VERITAS
2026-06-28 · The audit log

Landberg v City of New York: The Second Department's Sanction for AI-Hallucinated Citations

In late June 2026, the Appellate Division of the New York State Supreme Court, Second Department, directed an attorney to pay $8,000 and his firm to pay $2,500 to the Lawyers' Fund for Client Protection of the State of New York. The total was $10,500. By the reporting that followed, it was the second sanctions order against a lawyer in a US appellate court for AI-related citation problems in about three weeks. The first was the Ninth Circuit's order on June 3 in Lnu v. Blanche.

This is not a story about a careless lawyer. It is a story about a checkpoint that did not run. A draft was produced with a tool. The draft was filed. The step that was missing sat between those two events, and it is a step that resolves a narrow factual question before a brief is signed.

What the court described

The court's account is specific. The appellate brief, in the court's words, was prepared with the assistance of generative artificial intelligence and contained citations to nonexistent cases, purported Court of Appeals quotations contrary to actual law, and misrepresentations of what certain real cases held or decided. The matter began when the court, on its own motion, directed counsel to show cause under 22 NYCRR 130-1.1 why sanctions should not be imposed.

At oral argument in May, the panel raised the citations and quotations directly. Counsel later submitted an affirmation in which he took full responsibility for the deficiencies the court identified, and stated that when the court asked where the cited cases came from, he should have said they had been generated by artificial intelligence. The court directed the payments to the Lawyers' Fund for Client Protection and noted that whether further disciplinary action is warranted is a matter for the Attorney Grievance Committee.

We will not characterize the citations in our own words beyond what the court found. Veritas does not call a citation false. A court describes what it located in the record, and this one did. Our interest is narrower and more mechanical.

The two questions inside a citation

There are two questions buried in a citation, and they are not the same question. The first is mechanical. Does this case exist in the reporter, and does the quoted language actually appear at the cited page? That question has a yes-or-no answer. It does not require judgment. It requires looking. A citation either resolves to a real opinion in the reporter or it does not. A quotation either appears in that opinion or it does not.

The second question is legal. Assuming the case is real and the quote is accurate, does the case actually support the proposition the brief cites it for? That question requires a lawyer. It requires reading the holding, understanding the posture, and judging whether the authority does the work the argument needs. No tool replaces that, and the harder failure mode of proposition support is its own problem. The order in this matter describes failures in the first category. Citations that could not be located. Quotations that did not match the law. The first category is exactly the part that does not need a lawyer's judgment to catch. It needs a comparison against the reporter before the brief is signed.

Why the drafting tool does not matter

The detail that travels in coverage of cases like this is that the lawyer did not, at first, name the tool. For the verification question, that detail does not change anything.

A citation check runs against the court record, not against the drafting process. It does not ask how the text was produced. It asks whether each cited case exists and whether each quoted passage appears where the brief says it does. The answer is the same whether a person typed the citation from memory, a paralegal transcribed it, or a model generated it. The reporter is the fixed reference point. Everything is measured against it. This is the same place the federal sanctions orders have located the duty of inquiry, stated in New York's sanctions rule rather than in Rule 11.

What this does not change for lawyers

It would be wrong to read this as an argument against AI-assisted drafting, and it would be wrong to read it as fearmongering. Courts across several jurisdictions have now said the same thing in different words. AI changes the workflow, not the duty. The duty to confirm that the authorities in a signed filing exist and are cited accurately belongs to the lawyer whose name is on the brief, and it does not move.

What a mechanical check does is take the most catchable category of error off the lawyer's plate before it reaches the bench. It confirms existence. It confirms the quote. It leaves the legal judgment, whether the case supports the argument, where it belongs. The hard part of lawyering stays with the lawyer. The part that failed first in this matter was the part a check resolves in seconds.

The record before the filing

The representation is owed at the moment of signing. The question a later reviewer asks is not whether the citations could have been checked. It is whether they were, and what the check found. A Verification Certificate is the contemporaneous record of that check: each citation, its verification status, the sources consulted, a content hash, and a timestamp. It is filed in the matter file alongside the brief. When an order to show cause arrives, the record exists outside the memory of the people who produced the filing.

Landberg v City of New York now sits in the AI Citation Sanctions Tracker alongside the other matters on record. The common thread across them is not a machine inventing law on its own. It is a verification step skipped between drafting and filing.

Run a Filing Risk Scan

Filed under · Anatomy of a hallucination · Sanctions