AI Hallucination Sanctions: Mata v. Avianca and the Candor Duty

The most expensive lesson in legal AI

Every ethics opinion on the tracker cites the same cautionary tale, because it is perfect: experienced lawyers, a routine case, a tool that wrote beautifully, and a brief built on cases that never existed. Mata v. Avianca (S.D.N.Y. 2023) produced sanctions, international coverage, and a permanent change in how courts read briefs. United States v. Cohen (S.D.N.Y. 2024) demonstrated that even sophisticated parties submit fabricated citations, and Smith v. Farwell (Massachusetts Superior Court, February 2024) confirmed state courts will sanction the same conduct. The Ropes & Gray tracker now catalogs AI-related decisions continuously; fabricated-authority incidents stopped being newsworthy and started being a docket category.

The legal mechanism is old. Rule 3.3 prohibits knowingly making a false statement of law to a tribunal and requires correcting false statements previously made. Rule 11 and its state analogues make a signature a certification that the filing’s contentions are warranted by existing law after reasonable inquiry. A fabricated citation fails both, and “the computer wrote it” has persuaded no court anywhere, for the same reason “my paralegal wrote it” never has: the signature is the lawyer’s.

Why the machine lies so well

The failure mode is structural, not a bug being patched out next quarter. Generative language models produce statistically plausible text; a legal citation is among the most patterned text formats in existence, so a model can emit a reporter cite, pinpoint page, and convincing parenthetical for a case that was never decided. Worse, the surrounding prose is fluent, which defeats the skim-review lawyers actually perform under deadline. D.C. Opinion 388 makes this the core of its competence analysis: understanding that the tool predicts rather than retrieves is the minimum technical literacy Rule 1.1 now requires. The practical corollary: confidence and formatting are not evidence of existence. The only evidence of existence is the database.

What the courts and bars have converged on

Three layers now reinforce the same workflow. The ethics layer: every formal opinion requires output verification, with New Mexico’s 2024-004 specifying that verification means traditional databases and expressly rejecting self-confirmation by the same AI. The court-policy layer: New York’s Part 161 builds verification into the signature itself through its Appendix A model rule, under which signing certifies careful review and confirmation that the paper contains no fabricated cases, statutes, or other material, with sanctions under 22 NYCRR 130-1.1; the Illinois policy makes users fully accountable for final work product in the same breath that it authorizes AI use. And the judge-by-judge layer adds standing orders requiring certification in hundreds of courtrooms, mapped in the court orders guide.

Notice what none of these layers does: prohibit the tool. The regulation targets the moment of filing, which means the entire compliance burden compresses into one controllable workflow.

The verification workflow

For every AI-assisted filing: pull every cited authority in a traditional research database; confirm each case exists, each quotation appears where cited, and each authority stands for the proposition claimed; check the assigned judge’s standing orders for AI provisions; and have the signing lawyer perform or directly supervise the check, because the certification is theirs. Steps one through three take minutes per citation; the alternative cost is measured in sanctions, malpractice exposure, bar referrals, and a permanent search-result association between your name and the word “fabricated.”

Firms should not leave this to individual virtue. Write it as a non-negotiable rule in the firm AI policy, train it, and audit it; section 4 of the free policy template is that rule, ready to adopt.