# Your Judge’s AI Search Just Became Discoverable You don’t know what your judge typed into ChatGPT yesterday. Until April 29, 2026, that didn’t matter. Now it might be discoverable. That morning, a doctrinal paper posted to the Social Science Research Network argued that a judge’s AI research queries during a pending case waive the deliberative privilege at the moment of transmission. A March 2026 Northwestern University survey, the first random-sample study of federal judges, found more than 60 percent of responding judges had used at least one AI tool in their judicial work, 30 percent for legal research. Those vendors operate under terms authorizing retention, training use, and disclosure to governmental authorities. The paper predicts the first subpoena to an AI vendor for a judge’s interaction logs within twenty-four months. The first attorney to write that subpoena makes new law. # The Pattern In October 2025, Senator Grassley released letters from two sitting federal judges, Henry T. Wingate (S.D. Miss.) and Julien Xavier Neals (D.N.J.), admitting chambers staff had used generative AI to draft court orders containing fabricated citations. The judiciary’s Mata v. Avianca moment had arrived. Five months later, the Northwestern survey confirmed what Grassley had only glimpsed: AI use on the federal bench is not anomalous. It is majority practice. Six months later, the Sixth Circuit moved the consequence to the appellate level. In United States v. Farris , the court removed appointed counsel and denied Criminal Justice Act compensation after the attorney admitted that Westlaw CoCounsel, Thomson Reuters’s own AI platform, generated false quotations and misleading legal arguments in appellate briefs. Hallucination risk was not confined to consumer tools. It sat embedded in premium legal research infrastructure. [See Your AI Research Tool Fabricated the Quotation , The Technology Blind Spot (2026).] What Wingate, Neals, and the Farris attorney share matters more than what each did. Each transmitted analytical reasoning to a commercial AI platform. None realized transmission was the legal event. # The Doctrinal Map The paper that maps the consequence is Candor and the Algorithm: Deliberative Privilege and the Structural Vulnerability of Judicial AI , by my colleagues Alexis Austin Litle and Deepankar Das, and me. Its thesis is structural. Deliberative privilege, the qualified protection that shields pre-decisional governmental reasoning from compelled disclosure, exists because officials who fear disclosure self-censor, and self-censorship degrades decision-making. The privilege has never been tested against a technology that records every research query and routes it to a private server. A judge asking ChatGPT whether McDonnell Douglas requires judgment for the plaintiff when the employer’s stated reason is pretextual has revealed her analytical posture. Both prongs of In re Sealed Case are satisfied before she hits return. Judge Xavier Rodriguez of the Western District of Texas, a Northwestern survey co-author, has published a first-person account of using AI tools in a federal voting-rights bench trial. He used Claude through a retrieval-augmented platform to condense trial testimony and Clearbrief to verify citations, flagging the risks Candor now formalizes: automation bias, vendor training-use terms, confidentiality exposure. Rodriguez is the careful judge. His logs would be the most valuable subpoena target on the federal bench. # The Heppner Mechanism Waiver mechanics are established. In United States v. Heppner , Judge Jed Rakoff held that defendant Bradley Heppner’s exchanges with the Claude AI platform were protected by neither attorney-client privilege nor work product. Three independent grounds: Claude is not an attorney; Anthropic’s privacy policy authorizes collection, training use, and third-party disclosure of inputs and outputs; and Heppner had not communicated with Claude to obtain legal advice. Even assuming privileged information had been input, sharing with any third party waived it. [See Your AI Tool Doesn’t Keep Secrets , The Technology Blind Spot (2026).] Candor applies the same mechanism to judges. A judge’s transmission of pre-decisional reasoning to a consumer AI platform triggers two ALRM factors dispositively: the system processes inputs to improve its model, and vendor terms authorize training use. Both answer yes for every consumer AI account in commercial release. The judge has voluntarily disclosed deliberative material to a third party with divergent interests, and that disclosure is the waiver event under federal common law. Candor names the structural feature “waiver-before-balancing.” The privilege is qualified, not absolute, but the balancing test never applies if the privilege is already waived. The judge has lost it before any court has weighed competing interests. # The First Subpoena A litigant who learns, through public reporting, chambers staff disclosure, or the discovery requests this analysis enables, that a judge has consulted a consumer AI platform on a question relevant to a pending motion has acquired the predicate facts for a first-of-its-kind subpoena. The target is the AI vendor. The demand is the interaction log. The doctrinal predicate is Heppner by