Earlier this year, a financial executive opened an AI chatbot to think through his legal exposure before calling his lawyers, then watched a federal judge hand all 31 of those documents to prosecutors.
The lesson looked clear: A strategy session typed into a consumer AI tool is not privileged, not work product and fully discoverable.
A Texas court has now complicated that lesson.
In a commercial dispute before the Business Court of Texas, Kris Tate did much the same thing, using ChatGPT to work through a fight that had landed in litigation. When the conversations surfaced in discovery, Tate Group Automotive withheld them as attorney work product. The defendants moved to compel, arguing that a nonlawyer’s chats with an AI tool are not work product at all and that, in any event, feeding the material to ChatGPT waived any protection it carried. After reviewing the conversations in camera, Judge Grant Dorfman largely sided with the plaintiff. Most of the material stays protected.
Same conduct, two courts, opposite results — and not by accident. The federal court that decided United States v. Heppner and the Texas court that decided Tate Group Automotive v. Legacy Automotive Capital read their work product rules differently. Texas diverges from the federal courts on whether AI-assisted material is work product in the first place and lines up with the emerging federal majority on whether using AI waives it. Both moves favor protection. For Texas litigators, the gap between those two regimes is worth understanding.
Two Questions, Not One
It is tempting to file the Tate decision under “another court weighs in on AI and discovery” and move on. That framing, however, misses what is actually happening. The cases in this area are not fighting over a single question. They are fighting over two, and the two come apart.
The first question is whether AI-assisted material qualifies as protected work product at all. The second is whether sharing that material with an AI platform waives the protection. A court can answer yes to the first and no to the second, or the reverse, and the combination is what decides the case.
Sort the decisions that way and the picture clarifies. In Heppner, the federal court answered the first question no — the documents were not work product because the client prepared them without counsel’s direction — and never needed to reach waiver. In Warner v. Gilbarco and Morgan v. V2X, two federal courts took up the second question and answered it for what is becoming the majority view: Using a consumer AI tool is not the kind of disclosure that waives work product, because it is not disclosure to an adversary.
The Texas Business Court is the first to test this conduct against Texas’s own work product rule rather than the federal doctrine, and it answered both questions in favor of protection, though for different reasons on each. On the threshold question, it parted ways with Heppner. On waiver, it joined Warner and Morgan.
The Texas Rule Reaches Further
In Heppner, the court assumed the AI documents were prepared in anticipation of litigation and still found no work product protection because they “were not prepared by or at the behest of counsel” and did not reflect counsel’s strategy. Heppner’s own lawyers conceded he had run the AI sessions on his own initiative, without their direction. To Judge Jed Rakoff, that was disqualifying. The work product doctrine, he reasoned, exists to protect the lawyer’s mental processes — it is anchored to counsel, not to the client’s preparation.
Texas Rule 192.5(a)(1) defines work product to include “material prepared or mental impressions developed in anticipation of litigation or for trial by or for a party” — by or for the party, not merely the party’s attorney. The federal rule says nearly the same thing — it too protects material prepared “by or for” a party. The divergence is not in the words but in how courts read them. Heppner treated the doctrine as a shelter for the lawyer’s mental processes and, on that view, declined to extend it to a client’s preparation. Texas courts have not grafted that limitation onto the rule. Read for what it says, 192.5(a)(1) protects the party’s own work — and that is how Judge Dorfman read it.
Judge Dorfman applied the rule as written. Reviewing Kris Tate’s ChatGPT conversations in camera, he found they “plainly appear on their face” to fall within 192.5(a)(1): material prepared, and mental impressions developed, in anticipation of litigation, by a party. That the party was Tate rather than his lawyers — that the work ran through a chatbot rather than a legal pad — did not push the conversations outside the rule’s plain language. Most of the withheld material stayed withheld.
This is the heart of the split, and it is not a quarrel about facts. Heppner and Tate confronted the same conduct and reached opposite results because they read materially similar rules to draw the protective line in different places. Federal work product, as Heppner construes it, protects the lawyer’s thinking; the absence of an attorney’s hand defeated the claim. Texas work product protects the party’s thinking too; the presence of the party’s hand is exactly what the rule shields. For Texas litigators, the consequence is concrete: The strategy session that became government evidence in New York may remain protected in a Texas business dispute. The rule, not the conduct, is doing the work.
