Litigation Holds Now Have to Cover AI Chat Histories

Litigation Holds Now Have to Cover AI Chat Histories

If your litigation hold procedures were last updated before 2023, they almost certainly have a gap. A big one. Employees across your organization are using ChatGPT, Copilot, Claude, Gemini, and a growing list of internal AI tools to draft memos, analyze contracts, summarize documents, brainstorm strategy, and process sensitive data. Those interactions generate chat histories that are, under the Federal Rules of Civil Procedure, electronically stored information. And ESI is subject to preservation obligations the moment litigation is reasonably anticipated.

This is not a theoretical concern. Courts are already dealing with it, and the companies that get caught flat-footed are going to face sanctions, adverse inference instructions, or worse.

The Preservation Obligation Has Not Changed. The Data Landscape Has.

The duty to preserve relevant ESI under FRCP Rule 37(e) kicks in when litigation is reasonably anticipated, not when a complaint is filed. The 2015 amendments to Rule 37(e) gave courts explicit authority to impose sanctions, including adverse inference instructions, when a party fails to take reasonable steps to preserve ESI and the lost information cannot be restored or replaced through additional discovery. The standard is "reasonable steps," which means your preservation efforts are measured against what a reasonable party would do given the circumstances.

In Zubulake v. UBS Warburg (S.D.N.Y. 2004), Judge Scheindlin established the foundational framework: once a duty to preserve attaches, a party must suspend routine document destruction and put in place a litigation hold to ensure relevant documents are preserved. That framework has been extended repeatedly to cover new categories of ESI as technology evolves, from email to text messages to Slack channels to ephemeral messaging apps.

AI chat histories are the next logical extension. And frankly, they may be more discoverable and more damaging than any of those prior categories, because people tend to be remarkably candid with AI tools. They paste in confidential documents. They ask for help crafting arguments they would never put in an email. They use AI as a thinking partner in ways that reveal intent, knowledge, and decision-making processes with unusual clarity.

What Counts as an AI Chat History?

Broadly, any record of a human-to-AI interaction that is stored in a retrievable format. This includes:

• Conversations with public LLMs like ChatGPT, Claude, and Gemini, where the provider retains chat logs (OpenAI retains them for 30 days by default, longer if the user has not opted out)
• Interactions with enterprise AI tools like Microsoft Copilot, which are typically logged within the Microsoft 365 compliance boundary
• Prompts and outputs from internally deployed models, including RAG systems that query proprietary data
• AI-assisted drafting histories in tools like Notion AI, Google Docs with Gemini, or contract lifecycle management platforms with embedded AI
• API call logs where employees or systems programmatically interact with AI models

The scope is wide, and it is getting wider. The Sedona Conference's 2024 commentary on generative AI and ediscovery flagged AI interactions as a distinct category of ESI requiring specific attention in preservation plans. That is about as close to an industry consensus as you get in this space.

The Scope of an AI Chat History Hold

A litigation hold notice needs to specifically identify AI chat histories as a category of information subject to preservation. Generic language about "electronic communications" probably covers it in theory, but courts have shown little patience for parties who rely on ambiguous hold notices and then claim they did not realize AI chats were included. See Steves and Sons, Inc. v. JELD-WEN, Inc. (E.D. Va. 2018) for a $5 million sanctions award rooted in part in inadequate hold implementation.

Your hold should address several specific dimensions:

• Custodian identification. Which employees used AI tools in connection with the subject matter of the litigation? This requires knowing who has access to what AI tools, which is itself a governance problem many companies have not solved.
• Platform identification. Which AI tools are in use across the organization, both sanctioned and unsanctioned? Shadow AI is a real preservation risk. If an employee used a personal ChatGPT account to analyze documents related to the dispute, that history is potentially discoverable.
• Temporal scope. AI chat histories may predate the triggering event. If an employee used an AI tool to help develop the product, strategy, or communication at issue, those earlier interactions are relevant.
• Content scope. Both prompts and outputs matter. The prompt often reveals what the user knew, what they were trying to accomplish, and what documents they had access to. The output shows what information informed their subsequent actions.

The Technical Mechanics Are Harder Than You Think

Preserving AI chat histories is technically more complex than preserving email or Slack messages, for several reasons.

Retention Policies Vary Wildly by Platform

OpenAI's data retention for ChatGPT API usage is 30 days by default, but ChatGPT consumer and Plus accounts retain conversation history indefinitely unless the user manually deletes it or disables history. Anthropic retains Claude conversations for a limited window. Google's Gemini interactions within Workspace are subject to Vault retention policies, but personal Gemini use is not. Microsoft Copilot interactions within M365 are captured by Purview, but only if you have the right licensing tier and have configured retention policies correctly.

The practical upshot: if you do not act quickly after a hold triggers, data may auto-delete before you can preserve it.

Export and Collection Tools Are Immature

Most ediscovery platforms have not yet built robust connectors for AI chat history collection. Relativity announced early support for Copilot data in 2024, but coverage across the broader AI tool landscape remains spotty. For many tools, you are looking at manual export, API-based extraction, or screen captures, none of which scale well.

Metadata Matters

A chat transcript alone may not be sufficient. You need timestamps, user identification, session identifiers, model version information, and ideally the system prompts or configuration that shaped the AI's behavior. Without metadata, opposing counsel will argue the preserved data is incomplete or unreliable.

The Trade Secrets Angle

This is where it gets particularly sensitive. In trade secret litigation under the Defend Trade Secrets Act (18 U.S.C. 1836), AI chat histories can be both a sword and a shield. If a departing employee pasted proprietary formulas, customer lists, or source code into an AI tool, those chat logs are direct evidence of misappropriation. Conversely, if your own employees used AI tools to process or develop trade secrets, the chat histories may reveal the scope of what you consider confidential and how you protected it, or failed to.

In Brightmetrics LLC v. Ooma, Inc. (W.D. Wash. 2024), the court considered whether AI-generated analyses of competitive data constituted protectable trade secrets. The underlying AI interactions were central to the dispute. Expect more of this.

Practical Steps to Take Now

• Inventory all AI tools in use across your organization, including unsanctioned ones. You cannot preserve what you do not know exists.
• Update your litigation hold template to explicitly reference AI chat histories, prompts, outputs, and associated metadata.
• Configure retention policies on enterprise AI tools (Copilot, Gemini for Workspace) to align with your standard preservation periods.
• Establish a protocol for preserving data from consumer AI tools when custodians are identified, including instructions for disabling auto-delete and exporting history.
• Brief your ediscovery vendors on AI data sources and confirm their collection capabilities.

How FirmAdapt Addresses This

FirmAdapt's architecture was designed with exactly this kind of compliance requirement in mind. All AI interactions within the platform are logged with full metadata, including timestamps, user identity, session context, and model configuration. These logs are retained according to configurable policies and can be exported in standard ediscovery formats, which means when a litigation hold triggers, your AI interaction data is already preserved, indexed, and ready for collection without scrambling to pull data from a dozen different consumer tools.

More fundamentally, FirmAdapt gives you a single, governed environment for AI use across your organization, which reduces the shadow AI problem that makes preservation so difficult in the first place. If your people are doing their AI work inside a platform built for regulated industries, you know where the data is, you control how long it is kept, and you can respond to a hold notice the same day you receive it.