Your AI Chats Are Evidence: Litigation Holds Meet the Prompt Box
Team Sonomos · Sonomos
Team Sonomos writes about AI, privacy, and what happens to sensitive information at work.
In November 2025, FBI agents executed a search warrant at the home of Bradley Heppner, a former executive of GWG Holdings, and seized his documents and devices. Among them, his lawyers later told prosecutors, were about thirty-one documents recording conversations he had held with Claude. He had written them after receiving a grand jury subpoena, and they set out what he might argue about the facts and the law. His lawyers said the documents were privileged. In February 2026, Judge Jed Rakoff of the Southern District of New York ruled that they were not, on what he called a question of first impression nationwide.
Most of the commentary on the case was about privilege, which is the obvious lesson. The quieter lesson concerns what the documents were: a written record, made inside an AI tool, of a person thinking hard about a dispute. That is the raw material litigation runs on. Companies spent twenty years learning that email is evidence, then text messages, then Slack and Teams. AI conversations are next, and most companies stand roughly where they stood with text messages before regulators began fining firms over them. The material exists, it concerns the business, and it sits in accounts the company does not run.
Put plainly: once a dispute is reasonably foreseeable, the AI conversations in which employees did the company's work become evidence the company is expected to preserve and may have to produce, and because most of those conversations live in accounts the company cannot see or hold, the part of the exposure it can still control is decided at the moment of sending, by what went into the chat.
What the duty to preserve asks for
The federal rule is short. Under Federal Rule of Civil Procedure 37(e), if electronically stored information "that should have been preserved in the anticipation or conduct of litigation is lost because a party failed to take reasonable steps to preserve it," and it cannot be replaced through other discovery, a court may order measures to cure the prejudice. If the court finds the party acted "with the intent to deprive another party of the information's use in the litigation," it may presume the lost information was unfavorable, tell the jury so, or dismiss the case outright.
Two features of that text matter here. The duty begins with anticipation, which is usually well before anyone files a complaint: depending on the facts, a demand letter, a regulator's inquiry or a resignation that turns ugly can be enough. And the standard is reasonableness. The Advisory Committee's note says the rule "does not call for perfection," but it also tells judges to be "sensitive to the party's sophistication with regard to litigation." A company with in-house lawyers and a records team will get less patience than a sole trader.
An AI conversation is electronically stored information in the plainest sense. The prompt, any uploaded file and the model's answer are all stored somewhere, at least for a while. How long depends on the provider, the plan and the user's settings, which we track in Does ChatGPT save your prompts? and what Temporary Chat actually does. In an enterprise workspace, administrators may have export and retention tools, as we explain in Can my employer see my ChatGPT prompts?. In a personal account, the only person who can preserve the conversation, or delete it, is the employee.
The Google Chat case is the closest guide
The best indication of how a court will treat employee-controlled AI chats comes from a case about something else. In March 2023, Judge James Donato issued findings on chat preservation in the Google Play Store antitrust litigation, the consolidated cases brought by Epic Games, consumers, the attorneys general of 38 states and the District of Columbia, and others. Google's internal Chat had history off by default, and chats with history off were typically deleted after 24 hours. After the lawsuits began, Google did not suspend that deletion. It "chose instead to let employees make their own personal choices about preserving chats."
About 360 people were under the legal hold. Google had the technical ability to switch history on for all of them and did not. The court's account of what followed reads like a field study of human judgment under a hold notice. A former vice president who had received the hold notice testified that he kept history off for his entire time at Google and could not say in detail which topics were relevant. The court observed: "A lawyer working on this case would know that MADAs are relevant. Jamie Rosenberg did not." Another employee believed he had complied by copying the substance of chats into his own documents, in effect preserving his summary of the evidence instead of the evidence.
The judge set this against Google's handling of email, where custodians on hold "cannot override the automated preservation." For Chat, Google "gave each employee carte blanche to make his or her own call about what might be relevant." He called the result a "don't ask, don't tell" policy, found that Google had not taken reasonable steps, and found intent under Rule 37(e)(2). Part of that intent finding rested on Google's lack of candor with the court, so its facts were worse than most companies' will be. The reasoning about who decides is the part that carries over.
Personal AI accounts put a company in a harder position than Google's. Google could have flipped a switch. A company whose staff draft in their own ChatGPT or Claude accounts has no switch at all. A hold notice that tells people to "preserve relevant AI conversations" reproduces the structure Judge Donato rejected: the employee decides what is relevant, the employee decides what is kept, and the employee may not understand what the case is about.
Where the work is actually happening
None of this would matter if AI use at work were rare or visible. It is neither. In a global study by KPMG and the University of Melbourne, which surveyed more than 48,000 people in 47 countries between November 2024 and January 2025, 58 percent of employees said they intentionally use AI at work. Almost half admitted using it in ways that break company policy, including uploading sensitive company information to free public tools. Fifty-seven percent said they hide their use of AI. Only 40 percent said their workplace had any policy or guidance on generative AI.
Those figures describe the preservation problem precisely. Counsel collecting evidence interviews custodians about email, chat platforms and phones. An employee who hides AI use from a manager every day is unlikely to volunteer it in a custodian interview, and the questionnaire may not ask. That is how shadow AI, the use nobody approved or recorded, becomes a gap in a document production.
The content is the other half. Consider an HR manager who asks a chatbot to help word the performance review of an employee who later sues for discrimination. The drafts, and a prompt like "make this sound less like we are managing her out," are exactly what the plaintiff's lawyer will ask for. The same goes for a sales director rewriting a pricing email, or an analyst summarizing a complaint before the company responds to a regulator. People use AI for the passages they find hardest to write, and those tend to be the passages a dispute turns on.
Control is the unsettled question
Discovery under Rule 34 reaches material in a party's "possession, custody, or control." Whether an employee's personal AI account counts as within the company's control has no settled answer. The Sedona Conference's commentary on the question, republished for comment in January 2024, describes a split between federal circuits that ask whether a party has a "legal right" to obtain the material and those that ask whether it has the "practical ability" to get it.
The commentary's examples show which facts matter. One court applying the legal-right test refused to compel production of text messages on employees' personal phones, where the company had not issued the phones and the employees did not use them for work. The commentary also notes that no court has squarely held that the practical-ability test compels a company to produce data held by current employees, while warning that the test could put employers in the awkward position of asking for their staff's personal data.
One way to read this for AI: the employee who drafts client letters in a personal chatbot account fails the second condition in that phone case, because the account was used for work. A court could find the company controls material it cannot reach. And where it does not, the other side can subpoena the employee, or the provider. In the Heppner opinion, Judge Rakoff cited another judge in the same district who had observed that users do not have substantial privacy interests in conversations "which users voluntarily disclosed" to an AI platform and which the platform "retains in the normal course of its business."
Regulators have already run this experiment
Off-channel communication has a recent enforcement history. In September 2022 the SEC charged sixteen Wall Street firms, which agreed to pay combined penalties of more than $1.1 billion, after finding that their staff had "routinely communicated about business matters using text messaging applications on their personal devices." The Enforcement Director put the principle simply: "If there are allegations of wrongdoing or misconduct, we must be able to examine a firm's books and records." Those cases rested on recordkeeping rules specific to broker-dealers, so the liability does not transfer to every company. The pattern does.
The Justice Department has written that pattern into its guidance for prosecutors. The Evaluation of Corporate Compliance Programs, updated in September 2024, asks what communication channels a company and its employees use to conduct business, and "What preservation or deletion settings are available to each employee under each communication channel, and what do the company's policies require with respect to each?" It expects business communications to be "accessible and amenable to preservation by the company." The same document separately asks how companies manage the risks of artificial intelligence. A chatbot where employees draft business documents is a communication channel in every sense those questions care about.
Why the usual fixes fall short
A ban is the instinctive response, and the KPMG figures suggest what it produces. Use continues and goes quiet, which is the worst outcome for preservation, because the hidden channel is the one no hold notice names. Our AI acceptable use policy template is built on the assumption that people will use these tools whatever the policy says.
A rule telling employees to save their AI chats when a hold arrives recreates the Google arrangement. It depends on each person's judgment about relevance, and on their honesty about a habit they may have hidden for a year.
An approved enterprise tool with retention and export features helps more than anything else on this list, and companies that can buy one should. It preserves what happens inside it. It does nothing for the personal accounts that, as we discuss in the legal risk of unapproved AI tools, employees keep using when the approved option is slower or more restricted. Keeping everything forever also carries its own cost: every retained conversation is something to review, produce and secure.