Use Is Not Disclosure
If the threshold question is whether AI-assisted material counts as work product, the second question is whether using the AI waives it. Here Texas joins a developing federal majority and rejects the same authority it rejected on the first question.
The defendants argued, in the alternative, that even protected material lost its protection the moment Tate fed it to ChatGPT. Judge Dorfman disagreed, adopting the reasoning of Warner v. Gilbarco and Morgan v. V2X. Those courts drew a line that matters: Work product is waived by disclosure “to an adversary, or in circumstances that substantially increase the likelihood that an adversary will obtain the materials.” Under most circumstances, typing into a chatbot does neither.
The contrary view leans on the platform’s privacy policy: The terms say inputs may be collected, used for training and disclosed to third parties, so confidentiality is gone and protection with it. The problem is that this reasoning proves too much. That language is not unique to one chatbot. It is functionally identical across the consumer AI tools, cloud storage services and email providers that every modern litigant uses every day. If a generic reservation of the right to disclose were enough to waive work product, almost nothing prepared on a computer would survive. Warner, Morgan and now the Texas Business Court decline to read the doctrine so broadly.
Protected Is Not Privileged
A word of caution before anyone reads too much into Tate. The Business Court protected the chats as work product. But it said nothing about attorney-client privilege — and that silence matters.
Heppner rejected the privilege claim on grounds Tate never touches: A chatbot is not a lawyer, no attorney-client relationship forms by using one, and a client who turns to AI on his own is not seeking legal advice from counsel. Nothing in the Texas decision revives privilege for an AI conversation, and nothing in it suggests a Texas court would see the privilege question differently.
The distinction is more than academic. Privilege, once it attaches, is powerful but fragile; any disclosure to a third party can destroy it. Work product is qualified rather than absolute, but it tolerates exactly the kind of third-party exposure, a chatbot included, that privilege cannot. Texas may shield the client’s AI-assisted thinking as work product. It does not make the chatbot the client’s lawyer.
The practical lesson is the one from earlier this year, refined: If you want the most protection, the path still runs through counsel.
The Business Court as a Pragmatic Discovery Forum
Tate will draw attention, and most of it will cast the opinion as a doctrinal ruling — Texas’ work product rule applied to AI use. That framing captures part of the decision and misses the rest. How Judge Dorfman decided the issue is as significant as what he wrote about Rule 192.5(a)(1).
The Method, Not Just the Result
One striking feature of Tate is what the court declined to hold. Judge Dorfman confronted an issue he acknowledged was novel — “all case law authorities cited by the parties date[d] from 2026” — yet issued no sweeping rule about whether, in all instances, a nonlawyer’s AI chats qualify as work product. The court instead reviewed the withheld ChatGPT conversations in camera and resolved the dispute document by document, ordering specific pages produced because they did not constitute protectable work product, while permitting the remainder to be withheld. The court produced a process: That portion produced, that portion protected, each measured against the rule’s text.
The approach should be familiar to anyone who has litigated a complex discovery dispute in the Business Court. The court brought the same document-by-document, evidence-driven balancing to the confidentiality-designation fight in Westlake Longview Corp. v. Eastman Chemical Co. and to the trade-secret privilege dispute in SafeLease Insurance Services v. Storable. In Westlake, the court refused to rule in the abstract, requiring specific evidence of harm and a particularized need before it would decide. SafeLease ordered even assumed trade-secret material produced where the requesting party needed it, with the protective order supplying the protection. Tate extends the Westlake and SafeLease philosophy; the issue was novel, but the court’s method was not.
Protection of Content Is Not Protection of Conduct
Tate shows the same pragmatism in how the court separated the substance of the AI chats from their use. Tate’s work product victory brought him no clean exit. Drawing an analogy to Morgan v. V2X, which ordered the plaintiff to disclose the identity of the AI tool it had used, the court held that work product protection extends only so far: Tate has to disclose the AI tool he used and all discovery materials he fed to ChatGPT, including material produced under the protective order. The content of Tate’s strategy session remains protected; its existence and his use of the tool do not.
