AI citations and client data: lawyer responsible

You must now demonstrate two separate controls before submitting AI-assisted court papers: independent verification of every citation against authoritative sources, and advance contractual assurance of how client data will be stored, retained, trained on and shared.

An analysis of 3 November 2026 of source verification and data control in AI-assisted legal work argues that lawyers remain fully responsible for the accuracy and confidentiality of documents produced with generative AI, regardless of the tool's output. The New York State Unified Court System's October 2026 guidance on AI use in court papers explicitly warns against non-existent cases and incorrect citations, requiring independent checking through authoritative legal sources and recording of verification steps. In our assessment, the emergence of separate guidance from New York, Rhode Island and the New York State Bar Association signals that regulators now treat citation verification and data protection as two distinct professional obligations, each requiring its own documented workflow and each carrying separate liability if neglected.

What does the New York guidance actually require?

The New York State Unified Court System's official publication makes clear that the signing lawyer bears full responsibility for accuracy, even when an AI tool has produced fabricated or misrepresented case law. Before submission, citations, facts and statements of law must be independently verified through authoritative sources, and errors must be corrected. The guidance requires you to record the verification steps you have taken, making the distinction between a defensible and an indefensible process concrete. The substantive check must answer four questions: does every cited ruling actually exist, is the quoted passage authentic, does it support the asserted rule of law, and are the jurisdiction and currency correct?

Why is data control a separate obligation from citation checking?

Confidentiality does not depend on the accuracy of the output, but on what actually happens to the data you enter. The Rhode Island Supreme Court's interim guidelines require you to understand the terms, privacy policy, storage, retention, training, security incidents and third-party access before use. That is a different check from citation verification and requires different documentation. A case cited in the guidance—United States v. Heppner—illustrates the distinction: consumer use of an AI service without a lawyer's involvement was not treated as protected work product, partly because the service's privacy terms allowed inputs and outputs to be retained, used for training and disclosed to third parties. In other cases the outcome was different when the service was used under appropriate contractual restrictions. It follows that privilege does not arise automatically by using an AI tool: it depends on who uses the tool, for what purpose and under which terms.

What failure modes does this guidance address?

• Fabricated citations — AI systems generating case references that do not exist or misrepresenting their holdings.
• Misquotation and misapplication — authentic cases cited out of context or applied to jurisdictions where they do not hold.
• Unauthorised training on client data — confidential information entered into a service and subsequently used to train the provider's models.
• Uncontrolled retention — client data stored indefinitely or shared with third parties without contractual restriction.
• Loss of privilege through consumer terms — use of a service whose standard terms prevent work product protection.

Which concrete controls must you be able to demonstrate?

1. Record the model and version used — document which AI tool and which version generated each court paper before submission.
2. Document your verification sources — maintain a record of which authoritative sources you consulted to verify each citation and who performed the final check.
3. Assess contractual terms in advance — before entering any client data, review and document the service provider's terms on storage, deletion, incident notification and onward disclosure.
4. Secure separate data protection agreements — obtain written contractual restrictions on how client data may be stored, retained, trained on and shared, distinct from your citation verification records.
5. Maintain separate audit trails — keep distinct records showing how each of the two checks (citation verification and data control) was performed for each document.

How should you treat the different sources of guidance?

The Rhode Island Rule 1.1 amendment was adopted by the Supreme Court and carries the force of law. The interim guidelines adopted alongside it are advisory and have no binding force. The New York State Unified Court System's publication is official guidance from the court system itself. The New York State Bar Association's ethics guidelines are a professional standard. Their legal status differs, and that determines how binding you should treat each requirement. In jurisdictions where you practise, you should establish which of these sources carries mandatory force and which is advisory.

No tool can substitute for your own professional judgement about whether a citation is sound or whether a data protection term is acceptable. Automation can record that you have performed a check and can flag inconsistencies, but it cannot verify that a case actually supports your argument or that a contractual term adequately protects your client. Both checks remain your responsibility, and both must be demonstrably recorded before use.