analogy. Whichever court confronts that subpoena first writes the rule. There is no controlling authority. There is a Sixth Circuit per curiam saying attorneys cannot rely on AI without verification, a Northwestern survey saying 60 percent of judges use it anyway, and a doctrinal map on SSRN. Predict the source. The first subpoena will come from a party that has just lost a major patent case or other high-profile litigation. The adverse judgment will carry enough weight to overcome the judicial deference that has historically shielded deliberations from compelled disclosure. The demand will reach the judge’s reasoning and the AI chat history that informed it. # The Counterargument Strongest objection: no court has applied Heppner to deliberative privilege. Heppner was an attorney-client and work product case. Deliberative privilege waiver has historically turned on adversarial receipt. AI platforms are not adversaries. The distinction is real, and Candor meets it. Deliberative privilege waiver, unlike work product, does not require adversarial receipt. It requires only voluntary disclosure that destroys the reasonable expectation of confidentiality in pre-decisional reasoning. A vendor authorized by its own terms to retain, train on, and disclose inputs eliminates that expectation as a matter of contract. The objection narrows the doctrinal route; it does not close it. # Monday Morning By the start of next week, a federal litigator with active matters can take three actions. Pull every active case management order and read for the AI question. Most address the parties’ AI use. Few address the court’s. That asymmetry is the opening. Raise the question on the record. In your next status conference, scheduling order request, or pretrial filing, ask directly: does the court, or any chambers personnel, use generative AI tools in this matter, and if so, on which platforms and under what terms? The question is preservative, not adversarial. The record it creates is the record an appellate court will need. Stipulate where you can. In matters with confidential information already produced under protective order, draft a one-page Federal Rule of Evidence 502(d) supplemental stipulation barring the parties, and by negotiated extension chambers, from inputting protected materials into any consumer AI platform whose terms authorize training use. A 502(d) order does not bind a non-party judge directly. The request creates the contemporaneous record on which a subpoena rests. # The Keystrokes Return to the opening image. The judge’s keystrokes are not a court order. They are not a published opinion. They are not part of the case file. Until this year, they did not exist as a category of preserved record at all. Judges have always reasoned about cases between hearings, asked questions of clerks, talked through theories with colleagues, and read more than they cited. What changed is that one of those conversations now leaves a transcript on a server controlled by a private company. The first attorney to demand that transcript will discover what deliberative privilege was always supposed to protect, and, perhaps, that it no longer does. _______________________________ # About the Author JD Morris is Co-Founder and COO of LexAxiom, an Agentic AI platform for the business of law. Over a 25-year career, he has built and scaled enterprise technology products across Dell, EMC, VMware, and Cisco, including the first exabyte eDiscovery platform. He holds dual MBAs from Columbia Business School (Finance) and UC Berkeley Haas (Marketing), a Master of Legal Studies in Cybersecurity Law from Texas A&M, and a Master of Engineering from George Washington University. He writes The Technology Blind Spot on the intersection of emerging technology and law. Connect with him on LinkedIn at http://www.linkedin.com/in/jdavidmorris, on X at @JDMorris_LTech, or on Bluesky at @JDMorris-ltech.bsky.social. _______________________________ # References 1. Alexis Austin Litle, JD Morris & Deepankar Das, Candor and the Algorithm: Deliberative Privilege and the Structural Vulnerability of Judicial AI (Apr. 29, 2026) (manuscript), https://ssrn.com/abstract=6678078. 2. United States v. Heppner, No. 1:25-cr-00503-JSR, 2026 WL 436479 (S.D.N.Y. Feb. 17, 2026) (Rakoff, J.). 3. United States v. Farris, No. 25-5623, 2026 WL 915082 (6th Cir. Apr. 3, 2026) (per curiam) (recommended for publication). 4. Anika Jaitley, Daniel W. Linna Jr., Xavier Rodriguez, V.S. Subrahmanian & Siyu Tao, Artificial Intelligence in Federal Courts: A Random-Sample Survey of Judges, 27 Sedona Conf. J. ___ (forthcoming 2026), https://thesedonaconference.org/publication/Artificial_Intelligence_in_Federal_Courts. 5. Press Release, S. Comm. on the Judiciary, Grassley Releases Judges’ Responses Owning Up to AI Use, Calls for Continued Oversight and Regulation (Oct. 23, 2025), https://www.judiciary.senate.gov/. 6. Xavier Rodriguez, Judging AI: How U.S. Judges Can Harness Generative AI Without Compromising Justice, 109 Judicature 10 (2025). 7. In re Sealed Case, 121 F.3d 729 (D.C. Cir. 1997). 8. Fed. R. Evid. 502(d). 9. OpenAI, Terms of Use, “Our Use of Content” (effective Jan. 1, 2026), https://openai.com/policies/row-terms-of-use/.
Originally published on LinkedIn Newsletter: The Technology Blind Spot
