Litigation Discovery Requests for AI Tool Logs Are Now a Thing

Litigation Discovery Requests for AI Tool Logs Are Now a Thing

A colleague forwarded me a set of discovery requests from a breach of fiduciary duty case last month. Buried in the document requests, right between the usual email preservation demands and Slack channel exports, was this: "All prompts, inputs, outputs, and interaction logs from any artificial intelligence or large language model tools used by Defendant or its agents in connection with the matters at issue." The requesting party also asked for metadata, timestamps, user identification, and any system prompts or configuration settings applied to those tools.

This is not an outlier. It is becoming standard practice, and the implications for preservation obligations, production workflows, and privilege are significant.

Where This Is Coming From

Federal Rule of Civil Procedure 34(a)(1) has always allowed parties to request "electronically stored information" (ESI) in any medium. Courts have consistently interpreted ESI broadly. AI prompt and response logs fit comfortably within that definition, and litigators know it.

The catalyst was partly the judiciary's own attention to AI. Since Judge Brantley Starr's standing order in the Northern District of Texas (May 2023) requiring attorneys to certify that AI-generated content was verified by a human, courts have been increasingly aware that AI tools generate discoverable artifacts. The Sixth Circuit's Mata v. Avianca fallout, where attorneys submitted fabricated ChatGPT citations, made judges hyperaware that AI usage leaves a trail, or should.

By mid-2024, several Magistrate Judges in the Southern District of New York and the District of Delaware were addressing AI-related discovery disputes in meet-and-confer conferences. In Herrera v. Acuity Brands (N.D. Ga. 2024), the court addressed whether AI-assisted document review logs were discoverable, noting that the methodology itself could be relevant to completeness challenges. And in January 2025, a standing order from the Eastern District of Pennsylvania explicitly referenced AI tool outputs as ESI subject to standard preservation obligations under FRCP Rule 37(e).

The Preservation Problem

Rule 37(e) imposes sanctions when a party fails to take reasonable steps to preserve ESI that should have been preserved in anticipation of litigation. The advisory committee notes make clear that "reasonable steps" does not mean perfection, but it does mean having a system in place.

Here is where most organizations are exposed. Many AI tools, particularly SaaS-based LLM interfaces, do not retain interaction logs by default, or they retain them only for limited periods. OpenAI's ChatGPT, for example, allows users to disable chat history entirely. Microsoft Copilot logs vary depending on tenant configuration. Anthropic's Claude retains conversation data according to its retention policy, which is not designed around litigation hold requirements.

If your organization uses AI tools and a litigation hold trigger occurs, you now need to consider whether AI interaction logs are within scope. For most commercial disputes, employment cases, and IP matters, the answer is increasingly yes. The duty to preserve attaches when litigation is reasonably anticipated, per Zubulake v. UBS Warburg (S.D.N.Y. 2003), and that duty extends to all relevant ESI, including AI logs that may show how decisions were made, what information was considered, or how work product was generated.

The practical problem: if you do not have logging enabled, or if your AI vendor purges logs on a 30-day cycle, you may face a spoliation argument before you even realize the data was relevant.

Production Challenges

Assuming you have the logs, producing them raises its own issues. FRCP Rule 34(b)(2)(E) requires production in the form in which ESI is "ordinarily maintained" or in a "reasonably usable" form. AI interaction logs can be messy. They may include system prompts that reveal proprietary configurations, retrieval-augmented generation (RAG) source references, chain-of-thought reasoning artifacts, or embedded context from prior conversations.

A few specific challenges worth flagging:

• Format variability. Some platforms export logs as JSON, others as plaintext, others as structured API call records. Opposing counsel may not accept raw JSON as "reasonably usable," and you may need to convert or annotate.
• Scope creep. A request for "all AI interaction logs" could sweep in thousands of irrelevant interactions. Proportionality under Rule 26(b)(1) is your friend here, but you need to be able to articulate why a narrower production is appropriate.
• Mixed content. A single AI conversation might contain both responsive and non-responsive material, or a mix of privileged and non-privileged content. Redaction protocols for AI logs are not yet standardized, and courts have not issued much guidance.

The Privilege Question

This is the part that keeps general counsel up at night. If an attorney uses an AI tool to draft a legal memorandum, analyze case strategy, or evaluate risk, the prompts and outputs may be protected by attorney-client privilege or work product doctrine. But the protection is not automatic.

Work product protection under Rule 26(b)(3) applies to materials prepared "in anticipation of litigation." If an attorney's AI prompts reflect mental impressions, conclusions, or legal theories, they should qualify as opinion work product, which receives near-absolute protection under Hickman v. Taylor (1947). But if the AI tool is used for routine business tasks, like summarizing a contract or generating a compliance checklist, the work product argument weakens considerably.

Attorney-client privilege is even trickier. The privilege requires a confidential communication between attorney and client for the purpose of legal advice. When an attorney inputs client confidential information into a third-party AI tool, there is a real question about whether confidentiality has been maintained. Some courts have held that disclosure to a third party waives privilege unless the third party is a functional equivalent of an employee or agent. Whether an AI vendor qualifies depends on the terms of service, data handling practices, and whether the vendor can access or train on the inputs.

The ABA's Formal Opinion 512 (July 2024) addressed this partially, concluding that attorneys must exercise reasonable care when using AI tools and should evaluate whether use of a particular tool risks waiver. But "reasonable care" is doing a lot of work in that opinion, and the boundaries are still being litigated.

Practical Steps Right Now

If you are responsible for compliance or legal operations at a regulated company, a few things are worth doing immediately:

• Audit your AI tool inventory. Know which tools are in use, who is using them, and what logging capabilities exist. Shadow AI is a discovery liability.
• Enable and centralize logging. If your AI tools support interaction logging, turn it on. Route logs to a system that supports litigation hold and retention policies.
• Update your litigation hold procedures. Explicitly include AI interaction logs in your hold notice templates and custodian questionnaires.
• Establish privilege protocols. Create clear guidelines for when and how attorneys may use AI tools for privileged work, and ensure those tools are configured to minimize waiver risk.
• Negotiate discovery scope early. In meet-and-confer sessions under Rule 26(f), raise AI logs proactively. It is better to define scope collaboratively than to fight about it in front of a Magistrate Judge.

How FirmAdapt Addresses This

FirmAdapt was built with the assumption that every AI interaction at a regulated company is potentially discoverable. The platform maintains comprehensive, immutable interaction logs with timestamps, user attribution, and full prompt-response pairs. These logs are stored in a format designed for eDiscovery export, with support for litigation holds, retention policy enforcement, and role-based access controls that preserve privilege designations.

Because FirmAdapt operates within your compliance boundary rather than routing data to external model providers without controls, the privilege and confidentiality analysis is significantly cleaner. When opposing counsel requests AI tool logs, you have a defensible, organized record rather than a scramble across six different SaaS platforms with inconsistent retention settings. The logging infrastructure is not an add-on; it is core architecture.