AI Hallucinations in Legal Practice
Five named attorney sanctions in 30 months — from a $5,000 fine in Mata v. Avianca to a ~$60,000 sanction against an AmLaw 100 firm in December 2025.
The Founding Case — Mata v. Avianca, Inc.
On June 22, 2023, Judge P. Kevin Castel of the Southern District of New York sanctioned attorneys Steven A. Schwartz and Peter LoDuca of Levidow, Levidow & Oberman $5,000 jointly with their firm under Rule 11.
Schwartz submitted a brief in a personal-injury case citing six entirely fictitious decisions that ChatGPT had fabricated — and stood by the fake opinions when challenged by opposing counsel and the court. Mata v. Avianca, Inc., 678 F. Supp. 3d 443 (S.D.N.Y. 2023) became the inflection point — the moment AI hallucination in legal practice stopped being theoretical risk and started being a sanctioning offense.
The Case Pattern Since Mata
Federal and state, trial and appellate — every forum has now sanctioned an AI hallucination
People v. Crabill (Colorado)
Colorado Springs attorney Zachariah Crabill suspended for one year and one day (90 days to serve, remainder stayed on 2-year probation) for filing a motion citing ChatGPT-fabricated cases and initially blaming a legal intern.
Why it matters: State-bar discipline, not just monetary sanction. Career consequence.
Park v. Kim (2d Cir.)
The Second Circuit referred attorney Jae S. Lee to its grievance panel after she cited "Bourguignon v. Coordinated Behavioral Health Services," a nonexistent case ChatGPT generated, in a reply brief.
Why it matters: Federal appellate referral. The Mata pattern travels up the federal hierarchy.
United States v. Cohen (S.D.N.Y.)
Michael Cohen passed three Google Bard-hallucinated citations to attorney David M. Schwartz, who filed them in a supervised-release motion. Judge Furman declined to sanction (attorney had not directly used AI), but public exposure was severe.
Why it matters: High-profile defendant-driven hallucination — the firm cannot assume AI use is only on the attorney side.
Wadsworth v. Walmart (D. Wyo.)
Judge Kelly Rankin sanctioned three Morgan & Morgan / Goody Law Group attorneys ($3,000 / $1,000 / $1,000) and revoked Ayala's pro hac vice after a motion in limine cited eight ChatGPT-hallucinated cases.
Why it matters: Supervising attorneys sanctioned alongside the filer. Rule 5.1 failure case as much as Rule 3.3.
Goldberg Segalla / CHA
A federal judge sanctioned the No. 42 AmLaw firm by headcount and one of its attorneys roughly $60,000 over ChatGPT-generated fake citations in a lead-paint case.
Why it matters: Largest hallucination sanction to date; pattern has accelerated into BigLaw.
Discipline Beyond Money
Suspension (Crabill), pro hac vice revocation (Wadsworth), grievance referral (Park v. Kim). Career consequences, not just billable losses.
Why it matters: The cost is not the fine; the cost is the bar status and the firm reputation.
Why Hallucinations Happen — and Why RAG Does Not Fix Them
Large language models generate text by predicting plausible-sounding sequences. They do not have a concept of "this case exists" or "this holding is real." If a query implies a case probably exists with a particular name and holding, the model produces one — citation, court, year, page numbers, the works. The generated case looks real because everything about its surface form is real-looking. Only the case is not.
RAG-based legal AI (Westlaw AI, Lexis+ AI, Harvey) materially reduces this rate by grounding answers in retrieved sources, but does not eliminate it. Stanford's peer-reviewed study found 17–33% hallucination rates even on legal-specific RAG tools. See /legal-services/ai-legal-research/.
Why "I Did Not Know It Would Do That" Does Not Work
Every published sanction order has rejected the no-knowledge defense
Hallucinations have been publicly documented since mid-2023
The phenomenon is now a matter of professional knowledge. Courts treat ignorance of hallucinations as evidence of inadequate Rule 1.1 competence, not as a defense.
What courts say:
Every sanction order since Mata recites the public history.
What it means:
Lack of awareness compounds the violation rather than mitigating it.
ABA Op. 512 (July 2024) put the profession on formal notice
After Op. 512, every attorney using GenAI is on notice that competence (Rule 1.1) and candor (Rule 3.3) duties apply.
What courts say:
Post-Op. 512 sanction orders cite the opinion directly.
What it means:
Op. 512 marks the bright line between "novel issue" and "settled duty."
