirement. GOVERN 2.1 carries the gatekeeping role that no one actually held.
Everything else came back Insufficient Record Evidence for one reason. A sanctions opinion documents what the attorneys did, not what the firm's governance program looked like. That is not a finding that the firm had no policies. It is the record declining to say, and the distinction matters more than a scoring table usually gets credit for.
5. Claim to Evidence Analysis
Claim family in the record | What the record establishes | What it does not establish |
Fabricated authorities were filed | Nonexistent opinions with fake quotes/citations, generated by ChatGPT, appeared in the opposition | That any client other than Mata was affected |
Respondents stood by the material | They did not immediately withdraw after the cases were questioned | Intent to defraud beyond the court's "subjective bad faith" finding |
Sanctions imposed | $5,000 jointly and severally into the Registry; letters to Mata and each falsely named judge | Any monetary award to Avianca as a sanction |
Case outcome | The claim was dismissed as time barred under the Montreal Convention | Any ruling on the underlying injury |
6. Accountability and Responsible AI Assessment
Question | Evidence backed assessment | Confidence |
Was AI use itself the violation? | No. The court said reliable AI assistance is not inherently improper. | High |
What was the violation? | Filing nonexistent authority without verification, then standing by it. | High |
Who was accountable? | The attorney of record (LoDuca), the researching attorney (Schwartz), and the firm, jointly and severally. | High |
Is the responsibility movable to the tool? | No. AI changed the workflow. It did not move the professional's duty. | High |
Does this record establish bias, privacy, or security failures? | No. Those are outside the substantiated scope of this case. | High |
7. Controls That Would Have Changed the Outcome
Approved use policy for generative AI in legal and other high trust work, naming prohibited uses (e.g., treating a chatbot as a citator).
Source verification requirement for every external citation, quotation, and figure. Each one should be checked against an authoritative reporter, not the model.
Generation versus retrieval training so staff understand that a generative tool produces text, not verified authority.
A supervisory review gate where AI materially assisted a filing.
An escalation and incident rule: when a claim is challenged, stop, verify independently, correct the record, and disclose.
Retained evidence of the verification for high stakes work product.
WHAT GOOD LOOKS LIKE
Of the six controls above, only one would have stopped this on its own: independent source verification before filing. The rest reduce the odds of reaching that point. A single verification gate, owned by a named person and evidenced afterward, is worth more than a policy no one is accountable for executing.
8. Governing Law and Cross-Framework Mapping
Governing Law
Applicable law | What it required | Status | How to use it |
Fed. R. Civ. P. 11 (certification duty) | A filing's factual and legal contentions must be verified, not merely asserted. The sanctions order found subjective bad faith and imposed a $5,000 sanction jointly and severally, plus required notice letters. | Direct / binding rule. Decided by sanctions order | Use Rule 11 as the legal authority. Use the order for what it actually decided. It is not a bar-discipline finding. |
WHERE THE LAW LANDED
Rule 11 did not need updating for AI. The certification duty already covered this: sign a filing and you certify its contentions are verified. No new obligation was created, and none was waived because a tool was involved. The court held the researching attorney, the attorney of record, and the firm jointly and severally responsible.
Cross-Framework Mapping (voluntary, non-binding)
Status caveat. "Not Compliant" reflects an affirmative finding of failure in the record. "Insufficient Record Evidence" means the record did not address this control either way. Record-silent is not the same as non-compliant, and this table does not claim otherwise.
Framework | Closest mapping | Status | How to use it |
NIST AI RMF 1.0 | GOVERN 1.1/1.2/1.3/2.1; MAP 1.1/2.1/2.2/5.1; MEASURE 2.1/2.3; MANAGE 4.3 | Scored per subcategory in Section 4: 5 Not Compliant, 6 Insufficient Record Evidence, 0 Compliant, 0 Partial | Voluntary retrospective control analysis. Not violations. GOVERN 1.1 names the legal and regulatory tracking failure directly. See Section 4 for the per-subcategory evidence behind each status. |
NIST AI 600-1 | §2.2 confabulation | Not scored | §2.2 defines a risk rather than a control with an assessment question. Names the confidently false output mechanism. |
ISO/IEC 42001:2023 | 5.2 policy; 7.2 competence; 8.1 operational control; 9.1 monitoring | Not scored | ISO 42001 sets requirements for an AI management system. The record does not establish that the firm operated one, so every clause would return the same answer and a status column would tell a reader nothing. Read the mapping as a counterfactual. These clauses name what an AIMS would have placed between the draft and the filing. |
9. The Takeaway
A generative tool can assist the work, but a human still owns the evidence. Before a load bearing claim leaves the organization, someone accountable should be able to show the check. The evidence trail needs to name what was checked, which source owned the truth, and who verified it.
The Mata record supports a narrow, durable conclusion. The court sanctioned the respondents for filing nonexistent authority generated by ChatGPT and standing by it after its existence was questioned. The court also dismissed the underlying claim as time barred. The record does not support broader claims about the tool's general reliability or about outcomes for anyone other than the parties here.
Governance principle: The tool can help produce the work. It cannot own the accountability or carry the burden of proof. A human has to verify the source before the organization relies on the output. Governance built in, not bolted on
Read the Executive Brief
This case study has a companion Executive Brief ,a six-section one-pager distilling the governance signal, risk snapshot, recommended actions, and board discussion questions for executives who need the lesson without the full analysis.
Sources
Primary Sources
Mata v. Avianca, Inc., No. 1:22-cv-01461 (S.D.N.Y.), Opinion and Order on sanctions, June 22, 2023. This is the controlling document, reported at 678 F. Supp. 3d 443.
The same day Order granting Avianca's motion to dismiss. The claim was time barred under the Montreal Convention.
NIST AI 600-1 (Generative AI Profile), §2.2 (confabulation). This is used as an analytical lens, not law.
Federal Rules of Civil Procedure, Rule 11 (https://www.law.cornell.edu/rules/frcp/rule_11). This is the governing rule behind the certification duty and the sanctions order. The opinion above is its case-specific application.
S.D.N.Y. docket, Mata v. Avianca, Inc., No. 1:22-cv-01461, ECF entries (CourtListener docket no. 63107798). Used to confirm that ECF 21 (Affirmation in Opposition, March 1, 2023) and ECF 29 (Affidavit, April 25, 2023) are two separate filings.
Frameworks Used
NIST AI RMF 1.0
NIST AI 600-1
ISO/IEC 42001:2023
Legal status note: This is a sanctions opinion and dismissal record. The attorneys and their firm were sanctioned by the court. This case study makes no professional discipline claim beyond the court's order.
Legal Disclaimer: This is a governance analysis, not legal advice. The court's findings are the court's. The NIST mapping list is not exhaustive. It shows practical ways to apply governance controls to the facts the record gives us.
Verified: Material factual claims in this case study were checked against primary source documents. Framework mappings are our analytical read, not findings by a court or regulator. Last verified August 30, 2026.
We do not rise to the level of our AI capabilities. We fall to the level of our governance. Build it in from the start. Everything else is just damage control.
© 2026 UNITI Cyber Media LLC
Comments