ChatGPT Was Not the Witness

A New York court has drawn one of the more useful early lines in the law of generative AI and litigation. In Assini v. Hayward, Justice Rhonda E. Fischer of the Nassau County Supreme Court quashed a subpoena directed to OpenAI that sought a pro se defendant’s ChatGPT materials. The request covered prompts, inputs, uploaded documents, outputs, drafts, legal research, account materials, and materials tied to filings or communications in the case.

The ruling is narrow, but it is not trivial. It does not bless careless AI use in litigation. It does not convert ChatGPT into a lawyer. It does not say that every prompt is privileged. What it says is more precise: where a litigant uses AI to prepare for litigation, the opposing party does not automatically get to examine that process merely by calling it relevant.

For lawyers, that is the useful part of the decision. The court treated AI-assisted legal preparation as part of the litigation workspace. That workspace may be digital, outsourced to a platform, and imperfectly controlled, but the work-product question still begins with the function of the material. If the record reveals research, drafting, strategy, mental impressions, or litigation preparation, ordinary discovery relevance is not enough.

The Subpoena Sought the Litigation Workspace

The plaintiffs wanted OpenAI records associated with defendant John Recchio’s use of ChatGPT. According to the court’s decision, the amended subpoena sought prompts, inputs, uploaded materials, and corresponding outputs used to draft, revise, or generate filings, motions, sworn statements, or communications in the action. It also sought prompts, queries, uploaded documents, and outputs referencing the plaintiffs, Alpha Tech Lending LLC, claims or defenses in the case, and filings or correspondence generated for use in the litigation.

That is not a request for a finished filing. It is a request for the machinery behind the filing. In conventional litigation terms, the subpoena reached toward drafts, research notes, internal reasoning, and preparation materials. The fact that those materials were generated through a consumer AI platform did not make them less revealing.

The plaintiffs argued that the materials were relevant to the basis, accuracy, and authenticity of Recchio’s claims and defenses. That argument has intuitive force. If a party relies on AI-generated text, an adversary may want to know whether the filing was grounded in fact, whether the assertions were invented, and whether the litigant used the tool to construct a misleading position.

Justice Fischer did not treat that concern as enough. The court quashed the subpoena and denied the plaintiffs’ request to have it so-ordered. OpenAI was not required to produce the records absent a further court order.

Work Product Survived the Tool

The holding is best understood as a work-product ruling rather than an attorney-client privilege ruling. That distinction is essential in practice. ChatGPT is not counsel, and communications with a chatbot are not legal advice from a lawyer. But work-product protection is not limited to conversations with lawyers. It protects materials prepared in anticipation of litigation, including materials that reveal strategy, preparation, and mental impressions.

That framing allowed the court to avoid the weakest version of the AI-privilege debate. The question was not whether ChatGPT should be treated as an attorney. The question was whether the subpoena would expose litigation-preparation materials that discovery rules ordinarily protect from an adversary.

The court found the reasoning in Morgan v. V2X persuasive. It also addressed United States v. Heppner, where a federal court in the Southern District of New York had reached a different conclusion in a criminal context involving AI communications. Justice Fischer distinguished that case and proceeded within the civil-discovery frame before her.

The decision therefore adds to an early split in how courts analyze AI use. Some judges are wary that entering information into a public AI platform destroys confidentiality. Others are more willing to treat AI as a tool used in the preparation process, especially where the user is a self-represented litigant preparing claims, defenses, and filings.

The lesson is not that prompts are safe. The lesson is that courts are beginning to separate careless disclosure from protected preparation. That separation will likely become one of the central issues in AI-related discovery.

The Ruling Is Not a Shield for Bad AI Filings

This case should not be confused with the line of cases involving fake citations, hallucinated authorities, or sanctions for unverified AI-generated filings. Those cases concern the duty of candor, competence, and verification. A court does not need to inspect a litigant’s entire prompt history to sanction a filing that contains nonexistent cases or false representations.

Assini concerns a different problem. The plaintiffs sought access to the AI records behind the opposing party’s litigation work. The court had to decide whether the adversary could reach into that process through a subpoena to the AI provider.

That is why the ruling is legally elegant. It does not excuse bad outputs. It protects the preparation layer from indiscriminate discovery. A lawyer who files AI-generated nonsense may still face sanctions. A party who uses AI to commit fraud may still face discovery, evidentiary consequences, or worse. But an adversary cannot simply say “AI was used” and treat the prompt history as open terrain.

The court also made clear that AI use remains subject to court rules. Justice Fischer specifically referred Recchio to 22 NYCRR Part 161, New York’s rule on the use of artificial intelligence technology, and warned that failure to comply may result in sanctions. The court protected the subpoenaed materials while also signaling that AI-assisted litigation cannot proceed outside procedural discipline.

Why Prompts Can Reveal Legal Strategy

Lawyers should resist the temptation to describe prompts as mere search queries. A prompt can be a legal theory in rough form. It can reveal which facts the user believes are important, which arguments the user is testing, which weaknesses the user is trying to solve, and which narratives the user considered before choosing the final version.

Outputs can be equally revealing. They may show alternative theories, abandoned arguments, draft language, issue framing, credibility concerns, evidentiary gaps, or strategic uncertainty. In a litigation setting, the exchange between user and model often looks less like a database query and more like iterative drafting.

That is why this category of material does not fit neatly into older assumptions about digital records. A browser search history can be revealing, but prompt history can be more direct. It may contain the user’s reasoning, the user’s factual assumptions, and the user’s attempted legal architecture in one continuous record.

For law firms, that creates a governance problem. If AI tools become part of legal research and drafting, the prompt layer becomes part of the litigation file in substance, even if it sits inside a vendor’s system. Firms need to know which tools are being used, whether the tool retains prompts, whether the tool trains on user inputs, whether uploads are stored, who can access logs, and how those logs would be handled in response to subpoenas, litigation holds, regulator demands, or client disputes.

The Practical Consequence for Legal Teams

The strongest reading of Assini is not that AI records are categorically protected. The stronger reading is that courts will examine purpose, context, control, and procedural posture. A pro se litigant using ChatGPT to prepare filings presents one set of equities. A corporation feeding sensitive strategy into a consumer chatbot presents another. A lawyer using an approved enterprise tool under the firm's documented procedures presents yet another.

That variation is exactly why firms should not rely on after-the-fact privilege arguments. The defensible position has to be built before the dispute arises. Lawyers need policies that distinguish public AI tools from approved systems. Matter teams need rules for client confidential information, litigation facts, draft filings, deposition preparation, legal research, and attorney mental impressions. Firms also need retention logic. Keeping everything creates exposure. Deleting everything without regard to litigation holds creates a different one.

The operational issue is no longer whether lawyers use AI. Many already do. The issue is whether their use is sufficiently governed to preserve confidentiality, support work-product claims, satisfy court rules, and avoid creating an unnecessary evidentiary record for an adversary.

Assini is not a sweeping immunity decision. It is a useful early boundary. The court refused to let civil discovery become a back door into a litigant’s AI-assisted preparation. For legal professionals, the message is more disciplined than reassuring: prompts can be protected when they function as litigation work product, but protection is easier to defend when the use of AI is deliberate, documented, and controlled.