California Wants the Disclosure. New York Wants the Signature.

The same risk produced opposite rules

California enacted Senate Bill 574 on September 30. The law adds Business and Professions Code section 6068.1, which prohibits a lawyer from delegating the practice of law to generative AI. It also restricts the entry of confidential, identifying, and other nonpublic information into systems where access is not limited to authorized people; requires reasonable steps to verify outputs and correct erroneous or hallucinated material; and requires disclosure of generative AI use to the court for documents submitted. The citation rule is broader still: the attorney responsible for a court paper must personally verify every citation, whether AI supplied it or not.

New York's Part 161 begins from the same premise about responsibility and reaches a different procedural result. Its statewide policy says AI use in preparing court papers should not be prohibited, but attorneys and parties should not have to disclose that use merely because AI was involved. The model rule relies on existing signature and sanctions rules. A filer who signs a paper certifies that it contains no materially false factual statement or frivolous legal argument. If AI helped prepare it, the filer must carefully review the paper and independently ensure that it contains no fabricated authority or other fictitious material.

This is not a contest between a strict state and a permissive one. Neither rule tolerates fictional cases, undisclosed breaches of confidentiality, or the surrender of professional judgment. California treats disclosure as an accountability mechanism. New York worries that a tool-specific disclosure adds little because the filing obligation already exists. One regime asks the court to know that AI participated. The other asks the signer to own the result without making the software part of the record.

California's disclosure sentence is doing too much work

The difficult California language is short: a lawyer using generative AI to assist in legal practice must disclose its use to the court for all documents submitted. That sentence does not say what the disclosure must contain, where it belongs, whether it applies when AI assisted only with research or editing, or how much causal connection is required between the tool and the filed text. The California Lawyers Association noted the ambiguity in an October 27 analysis and predicted that courts or the State Bar may need to supply guidance.

A literal reading can become strange quickly. Modern legal platforms increasingly place generative features inside research, drafting, translation, document review, and ordinary office software. If a lawyer uses an AI summary to locate a case, rejects the summary, reads the opinion, and drafts the brief independently, was the tool used for the document? What if it suggests a heading, checks tone, or generates a chronology that never appears in the filing? A rule aimed at accountable advocacy needs a limiting principle, or disclosure may become both ubiquitous and uninformative.

The statute's other requirements are easier to connect to legal risk. A confidentiality rule governs what enters the system. A verification rule governs what leaves it. A nondelegation rule preserves the lawyer's judgment. Disclosure is different because it creates a communication duty to the tribunal even when the output is accurate and the client information protected. That may deter careless use, but it may also encourage defensive overdisclosure that tells the judge nothing about what the lawyer actually did.

The right implementation therefore cannot be a generic footer added to every filing. Until authoritative guidance arrives, California practitioners should identify the tool, the task it performed, the parts of the work it affected, the review performed, and the lawyer responsible. Whether every item belongs in the filed disclosure is a legal question still open. Each item belongs in the firm's internal record now.

New York puts the emphasis on authorship

New York's approach is appealing because it refuses to confuse provenance with truth. A lawyer can draft a false statement without AI, and a machine-assisted draft can be correct after disciplined review. Part 161 makes the signature carry the same meaning either way. The attorney is expected to understand the tool's limits, locate and read the cited material, and independently ensure that the final paper contains no fabrication. The court can sanction the failure without litigating which keystroke came from which system.

But New York has not created a no-disclosure zone. Part 161 permits individual courts to adopt a model rule, and judges retain authority over proceedings before them. The statewide policy rejects a blanket requirement; it does not erase a specific standing order or a judge's case-management direction. A lawyer who remembers the general rule but never checks the assigned judge's current requirements can still be wrong in the only courtroom that matters that day.

The deeper strength of New York's rule is its insistence that verification means more than confirming that a citation exists. The filer must independently ensure the paper contains no fictitious content or frivolous argument. A real opinion can be quoted inaccurately, applied outside its holding, or cited for a proposition it does not support. That is why the Journal's recent analysis of Florida's so-called AI slop focused on judgment rather than imaginary citations. Existence checking is clerical. Legal verification requires reading, context, and a decision about relevance.

The weakness is evidentiary. If a problem emerges, a rule that depends on independent review works best when the firm can show what review occurred. New York may not demand an AI disclosure on the docket, but a prudent firm still needs a traceable internal process. The absence of a public label is not permission to make the workflow invisible inside the office.

The compliance problem is no longer the hallucination

Firms should stop treating lawyer-AI policy as a warning not to cite fake cases. That was the first obvious failure mode, not the complete governance problem. California now regulates inputs, outputs, delegation, personal citation review, and disclosure. New York regulates competence, independent review, signature responsibility, and any additional rule adopted by the particular court. A single instruction to check the citations does not answer either jurisdiction.

The operational answer is a matter-level map of AI use. Firms need to know which approved systems were used, whether nonpublic information entered them, which tasks they performed, whether generated material reached the final work, who verified authorities and factual assertions, and which court-specific disclosure rule applied on the filing date. The record should be proportionate. It need not preserve every prompt forever. It must be good enough to reconstruct responsibility when a client, court, insurer, or disciplinary authority asks a precise question.

That map also helps firms avoid two bad extremes. One is secrecy, where lawyers use embedded AI features without recognizing them and cannot answer whether disclosure was required. The other is ritual, where every paper receives the same AI disclaimer and no one can explain what was reviewed. A useful control follows the work. It distinguishes search from synthesis, editing from authorship, a private enterprise system from a public tool, and a suggestion that was rejected from text that reached the court.

California and New York agree on the important point: the lawyer remains responsible. Their split reveals the next issue the profession has to solve. Responsibility is no longer just a statement in an ethics opinion. It must be visible in the way legal work is produced. California will make part of that visibility public. New York will leave more of it behind the signature. Either way, the firm needs to know what happened before the lawyer signed.

Disclosure and verification are different controls. A serious AI policy needs both questions answered.