Tate instructs practitioners on a practical point. A party that turns to a consumer AI tool in litigation must account for what it does with the tool. Tate does not shield a party from producing what it shared with a chatbot, and a party that cannot reconstruct what it shared may face a discovery problem — particularly when it received that information under a protective order.
The Protective Order Dimension
The last point led Judge Dorfman to flag a distinct concern. Feeding protected discovery material to a third-party AI tool may itself violate the protective order, independent of any work product question, and the court expressly reserved a ruling on that point. The court declined to resolve a problem not yet before it and instead urged the parties to confer and negotiate amendments to the protective order — amendments that would fix whether and how far protected material may be shared with any AI tool or large language model — pointing them to Morgan for guidance.
That recommendation invites counsel to draft, and counsel should take it up before a dispute arises. A workable AI-use provision addresses, at a minimum, which tools are permitted (and whether consumer services differ from enterprise ones), the tool’s training and retention terms, the handling of attorneys’-eyes-only and trade-secret designations, and a party’s notice and accounting obligations when it runs an AI tool on produced material — so that the disclosure Tate ordered follows as a matter of course rather than by motion. Drafting these terms into the protective order at the outset converts an open question into a governed one — precisely the result Tate encourages.
Don’t Overread Tate — the Procedural Posture
A final caution tempers everything above. Tate is not a final, appealable ruling. It is a minute entry under Business Court Local Rule 4, which, as the court took care to note, certifies the parties’ compliance with the local rule rather than resolving the dispute with finality. The court left any party that wants an ultimate ruling — particularly one eyeing appeal — with leave to file and set a motion to quash, for protection, or for reconsideration, without first proceeding under BCLR 4. Tate therefore signals how this court approaches a novel discovery question; it does not bind anyone as precedent that resolves one. That posture is a limitation of Business Court discovery practice as much as a virtue of it: BCLR 4’s letter process moves efficiently, but a complex question aired through informal letters and a nonfinal minute entry yields no definitive ruling until a party forces one through formal motion practice.
That distinction governs how much weight the decision can bear, but it does not diminish its practical significance. Tate offers a reliable read on the forum. A litigant weighing where to position a dispute over confidentiality and novel discovery questions can draw from Tate, Westlake and SafeLease a consistent expectation: the Business Court will engage closely with the record, calibrate relief to the specific materials and the specific showing, and resist both blanket secrecy and blanket disclosure. For the kind of dispute in which a strategy session typed into a chatbot might surface in discovery, that predictability is no doctrinal curiosity — it is a reason to choose the forum.
Conclusion: Forum Now Matters
The AI-discovery map is being drawn case by case, and Texas has placed a marker on the protective side of the line. Tate divides cleanly along the two questions that decide these disputes: On the threshold question, Texas reaches further than the federal courts; on waiver, it joins the emerging majority; on privilege, nothing has changed, and the answer remains no; and on conduct, even a party that wins must disclose the AI tool and account for its inputs.
Forum is now a strategic variable. Where a client litigates affects whether his AI-assisted thinking survives discovery, so the question belongs in the venue calculus from the outset, alongside the usual considerations of jury pool, docket speed and judicial experience. The lawyer’s own duties of competence and confidentiality, by contrast, do not turn on the forum — Texas Ethics Opinion 705 applies them to any use of generative AI, even where work product protection does not.
The two forums divide the same conduct. Earlier this year, a federal court handed an executive’s AI strategy session to the government. In Texas, the same session may survive — but with strings: disclosure of the tool, an accounting of the inputs and compliance with the protective order.
None of this is settled. Tate is a nonfinal minute entry, and more decisions are coming—from the Business Court, from the federal courts and eventually from appellate panels that will test these rules under pressure. The prudent course, though, does not change with the forum: involve counsel early, route sensitive analysis through vetted tools and write AI terms into the protective order before the dispute arrives.
Christopher Schwegmann is the managing partner of Lynn Pinker Hurst & Schwegmann LLP.
Ben A. Barnes is a partner in Dowd Bennett’s Dallas office.