Subpoenas and preservation

Can a Law Firm's AI Chats Be Subpoenaed?

Updated November 11, 2026

Short answer: What an AI provider stores can be sought through legal process, like other records, and a court can order a provider to keep data it would normally delete; that happened in NYT v. OpenAI, until the broad order ended on September 26, 2025, though data already kept, and some named accounts, stayed under preservation. What can be produced depends on what is stored, and where. A preservation order applies to data a provider keeps; a tool that clears client content leaves less kept data for an order to reach, but it does not protect privilege, and it does not change what your firm must keep in its own files. Whether a particular chat is privileged is a separate question that courts are only starting to answer.

Three places a chat can be sought from

When a lawyer uses an AI tool on a matter, the conversation can exist in up to three places:

  1. Your firm's own systems, if anyone saved, exported or copied it into the client file.
  2. The AI vendor, if its product stores prompts, documents and answers, as most legal platforms and business chat plans do.
  3. The AI model provider behind the vendor, unless a zero data retention agreement says it keeps nothing.

A subpoena or preservation order can be aimed at whoever holds the data. A tool that clears client content removes the second and third places once the content is cleared. It cannot remove the first, and should not, because your firm needs its records.

What NYT v. OpenAI showed

In the copyright case brought by The New York Times, a court ordered OpenAI in May 2025 to preserve user data it would normally delete. OpenAI says: "Our obligations under the earlier order ended on September 26, 2025." It says it still holds limited historical user data from April to September 2025 (OpenAI). The broad order ended then, though data already kept, and some named accounts, stayed under preservation (Duane Morris).

Who was covered at the time, in OpenAI's words: users of "ChatGPT Free, Plus, Pro, and Team subscription or if you use the OpenAI API (without a Zero Data Retention agreement)." Team was OpenAI's business plan at the time.

Who was not covered: "ChatGPT Enterprise or ChatGPT Edu customers", and API customers with zero data retention. Of those, OpenAI said: "we never retain the prompts you send or the answers we return. Because it is not stored, this court order doesn't affect that data."

Two lessons follow, and neither depends on the order still being in force:

  • A provider's normal deletion schedule can be overridden. A retention policy describes what a provider usually does. A court can order it to keep more.
  • Data that was never stored was not affected. OpenAI says the order did not touch zero data retention data because none was kept. An order applies to what exists, and a court could still order a provider to keep data from then on.

OpenAI was a party to the case, so its account is its own; the dates are checkable. Our companion site's Zero Data Retention AI guide covers the case in more depth.

Vendors' terms acknowledge legal process plainly. ChatGPT Enterprise keeps deleted conversations for up to 30 days, and longer where "we are legally required to retain them"; ChatGPT Business adds "or is reasonably necessary to protect our services or any third party from harm" (OpenAI). Harvey's public privacy policy reserves sharing with "law enforcement ... civil litigants" to "comply with ... court orders, subpoenas, legal process" (Harvey).

None of this is unusual. Any company that holds data can be asked to produce it. ABA Formal Opinion 512 says lawyers should make sure they "will be notified in the event of a breach or service of process regarding production of client information" (p. 11), which is why the vendor questions include it.

Is a chat with an AI tool privileged?

Courts have barely started on this. In United States v. Heppner (S.D.N.Y., February 2026), one federal trial court reportedly held that a criminal defendant's own conversations with a consumer AI tool were protected by neither attorney-client privilege nor work product, partly because the provider's privacy policy allowed disclosure. The ruling is disputed, it comes from a single trial court, and, as reported, it did not decide how enterprise tools, or AI used at a lawyer's direction, should be treated. Our account comes from law-firm and law-review commentary, not the court's own memorandum.

We do not repeat the analysis here. Our companion site, AI Ethics for Lawyers, covers Heppner and what it does and does not decide, with its sources.

What matters for choosing a tool is narrower: no AI tool, however it is built, makes a conversation privileged. Privilege depends on the law and the facts, not on the vendor.

What a tool that clears content does and does not do

Some tools clear client content rather than storing it. Faradex holds files and conversations in memory, and they are gone when the Workspace is reset. Its AI model: Currently Opus via AWS Bedrock, more information is available on our website. AWS does not store prompts or responses, and model providers have no access to them. Proton says of Lumo for Business: "After Lumo processes your query and generates a response, the data is erased" (Proton).

What it does: once the content is gone, there is no copy at the provider to preserve, produce or lose in a breach. A preservation order applies to data a provider keeps, so there is less kept data for an order to reach. While content is still held, for example in memory before a reset, it exists.

What it does not do:

  • It does not make anything privileged, or protect privilege that would otherwise be lost.
  • It does not stop legal process against your firm. Anything your firm saved can be sought from your firm, as before.
  • It does not change your firm's own duties to keep records, or to preserve what it holds when litigation is expected. Ask your counsel how those duties apply to AI work product.
  • It does not mean nothing exists. Whatever your lawyers saved still exists, and a provider may keep account and billing records. Clearing content is a statement about conversation content at the provider.

Where the common tools sit

Tool type What exists at the provider Example terms
Tools that clear client content Held in memory, gone on reset; or erased after each response Faradex: held in memory until the Workspace is reset. Lumo: "the data is erased" after each response
Legal platforms with zero retention upstream Your content, at the vendor, often with backups; nothing at the model provider CoCounsel: US backups; Clio: "regular automatic data backups"; Spellbook: stored in Canada and the US
Business chat plans Your chats, until deleted, then up to 30 days, with legal exceptions ChatGPT Business and Enterprise; Claude Enterprise "retained indefinitely unless a custom retention period is set"
AI inside Microsoft 365 Prompts and responses, stored with your other Microsoft 365 content Copilot: retention set by your admin in Purview

Our reading of the last row: because Copilot stores interactions with your firm's other Microsoft 365 content, they are within reach of the same e-discovery your firm runs or is ordered to run. That can be useful for your records. It also means AI history is part of what can be produced.

Your firm's own record duties

A tool that keeps less makes your own file more important. Firms often ask what happens to the record if the AI tool does not keep it. The answer is that the record is whatever your firm saves: the advice sent to the client, the draft that was filed, the research memo in the matter file. Decide before you start:

  1. What must be kept, under your rules and your firm's policies, when AI helps produce it.
  2. Where it goes: your document management or practice-management system, the same as work done without AI.
  3. Who saves it, so nothing depends on a chat history that may not exist later.

Practical steps

  • Know what your tool keeps, and where. Use the vendor questions, especially questions 2, 3 and 5.
  • Ask to be told about legal process. Put notice of subpoenas and preservation demands in the contract where you can.
  • Do not count on any tool for privilege. Treat AI conversations as you would any other communication with a third-party service.
  • Keep the record in your own systems, whatever the tool keeps.
  • Talk to clients about the tools you use. Our companion site covers when consent is needed.

This guide is general information, not legal advice. Whether a communication is privileged, and what your firm must preserve, depends on your jurisdiction and your facts.

About Faradex. Faradex is a private AI workspace for small and mid-sized firms that handle confidential client work. Each firm gets its own Workspace on dedicated infrastructure in the United States. Files and conversations are held in memory and are gone when the Workspace is reset. AI model: Currently Opus via AWS Bedrock, more information is available on our website. AWS does not store prompts or responses, and model providers have no access to them. Pricing: from $120 a month, no seat minimum; see faradex.ai/pricing.