New York's statewide court system has adopted a layered set of policies governing how judges and litigators may use generative AI for legal research, writing, and court filings, balancing efficiency gains against well-documented risks of fabrications and ethical breaches. The policies, outlined in a detailed analysis by Manhattan Supreme Court Justice Gerald Lebovits, establish mandatory training for judges, restrict which tools they can use, and set a baseline for attorney submissions without requiring blanket disclosure of AI use.
More than 60% of federal judges surveyed in December 2025 by Northwestern University reported using at least one AI tool in their judicial work, most often for legal research (30%) and document review (15.5%). Justice Lebovits' article, which synthesizes policy guidance from New York's Unified Court System, describes how generative AI can handle discrete tasks such as generating Boolean searches, identifying leading cases, and organizing arguments under the IRAC format. But it is best used for discrete tasks, not as a substitute for independent judgment, the piece states.
Ethical boundaries and red flags
The article identifies recurring signs that a legal document may have been produced with minimal human oversight: overuse of em dashes, formulaic parallel phrasing, absence of typical human errors like typos, and reliance on cliché transitions. More substantive red flags include hollow reasoning that summarizes both sides without genuine analytical friction, invented authorities, and jurisdictional slippage where a tool treats federal or another state's rule as New York law.
Justice Lebovits writes that a 2024 study of public-facing models found legal hallucinations in at least 58% of tested responses overall, and at least 63% for questions about a case's central holding. "Generative AI's greatest weaknesses are its inability to authenticate information, reason reliably, and exercise judgment," he writes.
New York's two-track policy
The New York State Unified Court System adopted an Interim Policy on the Use of Artificial Intelligence in October 2025. It applies to judges and staff, requires formal training before use, and mandates that only tools approved by the Division of Technology and Court Research can be used on UCS devices. Microsoft 365 Copilot Chat is currently approved; ChatGPT has been removed and blocked on court-owned devices.
For attorneys, the separate 22 N.Y.C.R.R. Part 161 took effect June 1, 2026. It sets a statewide baseline without requiring attorneys to disclose AI use in filings. "Existing duties already require accuracy," Lebovits writes. He notes that certifying a filing already constitutes a statement of non-frivolity after reasonable inquiry, making a separate AI disclosure requirement unnecessary in his view. Several individual judges in Brooklyn, Manhattan, Queens, and Staten Island have adopted part-specific rules requiring varying levels of disclosure and certification.
Auditing AI-generated work
The article lays out concrete procedures for verifying AI-assisted content. Legal professionals should confirm all cited material, ensure seminal cases are included, check internal consistency, and revise until the output sounds like the author's prior work. "Confirm all cited material, including givens, such as legal standards," Lebovits writes.
The piece also addresses billing implications. If generative AI produces a document in one hour that an attorney would need three hours to produce, the rule generally allows billing only the time actually spent "including reasonable time reviewing and revising the output," not the time the task would have taken without AI, the piece states.
Why this matters for legal professionals
For New York judges and litigators, the article's most practical takeaway is that the court system expects all drafts to receive human editorial scrutiny. Sanctions remain a real possibility. A Connecticut Supreme Court case decided in July 2026 ordered an attorney to complete additional CLE ethics training, donate $1,000 to the bar as well as from his firm, and comply with filing deadlines after hallucinated citations appeared in two separate cases. The judge who wrote the decision named the attorney 23 times in the text of the order.
Your membership also unlocks:
