# Open AI Just Lost in Court. Your Discovery Practice Didn’t Get the Memo ### JD Morris Published Apr 14, 2026 The Technology Blind Spot Jeffries v. Harcros Chemicals, Inc. is a toxic tort class action. The allegation: decades of carcinogenic emissions, including ethylene oxide, from a Kansas City, Kansas chemical plant caused cancer clusters, reproductive harm, and serious disease across surrounding neighborhoods. The case has nothing to do with legal technology. And yet, buried inside a March 25, 2026 protective order ruling, a federal magistrate judge drew a line that will reshape how every attorney in this country manages discovery materials. The line is simple: open artificial intelligence (AI) tools are out. Closed AI tools are in. And the reasoning behind that line goes deeper than most practitioners realize. What the Court Actually Held Magistrate Judge Angel D. Mitchell of the United States District Court for the District of Kansas granted defendants’ motion to amend the protective order in Jeffries v. Harcros Chemicals, Inc., Nos. 25-2352-KHV-ADM, 25-2569-KHV-ADM (D. Kan. Mar. 25, 2026). The ruling did not restrict AI use generally. It drew a specific, operational distinction between two categories of tools. The court used “open AI Tools” and “closed AI Tools” as defined legal terms throughout the opinion, not references to OpenAI the company. “Open AI Tools” are publicly accessible generative AI systems that may retain and use submitted data to train their underlying models. “Closed AI Tools” are systems that operate in a controlled environment, do not use submitted data for model training, and provide contractual guarantees over data handling and deletion. The court’s holding covered all discovery materials, including specifically those that are not designated confidential. That scope is the headline. An attorney who uses a consumer-facing AI tool to analyze the other side’s document production is now, in this court’s view, doing something the rules do not permit. Why This Ruling Is Not an Isolated Data Point Jeffries did not emerge from a vacuum. It arrived in the same six-week window as three other significant federal rulings. Together, the four decisions address two distinct questions that practitioners need to keep separate. The first question is whether using a consumer AI tool destroys privilege or work product protection. Federal courts answered that question on February 10, 2026, and reached opposite conclusions the same day. In United States v. Heppner, 25 Cr. 503 (JSR) (S.D.N.Y. Feb. 10, 2026) (written opinion issued Feb. 17, 2026), Judge Jed S. Rakoff declined to extend attorney-client privilege or work product protection to documents a defendant generated using the consumer version of Anthropic’s Claude platform, outside the direction of counsel. [See “Your AI Tool Just Waived Attorney-Client Privilege,” The Technology Blind Spot, February 2026.] The court’s reasoning rested on three independent grounds: no attorney created the documents, the consumer platform’s terms of service gave users no reasonable expectation of confidentiality, and the documents did not reflect defense counsel’s strategy. Consumer Claude, the court found, was not a privileged channel. On the same day, Magistrate Judge Anthony P. Patti of the Eastern District of Michigan reached the opposite conclusion in Warner v. Gilbarco, Inc., No. 2:24-cv-12333, 2026 WL 373043 (E.D. Mich. Feb. 10, 2026). The plaintiff in Warner was a pro se employment discrimination litigant who used ChatGPT to prepare her litigation materials. When defendants moved to compel production of everything she had generated using AI tools, the court denied the request. It held that a pro se litigant’s AI-generated materials are protected under the work product doctrine because they reflect mental impressions and litigation strategy, and that using an AI tool does not constitute disclosure to an adversary. “ChatGPT (and other generative AI programs) are tools, not persons,” the court wrote. Using a tool does not waive a privilege. The court’s closing line states it plainly: “In the end, both sides of this dispute seek to obtain each other’s thought processes, while shielding their opponent from discovery of their own. The Court will uphold the protections afforded the thought processes and litigation strategies of both sides and will order production of neither.” The apparent contradiction between Heppner and Warner resolves when you examine what each court actually measured. Three variables drove both outcomes: who directed the AI use, whether a reasonable expectation of confidentiality existed given the platform’s terms of service, and whether the materials were prepared in anticipation of litigation. Heppner failed on all three: the defendant acted without counsel’s direction, the consumer platform’s terms of service eliminated any confidentiality expectation, and the documents did not reflect attorney strategy. Warner satisfied all three in the specific context available to a pro se litigant: she was her own representative, the work product doctrine covers materials prepared by a party directly, and the court found no evidence she had uploaded documents covered by the protective order into the AI platform. The cases do not contradict each other. They describe a framework. The practical question is whether represented clients can reliably operate within that framework on consumer platforms. In most discovery contexts, they cannot. A second, distinct question is whether courts should use protective orders to restrict open AI use in discovery entirely, regardless of what privilege or work product might otherwise protect. That is the question Jeffries and Morgan answer, and it is a separate question from the first. Jeffries and Morgan v. V2X, Inc., 2026 WL 864223 (D. Colo. Mar. 30, 2026), came down five days apart in late March and should be read together. In Morgan, Magistrate Judge Maritza Dominguez Braswell of the District of Colorado addressed a protective order dispute in an employment discrimination case and drew her own open/closed distinction, requiring that any AI tool used to process confidential information be subject to contractual prohibitions on model training, restrictions on onward disclosure, and enforceable deletion rights. The court was direct about the