AI chat logs, privilege and discovery: nine 2026 decisions and what they mean in NSW
Between January and June 2026, nine United States courts ruled on whether a litigant's AI conversations can be discovered, subpoenaed or read in evidence. None binds an Australian court. The same questions are arriving here, and our rules answer some of them and not others.
Four answers first.
A client's own ChatGPT conversations about their case are probably not privileged in NSW. Treat them as discoverable and liable to subpoena.
In all six privilege cases, protection turned on whether a lawyer directed the AI use.
No Australian court has decided the privilege question. The one Australian passage on it is obiter, and it points the same way.
An expert's AI prompts are discoverable. In NSW they also need prior leave and an annexed record of the prompts.
The split turns on who directed the use
In United States v Heppner (SDNY, 17 February 2026) a represented securities fraud defendant ran his own defence strategy through Claude after receiving a grand jury subpoena. Rakoff J held the conversations were neither privileged nor work product. Claude is not a lawyer, the provider's terms allowed training on inputs and disclosure to third parties, and counsel "did not direct" the searches. Handing the outputs to his solicitors afterwards did not change their character. Shealy v Seaside Investments (Mass Super Ct, June 2026) is to the same effect, on facts involving a represented plaintiff whose partner ran his documents through ChatGPT.
Four decisions went the other way, and every one involved a self-represented litigant. Warner v Gilbarco (ED Mich, 10 February 2026) refused to compel AI records because "ChatGPT (and other generative AI programs) are tools, not persons". Morgan v V2X, Tate Group Automotive v Legacy Automotive Capital and Assini v Hayward, which quashed a subpoena issued to OpenAI, followed.
Every case that protected the material involved a litigant with no lawyer to direct anything. Every case that refused protection involved a client who had one and went around them.
In NSW a client's own AI chats are probably not privileged
The protecting decisions relied on rules we do not have. FRCP 26(b)(3) and Texas Rule 192.5(a)(1) protect material prepared in anticipation of litigation "by or for a party", which Judge Dorfman held covered a non-lawyer principal's ChatGPT conversations. There is no Australian equivalent. Sections 118 and 119 of the Evidence Act 1995 (NSW) each require a lawyer, as does the dominant purpose test in Esso. So Heppner and Shealy are the decisions whose reasoning transfers, and both refused protection. Heppner also raises waiver, since s 117 requires an obligation not to disclose before a communication is confidential and s 122(2) covers inconsistent conduct.
The Australian position
No Australian court has ruled on the point. In Helmold & Mariya (No 2) [2025] FedCFamC1A 163 the Full Court of the Federal Circuit and Family Court said at [9]:
input of documents arising out of the proceedings into a generative AI program which stores, collates and replicates data may waive privilege or fall foul of the requirements that certain matters be treated as commercial in confidence. These issues warrant extreme caution.
The Court also said that inputting court documents into an open AI program may fall foul of the publication prohibitions in Pt XIVB of the Family Law Act 1975 (Cth). That is obiter, in a judgment about fabricated citations, but it is the Heppner reasoning reached independently.
Everything else Australian courts have said about AI is about verification. In May v Costaras [2025] NSWCA 178 Bell CJ held at [15] that self-represented litigants are subject to SC Gen 23 whether or not they know its terms, and described verification at [17] as an "absolute necessity". Practitioners have been referred to their regulators in Valu (No 2), Mertz (No 3) and Dayal.
Where NSW is ahead
Paragraph 9A of Practice Note SC Gen 23 provides that material subject to a suppression order, the implied Harman undertaking or a statutory publication prohibition, and material produced on subpoena, must not go into a Gen AI program unless the practitioner is satisfied it stays in a controlled environment, is contractually protected from release, is used only in that proceeding and does not train the model.
Two US courts spent this year improvising a weaker version, the magistrate judge in Morgan writing it into the protective order from scratch. Paragraph 9A has applied since 3 February 2025. If you rely on enterprise terms to satisfy yourself under it, keep those terms on the file.
Expert prompts are discoverable
Conservation Law Foundation v Shell Oil (D Conn, 18 May 2026) held that an expert's AI prompts fall within the scope of discovery, because culling a document production down to a working subset forms part of the expert's methodology.
NSW is stricter already. Paragraphs 20 to 22 of SC Gen 23 require prior leave, disclosure of which parts of a report were prepared with Gen AI, and an annexed record of the prompts and variables used. Put it in the letter of instruction.
The logs are evidence
In Fortis Advisors LLC v Krafton, Inc. (Del Ch, 16 March 2026) a chief executive facing a projected earnout of about US$191.8m asked ChatGPT for help and received a "pressure and leverage package". Vice Chancellor Will found Krafton followed most of the recommendations, held the terminations pretextual and reinstated the founder. A footnote records that he admitted deleting the relevant logs.
Providers keep this material and courts will order its production. In In re OpenAI Copyright Infringement Litigation (SDNY, 5 January 2026) Stein J affirmed orders compelling a sample of twenty million consumer conversations. Destroying material with intent to mislead a judicial tribunal is an offence under s 317 of the Crimes Act 1900 (NSW) carrying ten years, and a litigation hold that does not name AI prompts and outputs will not catch them.
What to do
Tell clients in writing not to discuss the matter with any AI tool unless instructed. Name AI prompts and outputs in litigation holds, including personal accounts and deletion defaults. Put SC Gen 23 in expert instructions and seek leave first. Keep the documentation paragraph 9A requires.