Within days of each other in early June, two courts answered a question litigation teams have been asking more frequently of late: Can a party’s AI prompts be pried open in discovery?
On June 3, Judge Grant Dorfman of the Texas Business Court held that a litigant’s ChatGPT conversations were protected attorney work product and that using ChatGPT did not waive that protection. Tate Group Automotive, LLC v. Legacy Automotive Capital, LLC, No. 25-BC11B-0020 (Bus. Ct. Tex. June 3, 2026). The next day, Justice Rhonda Fischer of the New York Supreme Court, Nassau County, quashed a subpoena seeking a litigant’s entire ChatGPT account, holding that interacting with an AI tool neither surrenders confidentiality nor waives work-product protection. Assini v. Hayward, 2026 NY Slip Op 26086 (Sup. Ct. Nassau County 2026). For readers who followed our earlier coverage warning that AI chats are not automatically privileged, this is the other half of the story, and it now comes from two separate court systems.
What the Texas Court Held in Tate
The defendants made two arguments: that the work-product doctrine does not reach a non-lawyer’s chats with an AI tool, and that uploading legal analysis to ChatGPT waived any protection. Judge Dorfman rejected both. He followed Warner v. Gilbarco, Inc., 2026 WL 373043 (E.D. Mich. Feb. 10, 2026), and Morgan v. V2X, Inc., 2026 WL 864223 (D. Colo. Mar. 30, 2026), which hold that work product is waived only by disclosure to an adversary, or where the risk of adversarial access rises substantially. On the record before him, AI chats did not meet that standard.
Tate matters because it is the first decision to apply the doctrine in an ordinary commercial setting, where a represented party uses AI as a litigation-support tool. The governing rule does the work: Texas Rule of Civil Procedure 192.5(a)(1) protects material prepared “by or for a party” in anticipation of litigation. That language answers the objection that a non-lawyer’s notes cannot be work product. It also describes the scenario in which most disputes will actually arise.
“The issue presented is a novel one, as evidenced by the fact that all case law authorities cited by the parties date from 2026.”
— Judge Grant Dorfman, Tate Group Automotive v. Legacy Automotive Capital (Bus. Ct. Tex. June 3, 2026)
What the New York Court Held in Assini
In Assini, the plaintiffs skipped asking the self-represented defendant what he had withheld and instead subpoenaed non-party OpenAI directly, demanding the defendant’s full account: every prompt, input, upload, and output used to draft his filings. Justice Fischer quashed the subpoena. A self-represented litigant can claim work-product protection for AI-assisted preparation, she held, and using ChatGPT does not waive it. The defendant analogized his use to “scribbling notes on a legal pad or brainstorming strategy with a colleague,” and the court agreed that confidential, strategy-focused work is exactly what trial-preparation protection shields.
“It is entirely reasonable for a person to expect some privacy and confidentiality when interacting with these tools.”
— Justice Rhonda E. Fischer, Assini v. Hayward (Sup. Ct. Nassau County 2026)
Assini adds what the federal cases could not. It is the first state-court decision of its kind, which matters because most civil litigation lives in state court, and the first to confront a subpoena aimed squarely at the AI platform rather than the litigant. Two practical points stand out for in-house teams. First, the court equated AI use with cloud storage: third-party technical access to data does not automatically end its protections. Second, New York’s new AI practice rule, 22 NYCRR Part 161, effective June 1, 2026, requires practitioners to confirm their submissions contain no fabricated citations and to certify compliance. Confidentiality and a duty to verify now travel together.
The Federal Foundation
The two federal cases the state courts leaned on draw the same line from both directions. In Warner, an Eastern District of Michigan employment case, Magistrate Judge Anthony Patti refused to compel production of a self-represented plaintiff’s AI materials, reasoning that AI programs are tools, not adversaries, and that the contrary rule “would nullify work-product protection in nearly every modern drafting environment.” Morgan, a District of Colorado employment case, protected the litigant’s AI-assisted analysis but policed the inputs: the court ordered the plaintiff to disclose the name of any AI platform used for confidential information and amended the protective order to bar confidential material from consumer AI tools unless the provider is contractually prohibited from training on or disclosing it. Morgan protects the thinking; it sets hard guardrails on what gets typed in.
The contrary federal decision, United States v. Heppner, 2026 WL 436479 (S.D.N.Y. Feb. 17, 2026), is the lone outlier, and the civil courts have confined it to its facts: a criminal matter in which a counseled defendant, later convicted of fraud, used a consumer tool on his own, without telling his attorneys. That gap between party and counsel does not exist in ordinary civil litigation, and all four civil decisions distinguish Heppner on that basis. As we explained in our earlier analysis of why Heppner and Warner both got it right on their respective facts, the tension is more about differing circumstances than a genuine split.
The Fault Line: Confidentiality Is Not Waiver
The through-line is that confidentiality and waiver are two different questions. Courts siding with the litigant treat typing into an AI tool as neither an adversarial disclosure nor a surrender of confidentiality. Four civil decisions across two court systems now keep those questions separate. None of them, however, is a license to feed anything into any tool. Tate ordered production of pages that were not genuine work product and, borrowing from Morgan, required the litigant to identify every discovery document it had shared with ChatGPT. The doctrine protects your analysis. It does not cure a protective-order breach, excuse an unverified citation, or shield inputs you had no business pasting in.
Three Key Takeaways:
- Protection now spans court systems. Federal courts in Michigan and Colorado, the Texas Business Court, and a New York trial court have all held that AI-assisted work product is not waived by using the tool.
- Govern your inputs as carefully as your analysis. Reserve confidential and opposing-party material for enterprise-tier platforms contractually barred from training on or disclosing your data, and document those protections, the standard Morgan imposes.
- Verification duties are rising. Courts will still compel non-work-product pages and the identity of the tool, and rules like New York’s Part 161 require you to check AI output before filing.
