AI Hallucinations in Court: Sanctions, ABA Op. 512, and Michigan (2026)

By Attorney Manny Chahal · Updated May 2026 · Reading time: ~9 min

In May 2023 a Manhattan federal judge fined two New York lawyers $5,000 for filing a brief stuffed with cases that did not exist. Three years later, the same mistake has cost lawyers in at least 128 reported United States proceedings, drawn formal guidance from the American Bar Association, and triggered local-rule proposals in the Eastern District of Michigan. This is the state of AI hallucinations in court — and what it means for Michigan lawyers, their clients, and anyone whose personal-injury case may be shaped by a tool that confidently invents authority.

The Founding Case: Mata v. Avianca (2023)

Roberto Mata sued Avianca Airlines for a personal-injury claim arising from a service-cart incident. When the airline moved to dismiss on limitations grounds, plaintiff's counsel filed an opposition brief citing six federal appellate decisions: Varghese v. China Southern Airlines, Shaboon v. Egypt Air, Petersen v. Iran Air, Martinez v. Delta Airlines, Estate of Durden v. KLM, and Miller v. United Airlines. None of those cases existed. Each was fabricated by ChatGPT, complete with fictitious quotations and attributed to real federal judges.

In Mata v. Avianca, Inc., 678 F. Supp. 3d 443 (S.D.N.Y. 2023), Judge P. Kevin Castel imposed $5,000 in Rule 11 sanctions on the attorneys, finding subjective bad faith in their failure to verify the citations before filing and their later attempts to explain the mistake away. The sanctions order also required counsel to send personally signed letters to every real judge whose name had been falsely attached to a fake opinion.

ABA Formal Opinion 512: The Six-Duty Framework (July 2024)

On July 29, 2024, the ABA Standing Committee on Ethics and Professional Responsibility issued Formal Opinion 512: Generative Artificial Intelligence Tools. The opinion maps six existing Model Rules onto every lawyer's use of generative AI. Michigan has adopted Model Rules language closely; the same analysis flows through the Michigan Rules of Professional Conduct.

Competence — MRPC 1.1

Lawyers do not need to become AI experts, but they must have a reasonable understanding of the capabilities and limitations of the specific tool they use. The duty is ongoing: a one-time training in 2024 does not satisfy MRPC 1.1 in 2026 when the model has been updated five times since.

Confidentiality — MRPC 1.6

Pasting client facts, deposition transcripts, or medical records into a public generative-AI tool may waive privilege and breach confidentiality. Enterprise-grade tools with no-training and no-retention contractual protections are not a complete answer; the lawyer must understand the data flow before any client information goes in.

Communication and Fees — MRPC 1.4 and 1.5

Clients should know when generative AI is doing material work on their matter. Fee bills cannot pass through AI subscription costs as a separate disbursement absent clear disclosure, and time saved by AI cannot be billed as if the lawyer performed the task by hand.

Candor to the Tribunal — MRPC 3.3

This is the rule Mata implicates most directly. A lawyer who files a brief without independently verifying every citation violates the duty of candor. Hallucinated authority is a false statement of law within the meaning of the rule, regardless of the lawyer's subjective belief.

Supervision — MRPC 5.1 and 5.3

Managing partners must establish written policies on permissible AI use and train both lawyers and nonlawyers. A firm cannot delegate AI compliance entirely to a junior associate; the duty runs from the top.

The 2025–2026 Sanctions Wave

The volume of hallucination-driven sanctions orders accelerated through 2025 and into 2026. Public trackers now list more than 480 documented matters worldwide and more than 320 in United States federal, state, and tribal courts, with at least 128 individual lawyers personally identified and two judges whose chambers staff used AI without disclosure.

Representative orders include:

• D. Ariz., August 2025. Judge Alison Bachus sanctioned counsel after finding 12 of 19 cited cases were fabricated, misleading, or unsupported, with patterns consistent with generative-AI hallucination.
• N.D. Ala., July 2025. In Johnson v. Dunn, the court declined to sanction a defending firm despite its counsel filing a brief with fabricated authority, citing the firm's written AI policy and demonstrated supervision as mitigating factors. This is the rare case where supervision discipline saved the lawyer.
• E.D. Pa., April 2026. Judge Kai N. Scott sanctioned attorney Raja Rajan $5,000 after a prior $2,500 sanction for the same conduct, with an express warning that a third occurrence would draw a referral to the Pennsylvania disciplinary board.
• D.N.J., April 2026. A Cherry Hill attorney was sanctioned a second time for filing a brief containing AI-hallucinated citations.

Michigan's Framework

Eastern District of Michigan Proposed Local Rule 5.1(a)(4)

On December 8, 2023, the judges of the United States District Court for the Eastern District of Michigan published a Notice of Proposed Amendments that included a new Local Rule 5.1(a)(4), “Disclosing Use of Artificial Intelligence.” The proposed rule reads, in operative part, that if generative AI is used to compose or draft any paper presented for filing, the filer must disclose its use and attest that citations of authority have been verified by a human being. Practitioners filing in the Eastern District should treat the proposal as the floor of expected practice until the final rule is adopted.

State Bar of Michigan Guidance

The State Bar of Michigan's Ethics Department has published an “Artificial Intelligence for Attorneys — Frequently Asked Questions” resource that anchors the analysis in MRPC 1.1. The State Bar's June 2025 “Age of AI” report extends that guidance: a Michigan lawyer's duty of competence includes a working knowledge of the technology, an obligation to verify every AI output before it goes to a client, opposing counsel, or a court, and an affirmative duty to update that knowledge as the tools change.

Work-Product Protection for AI-Assisted Drafts

Michigan federal courts have already addressed discovery exposure for AI-assisted attorney work. A 2025 Eastern District ruling extended traditional work-product immunity to attorney prompts, drafts, and notes generated with the assistance of generative AI, provided the materials were prepared in anticipation of litigation. Lawyers should still segregate AI prompt logs from substantive case files and be prepared to log AI-assisted documents on a privilege schedule.

Notable AI Sanctions Cases at a Glance

What This Means for Injury Clients

The cautionary tale is not abstract. Mata v. Avianca was itself a personal-injury case. The plaintiff lost his suit not because his factual claim was weak but because his lawyers filed an unverified AI brief. Clients should expect their counsel to use technology, including generative AI, where it speeds work without sacrificing accuracy — document review, deposition summaries, timeline construction. Clients should not expect, and should not pay for, a brief that no human cross-checked before filing.

Three concrete client questions to ask any lawyer in 2026:

• Do you use generative AI on my file? If yes, which tool, and how do you protect my information?
• What is your verification policy? A defensible answer references both the citation-checking workflow and the supervisory lawyer who signs off before filing.
• Do you disclose AI use to the court? In jurisdictions with standing orders, the answer must be yes — in writing, on the face of the filing.

Practical Checklist for Michigan Practitioners

1. Written firm policy. Identify approved tools, prohibited inputs (PHI, settlement figures, attorney-client privileged material), and the citation-verification workflow.
2. Citation-verification step before filing. Every cite the AI produced gets pulled from Westlaw, Lexis, or PACER and compared to the AI output before signature.
3. Disclosure habits. In every federal forum, check the assigned judge's standing orders for AI-disclosure language before filing.
4. Training cadence. Annual MRPC 1.1 training that includes the current generation of AI tools, not a one-and-done module from 2023.
5. Privilege segregation. AI prompt logs are work product when prepared for litigation; treat them like any other privileged work file.
6. Client communication. Engagement letters that affirmatively address AI use and verification responsibility.

Frequently Asked Questions