practical consequence: “this provision will (at least for now) bar the parties from using most, if not all, mainstream low-to-no-cost AI to process Confidential Information.” That Kansas ruling extends Morgan by one significant step: it applies the open/closed restriction to all discovery materials, not just those marked confidential. Taken together, these four decisions do not establish a trend. They establish a trajectory. Courts are not experimenting with AI governance. They are building doctrine. The Quote That Isn’t the Judge’s Before going further, a correction to commentary circulating in legal technology publications is necessary. The line being widely attributed to Judge Mitchell is this: “what is clear is that the way AI Tools function poses a potential threat to the integrity and security of data produced in litigation.” That language is not hers. It came from defendants’ brief. The court quoted and adopted the reasoning, which carries legal significance because judicial adoption gives the argument authority, but the words belong to Harcros and its co-defendants. This matters if you cite this case in a filing or a client memo. Misattributing advocacy to a holding is an error opposing counsel will catch. What Judge Mitchell actually said, in her own voice, is worth more than the misattributed line: “If the court were to allow a party to upload all non-confidential documents and materials produced by another party to an open AI Tool, thereby making all such data amenable to public consumption, parties may err on the side of under-producing potentially responsive documents or seek to make extensive redactions of irrelevant or non-responsive information. Defendants’ proposed amendment allowing the use of closed AI Tools, as opposed to open AI Tools, will facilitate discovery by incentivizing more fulsome document productions.” That paragraph is a judicial finding about attorney behavior, not a technology policy preference. It says that if open AI tools remain available, producing parties will produce less and redact more. The remedy for that perverse incentive is restricting open tools while permitting closed ones. That is the court’s reasoning. Get it right when you use it. The Mechanics Behind the Ruling To understand why the court reached this result, you have to understand how open AI tools actually work. The court explained it plainly: these systems identify and encode patterns in massive volumes of data, and they use the data submitted to them to continually develop and improve the underlying model. That process runs in one direction. Once a document enters the training pipeline, a court order cannot retrieve it. The Jeffries court stated this directly: defendants argued it is “practically impossible” to claw back data later determined to be privileged, or to delete data from an open AI tool at the conclusion of the action, as protective orders routinely require, because the information has already been absorbed into the model’s training. Judge Mitchell found that argument unrefuted. Plaintiffs, she noted, “never directly address Defendants’ argument that clawing back or deleting data submitted to an open AI Tool is impossible as a practical matter.” You cannot delete what has already been learned. Meet Dana Schreiber. She is a composite: the name is invented; the scenario draws from practitioner accounts describing variations of the same mistake. Her firm is real in type: a twelve-attorney regional practice, one associate producing for a national manufacturing client in a commercial dispute. It is 11:15 on a Tuesday night. She has 40,000 documents in a rolling production from opposing counsel, a privilege review deadline at noon, and a free-tier AI platform open in her browser. She pastes the first 200 pages into it to build a privilege log summary. Among those pages: internal email chains between the client’s European business partners and their legal departments. By the time Schreiber’s alarm goes off Wednesday morning, that material has entered a training pipeline. No court order issued later that week can retrieve it. The Jeffries court treated the sequence Schreiber set in motion not as a privacy risk but as a discovery integrity problem, and the distinction is what makes the ruling consequential. On the GDPR point, the court went further than most practitioners will expect. The defendants in Jeffries included European corporate entities subject to GDPR jurisdiction. The court found that wholesale submission of discovery materials to an open AI tool could violate GDPR’s consent requirements, noting that defendants’ employees, contractors, and correspondents had not consented to their information being used to train a public model. GDPR Article 4(11) requires documented consent that is “freely given, specific, informed and unambiguous.” A proposal to redact certain personally identifiable information before uploading materials did not satisfy that standard. For any attorney handling matters that touch European parties, subsidiaries, or employees, that holding is not peripheral. The Professional Responsibility Angle Judge Mitchell’s analysis extended into the ethical obligations of counsel, adopting defendants’ argument that the inability to claw back or delete materials from an open AI tool creates risk of violating counsel’s professional duty to safeguard information relating to the representation of a client. That analysis maps directly to American Bar Association (ABA) Formal Opinion 512, issued July 29, 2024, which addressed attorneys’ ethical obligations when using AI tools in their practice. The Opinion acknowledged that while many lawyers already use AI-based technologies to improve efficiency, the emergence of generative AI tools raises novel concerns about client confidentiality. ABA Opinion 512 does not prohibit AI use. It requires that attorneys understand the tools they use well enough to protect client information. Jeffries gives that standard concrete application in a discovery context. Understanding your AI tool is no longer a general competence obligation sitting in a bar association opinion. It is now a specific discovery compliance question with case law behind it: is the tool you are using open or closed, and can you prove it?

Originally published on LinkedIn Newsletter: The Technology Blind Spot