June 1, 2026 (in 3 days): New York: 22 NYCRR Part 161 takes effect, system-wide AI policy for all UCS courts

Kaur v. Desso

U.S. District Court, Northern District of New York · N.D.N.Y. · New York bar guidance

Court sanction

Verified April 26, 2026

Citation
Kaur v. Desso, No. 9:25-cv-00726-AMN (N.D.N.Y. July 9, 2025) (Dkt. No. 24)
Decided
July 9, 2025

Summary

Petitioner's counsel Dennis Desmarais used Claude Sonnet 4 to draft a supplemental brief in an emergency habeas corpus matter and submitted multiple fabricated quotations attributed to DHS v. Thuraissigiam, Landon v. Plasencia, Reno v. American-Arab Anti-Discrimination Committee, and Wong Wing v. United States. Even after the Government's responsive brief flagged the fabrications, Desmarais took no remedial action until the court raised the issue at a hearing. Judge Anne M. Nardacci found subjective bad faith based on conscious avoidance and imposed Rule 11 sanctions sua sponte.

AI tool:
Claude Sonnet 4
Sanction amount:
$1,000
This case summary is informational only. Verify the underlying opinion or order against the primary source before relying on it in any filing or client matter.

What sanction did the court impose?

$1,000 monetary penalty payable to the Clerk of Court within fourteen days, plus required completion of a CLE-credited program on the use of artificial intelligence in legal writing and proof of service of the order on the petitioner, both due by September 1, 2025.

Why does Kaur v. Desso matter for law firms using AI?

Kaur v. Desso is the first published sanctions order in this tracker to name Anthropic’s Claude (specifically Claude Sonnet 4) as the generative AI tool behind the fabricated authority, a notable data point for firms that have standardized on Claude under the assumption that more recent frontier models meaningfully reduce hallucination risk in legal research. The court’s reasoning is also instructive for managing partners: Judge Nardacci treated fabricated quotations from real cases as functionally identical to fabricated cases, and found “conscious avoidance” sufficient for subjective bad faith once an attorney knows generative AI can hallucinate and files anyway without verification. Mitigating factors (illness, expedited timeline, prompt confession, voluntary CLE enrollment) reduced the fine below the $1,500-$15,000 range the court identified as typical, but did not avoid sanctions, a CLE requirement, or a published opinion naming the attorney.

Sources

Primary sources

Further reading

Source PDF is a Westlaw printout mirrored from the Damien Charlotin hallucination database. We are working to add the underlying court docket (PACER, CourtListener, or court website) as a second source.