Every state bar AI opinion warns about it
California, Florida, NYC Bar, DC, Texas, and the ABA have all flagged hallucinations. The warning is uniform; the duty is unambiguous.
What courts say:
State opinions cite hallucination cases as the operating risk.
What it means:
Attorneys in every state operate under explicit notice.
Citation verification has been a fundamental Rule 3.3 duty since long before AI existed
The duty to verify factual and legal assertions in court filings predates AI by decades. AI did not create the duty; it just produced a new way to violate it.
What courts say:
Pre-AI sanctions for inaccurate citation are common in the case law.
What it means:
AI hallucinations are old violations in new clothing.
The Two-Layer Defense That Works
Tool-level controls + workflow controls. Both layers are required
Use enterprise-grade legal research AI
Not general-purpose ChatGPT, for any research that will end up in a brief. Westlaw AI, Lexis+ AI, Harvey, vLex.
Configure source-grounded answers
Surface source documents alongside generated answers. Set firm-default settings to maximum verification mode where the tool supports it.
Verify every citation in every brief
Pull the case. Read the holding. Confirm the proposition the AI attributed to the case is in the case.
Document verification under Rule 5.1
Citation verification recorded in the matter file. Supervising attorney sign-off on AI-assisted brief work.
Random-sample audits
Sample completed briefs for citation accuracy. Use the audit signal to refine training and process.
Mandatory annual Rule 1.1 training
Hallucination phenomenon, verification protocols, the published case pattern, and tool-specific failure modes.
Incident response plan
For when a hallucinated citation makes it into a filing — opposing counsel notification, court notification, withdrawal motion where needed.
AI Hallucinations in Legal Practice — FAQ
What is an AI hallucination in legal practice?
An AI hallucination is when a generative AI tool produces text that looks like a real legal citation, holding, or quote but is fabricated — the case, court, or holding does not exist or does not say what the AI claims. The phenomenon has been publicly documented since at least mid-2023 and is the subject of multiple sanctioning orders.
What is Mata v. Avianca?
Mata v. Avianca, Inc. (678 F. Supp. 3d 443, S.D.N.Y. 2023) is the founding hallucination case. On June 22, 2023, Judge P. Kevin Castel sanctioned attorneys Steven A. Schwartz and Peter LoDuca of Levidow, Levidow & Oberman $5,000 jointly with their firm under Rule 11, after Schwartz filed a brief citing six entirely fictitious cases ChatGPT had fabricated.
Have attorneys been sanctioned for AI hallucinations since Mata?
Yes — repeatedly. Confirmed cases include People v. Crabill (Colorado, state-bar suspension), Park v. Kim (2d Cir., grievance referral), Wadsworth v. Walmart (D. Wyo., $5,000 across three attorneys plus pro hac vice revocation), and the Goldberg Segalla / Chicago Housing Authority case (~$60,000 sanction against an AmLaw 100 firm in December 2025). The case pattern has accelerated, not subsided.
Do not tools like Westlaw AI and Lexis+ AI eliminate hallucinations?
No. Stanford RegLab's peer-reviewed 2024 study found Lexis+ AI and Westlaw AI-Assisted Research each hallucinated between 17% and 33% of the time on a benchmark of legal queries — despite vendor 'hallucination-free' marketing claims. RAG-based legal AI is a meaningful improvement over general-purpose LLMs but is not a substitute for citation verification.
What is the standard workflow to prevent AI hallucinations from reaching filings?
Two-layer defense — use enterprise-grade legal research AI (not general ChatGPT), then verify every citation before filing. Pull the case, read the holding, confirm the proposition the AI attributed to the case is actually in the case. Document the verification under Rule 5.1 supervision and Rule 1.1 competence. Supervising attorneys must sign off on AI-assisted brief work.
Related Resources
Continue across the silo or bridge to a core hub
AI for Legal Research
Stanford's 17–33% hallucination finding on leading legal RAG tools
Read article →ABA Formal Opinion 512
Rule 3.3 candor and Rule 5.1 / 5.3 supervision in sanctioning orders
Read article →Attorney-Client Privilege and AI
The Rule 1.6 layer that compounds the hallucination problem
Read article →Shadow AI Hub
Where unsanctioned LLM use creates the hallucination risk in the first place
Read article →AI Observability and Compliance
The audit-log layer that supports Rule 5.1 verification documentation
Read article →Operationalize Hallucination Prevention Across Your Firm
Free Shadow AI Risk Check audits your verification workflow, your supervising-attorney documentation, and your incident-response plan.