Discovery of AI Prompts and Chat Logs: What the Record Shows

The answer has depended on who did the prompting and why. In Villanueva v. Las Vegas Metropolitan Police Department, a federal district judge ordered an expert's ChatGPT case summaries produced as facts or data considered but protected his prompts and draft output. In Conservation Law Foundation v. Shell, a magistrate judge ordered an expert team's prompts produced as methodology; that order is stayed. For parties, United States v. Heppner rejected protection, while several trial courts protected party use as work product. No appellate court has resolved the question.

Where the reported AI discovery orders stand

The orders below are trial-level rulings, most from magistrate judges, and several are interlocutory (entered before final judgment and open to revision). They are set side by side so each can be read against the others, not as settled law.

A retained witness's chat history: considered material or drafts

In US federal courts, Federal Rule of Civil Procedure 26(a)(2)(B) requires a retained witness's report to disclose the facts or data considered in forming the opinions, and Rule 26(b)(4)(B) protects drafts of that report. Those provisions assume separate artifacts. One chat session can hold research, drafting and deposition preparation interleaved, so the recent orders have had to sort a single file into categories.

Villanueva shows the sorting done by hand. The ChatGPT use surfaced only at deposition, after a subpoena for "all data, documents, and other information considered by expert witness in forming the expert's opinions." Magistrate Judge Albregts ordered the whole log produced without reviewing it. On objection, District Judge Anne R. Traum reviewed it in camera and divided it. The order states that "The legal case summaries generated by ChatGPT constitute 'facts or data' that Melton considered and must be disclosed under Rule 26." The prompts and deposition outline stayed protected, and so did drafting output, as draft material "regardless of the form in which the draft is recorded." The judge found no waiver, because the program is not a person, and left methodology concerns to impeachment.

CLF v. Shell drew the line elsewhere. Magistrate Judge Thomas O. Farrish treated the prompts used to filter a document production as part of the method, and method is fair ground for discovery. That is the territory of Clause 5, reproducible analysis: a culling step that was never preserved cannot be rerun. The parties' stipulation excluding notes and drafts did not clearly reach prompts, and one reference to a "prompt" in a research assistant's declaration justified a further search. The two rulings are reconcilable on their facts, since choosing what to read is method while drafting prose resembles a draft. Neither claims to state a general rule.

Once the material exists, it is used. In the Watson Grinding trial, as reported, logs behind a 3M engineer's report were produced after a demand at deposition and became the core of his examination; accounts differ on how they were obtained. In Matter of Weber, a Surrogate's Court rejected a damages analysis whose author could not recall the prompts he had given Microsoft Copilot. And in TRAX, the Court of Federal Claims held that whether evaluators found AI outputs useful "has no bearing on whether it was considered."

A party's own chatbot sessions: privilege, work product and waiver

Two protections are in play when a party has used a chatbot. Attorney-client privilege covers confidential communications between client and lawyer for legal advice. Work product (material prepared in anticipation of litigation) shields a party's preparation from an adversary. Waiver (loss of a protection by disclosure) operates differently for each, and that difference has driven most results.

In Heppner, Judge Jed S. Rakoff held that a securities-fraud defendant's Claude documents, prepared on his own initiative and later shared with counsel, were neither privileged nor work product. The absence of any attorney disposed of privilege; the provider's privacy policy defeated any expectation of confidentiality; and the memorandum noted the analysis might differ had counsel directed the use.

Other trial courts reached protective results. Warner held that generative programs are tools, not persons, and that work product is waived only by disclosure to an adversary or in a way likely to reach one. Morgan held that a self-represented plaintiff, being party and advocate at once, has some work product protection over his AI use, yet ordered him to name the platform. Tym adopted Morgan and required a privilege log. Tate Group sustained protection under the Texas rule, which covers material prepared by or for a party, and said in terms that it disagreed with Heppner.

Shealy marks a boundary. A party's partner ran loan documents through ChatGPT with no lawyer involved, and the court held the queries and output unprotected unless performed at counsel's direction. Jones v. Delta answers a narrower question: a deponent with no attorney could not invoke attorney-client privilege, and could not consult a system while testifying. Across these orders the deciding facts were who prompted and whether counsel directed it, not which product was used.

Going to the provider: subpoenas, seizure and certifications

A session record usually sits on the provider's servers, which makes the provider a third party from whom records can be sought without the user's participation. Assini v. Hayward is the reported test. The plaintiffs subpoenaed OpenAI OpCo, LLC for a pro se defendant's prompts, outputs, uploads, drafts and account data. The Nassau County court quashed it, reasoning that the contents resemble the confidential, strategy-laden iterative work protected by CPLR 3101(d), and declined to follow Heppner.