Data loss prevention tools were built to watch files and email leaving the network, and text typed into a browser tab is a harder thing to see. Deletion is no fix either. Before a duty attaches, routine deletion is lawful. After it attaches, deletion is the problem Rule 37(e) exists to punish.
What remains is the content. A preserved or produced conversation is evidence of what an employee did, and nothing should change that. But a conversation in which the client's name, the patient's details and the account numbers were replaced before sending puts far less of other people's confidential information into a third party's systems, and later into a production set, a privilege log and an opposing lawyer's review platform. The duty to preserve stays the same. The harm that preservation and production spread is smaller.
Frequently asked questions
Do employees' AI conversations belong on a litigation hold?
If they are relevant to the dispute and the company can reasonably preserve them, they should be treated like any other electronically stored information. Hold notices should name AI tools explicitly, custodian interviews should ask about them, and enterprise workspaces should have retention set before the hold arrives. For personal accounts, the realistic step is a specific instruction not to delete, plus a record that it was given.
Is it spoliation to delete a chatbot conversation after receiving a hold notice?
It can be. Rule 37(e) applies to information that should have been preserved once litigation was anticipated. Deleting a relevant conversation after that point risks curative measures, and if a court finds intent to deprive the other side of it, an adverse inference or worse. Deleting old chats in the ordinary course before any duty arises is a different matter.
Are conversations with a public AI tool privileged?
In the most prominent federal decision so far, no. Judge Rakoff held in Heppner that a defendant's own exchanges with a consumer AI tool were neither privileged nor work product, partly because the provider's terms allowed disclosure to third parties. He suggested the answer might differ if a lawyer had directed the AI use. It is one trial court's view, and other courts may reach different results. Our explainer on AI and the work product doctrine covers the lawyer-directed case.
Does an enterprise AI plan solve the preservation problem?
It solves much of it for conversations that happen inside the enterprise workspace, provided retention is configured and someone knows how to export. It does not reach employees' personal accounts, which the surveys suggest many people still use.
Decide what goes in before a court decides what comes out
The practical program follows from the argument. Name AI tools in hold notices and custodian questionnaires. Give people an approved tool good enough that they stop reaching for a personal one, and set its retention deliberately. Accept that some work will still happen in accounts no one in the company can see, and treat every prompt as a document that may one day be read by a stranger with a subpoena. Under that assumption, the most useful habit is data minimization at the point of use: the conversation can still be helpful without the client's real name or the patient's real record in it. For firms where this is daily work, our legal use case shows how it plays out.
That last step is where Locke fits. It runs on your device, detects names, account numbers and other sensitive values in the AI tools it screens, and swaps them for placeholders before the prompt leaves, and the unmasked content is never sent to Sonomos. Download Locke and try it on the work you already do with AI.
Sonomos publishes this blog and makes Locke, which detects and masks sensitive information on your device before it reaches an AI model. How Locke works
This article is general information about privacy and AI, not legal advice, and reading it does not create a solicitor–client or attorney–client relationship. Regulatory requirements turn on your specific facts and jurisdiction and change over time. Consult qualified counsel before relying on anything here, and check the publication and last-reviewed dates above. We may revise posts, but we do not undertake to keep any post current as the law changes.
Reference pages for this post
Related Articles
The Breach Is Contractual First: Client Data in AI Tools
Before any regulator gets involved, pasting a client's data into an AI tool usually breaches the NDA, DPA or BAA it arrived under. Here is how those clauses read, why the client is the first to have a claim, and where the usual fixes leave gaps.
The Paste Is the Disclosure: The Legal Risk of Unapproved AI Tools
When an employee pastes company data into an AI tool nobody approved, the law does not ask whether AI was involved. It asks who you told. Here is how one paste can weaken a trade secret, waive privilege, or start a HIPAA or GDPR breach clock.
The Anthropic DPA Explained: Claude Data Processing Agreement Guide (2026)
Anthropic offers a DPA on Claude Team, Enterprise, and the API — not on Claude.ai Free or Pro. Here is what the Anthropic DPA covers, how it handles EU-US transfers via SCCs and the DPF, what to verify before signing, and where a DPA stops.