That reasoning rested on litigation strategy. A retained witness's prompts are more often treated as method, as in CLF v. Shell. Two other routes appear in the record. In Heppner, agents simply seized printed documents in a search. In Fry v. Fry, the Eastern District of Pennsylvania required a self-represented petitioner to certify which tools he used and list every prompt and document entered; his disclosure that redacted filings had gone to non-enterprise Claude then limited him to narrow redactions, because the court would not seal what had already been given to a private company.

Protective orders that now name generative tools

A protective order governs how produced material may be used and by whom, and it binds any retained specialist who receives that material. In 2026 several began to name generative tools.

• Jeffries v. Harcros. Magistrate Judge Angel D. Mitchell extended AI restrictions to all discovery materials, barring open-loop tools while permitting closed ones, reasoning that the rule would "facilitate discovery by incentivizing more fulsome document productions."
• Dunn v. LexisNexis. Magistrate Judge Maritza Dominguez Braswell entered an agreed AI provision without requiring consent before each new tool, while recognizing that "Knowing how confidential information is handled is a legitimate interest."
• Morgan v. V2X. The amended order bars Confidential Information from any platform whose provider is not contractually prohibited from storing or training on inputs.
• Tate Group. The court recommended an AI amendment to the protective order and meanwhile required a list of everything shared with ChatGPT, including protected productions.

None has been reviewed on appeal. In each, the question of what may be submitted to which tool is settled by the order in the matter, and it attaches to the material rather than to the task.

What an expert should keep, and expect to be asked for

The Standard does not predict rulings. It sets a working posture: assume the record of system use will be produced, because the cost of being wrong runs in one direction. Clause 7, contemporaneous record, requires a record, made as the work proceeds, of each system used, its purpose and how output was verified. Clause 8, disclosure, requires stating at retention that use will be disclosed and the record produced on proper request.

• Keep one log per matter. The AI-use log template records date, system and version, purpose, what was submitted and how output was verified. Morgan, Tym and Tate Group each demanded part of that information.
• Separate sessions by purpose. Villanueva divided one log into research, deposition preparation and drafting after the fact. Sessions kept apart from the start can simply be described.
• Export on a schedule. The preservation and production protocol calls for periodic exports in the provider's own format into the matter folder, so the record survives the account.
• Ask counsel about the protective order first. Clause 6, custody and confidentiality, settles what may be submitted before anything is. Jeffries, Morgan and Fry show the cost of asking late.
• Verify what a system summarizes. Villanueva ordered the case summaries produced partly so hallucinated citations could be found. Clause 3, independent authority, requires each characterization of a source to be confirmed against it.
• Log test runs and recipients. TRAX turned on output generated only to test a tool, and on who had seen it.

Expect, at minimum, questions about which systems were used, for what and on what material. Whether session content must also be produced has been answered differently, and retaining counsel is the right person to address it in a given matter. The expert-witness matters are updated as these orders move.

Matters this guide relies on

• Villanueva v. Las Vegas Metropolitan Police Department September 29, 2026 · Expert witnesses and AI Primary document read
• TRAX International Corp. v. United States September 22, 2026 · AI-generated and altered evidence Primary document read
• Watson Grinding explosion trial (3M expert) August 2026 · Expert witnesses and AI Reported; document not located
• AI clauses in protective orders (Jeffries; Dunn) August 31, 2026 · Rules, orders and guidance Primary document read
• Fry v. Fry (Claude upload and sealing) August 27, 2026 · AI-generated and altered evidence Primary document read
• Shealy v. Seaside Investments June 2026 · AI-generated and altered evidence Primary document read
• Assini v. Hayward (subpoena to OpenAI quashed) June 4, 2026 · AI-generated and altered evidence Primary document read
• Tate Group Automotive v. Legacy Automotive Capital June 3, 2026 · AI-generated and altered evidence Primary document read
• Conservation Law Foundation v. Shell (expert AI prompts) May 18, 2026 · Expert witnesses and AI Primary document read
• Tym v. Cerno (AI work product) April 22, 2026 · AI-generated and altered evidence Primary document read
• Jones v. Delta Air Lines (ChatGPT at a deposition) April 22, 2026 · AI inside the courtroom Primary document read
• Morgan v. V2X (AI work product) March 30, 2026 · AI-generated and altered evidence Primary document read
• United States v. Heppner February 17, 2026 · AI-generated and altered evidence Primary document read
• Warner v. Gilbarco (ChatGPT work product) February 10, 2026 · AI-generated and altered evidence Primary document read
• Matter of Weber October 10, 2024 · Expert witnesses and AI Primary document read

Questions