Do litigation holds need to cover AI chat logs in 2026?

The direct answer: yes, if the logs may matter

By September 16, 2026, the question is no longer whether AI chat logs can be discovered. The better question is whether your litigation hold captures them before they vanish. Law360's headline put it plainly: Your Next Litigation Hold Should Cover AI Chat Logs. JD Supra reached the same conclusion: AI Chats Are Discoverable—And Trigger Preservation Obligations. Under Federal Rule of Civil Procedure 34, discoverable electronically stored information includes any stored data that can be retrieved and translated. Prompts, model outputs, timestamps, account identifiers, and attached files in ChatGPT, Microsoft Copilot, Google Gemini, Anthropic Claude, or an internal large language model all qualify. If a chat log contains facts about a dispute, a party's state of mind, a product defect, an employment decision, or a contract negotiation, it is relevant. Relevance does not require the log to be admissible at trial. It only requires that the log may lead to admissible evidence. The duty to preserve attaches when litigation is reasonably anticipated, not when a complaint is filed. That duty runs to the client, not just outside counsel. A litigation hold that ignores AI chat logs is incomplete in 2026.

Also worth reading: How do defensible AI eDiscovery audit logs work and why are they essential for modern litigation? · how to draft an AI eDiscovery litigation hold notice? · How do legal teams build defensible AI discovery workflows for litigation in 2026?

Why AI chat logs are ordinary ESI with unusual risks

Courts have treated chat logs as ESI for years. Chelsea Manning's chat logs were published by Wired in 2011. WhatsApp chat logs have been used in criminal and civil cases. What changed after 2022 is the volume and location of AI conversations. Employees now use chatbots for drafting, coding, research, and legal questions. Some of those conversations contain candid admissions. In one widely reported criminal case, a defendant consulted ChatGPT about weapons and timing before an attack. The chatbot logs became evidence. In civil litigation, plaintiffs have sought AI prompts to show what a defendant knew and when. The risk is not only content. AI platforms often store metadata: model version, session ID, IP address, device, and deletion settings. That metadata can authenticate a log or show spoliation. Unlike email, AI chats may live in a vendor cloud outside the company's direct control. A user can delete a conversation with one click. An administrator can change retention settings. A vendor can update terms of service. Those features make preservation urgent. The 2015 amendments to Rule 37(e) provide a safe harbor only if a party took reasonable steps to preserve ESI. Waiting for a preservation letter is not reasonable.

How AI chat logs compare with email, texts, and ephemeral messages

The table shows why a litigation hold cannot treat AI chat logs as a minor add-on. Email preservation is mature. Most organizations have journaling, legal hold tools, and retention policies. AI chat logs are newer. Microsoft 365 Copilot logs may sit in Exchange or SharePoint. ChatGPT Enterprise logs may sit in a separate admin console. Consumer ChatGPT accounts may not be visible to the employer at all. Ephemeral messages like Signal or WhatsApp disappearing messages create similar problems, but AI logs often contain richer reasoning. A single prompt can reveal a decision-maker's assumptions. A single output can show what advice was received. That makes AI chat logs more like a memo than a text. The comparison also shows cost. Preserving email at scale is expensive. Preserving AI logs may be cheaper per custodian, but the privilege review is harder. Courts are split on whether chats with a public AI tool are confidential. That split increases the cost of review because every log may need a privilege call.

Practical steps to issue and manage a litigation hold for AI chat logs

First, identify the AI systems in use. That includes sanctioned tools like Microsoft 365 Copilot, Google Workspace Gemini, ChatGPT Enterprise, Claude for Work, and internal models. It also includes shadow AI: personal ChatGPT accounts, browser extensions, and AI features inside Slack, Zoom, and Adobe. Second, name custodians. The hold should cover employees who used AI for the disputed subject matter. That may include sales, engineering, HR, finance, and legal. Third, suspend auto-deletion. For ChatGPT Enterprise, an admin can disable chat history off or set retention to indefinite. For Copilot, preserve the underlying mailbox and SharePoint content. For consumer accounts, ask the custodian to export data and stop deleting. Fourth, issue a written hold notice. The notice should define AI chat logs, give examples, and tell custodians not to delete, edit, or use disappearing-message features. Fifth, collect. Use vendor export tools, eDiscovery APIs, or forensic imaging. Capture native format, metadata, and thread context. Sixth, document. Keep a preservation log with dates, custodians, systems, and actions. Seventh, repeat. AI tools change monthly. A hold issued in January 2026 may miss a tool adopted in March. Review the hold every 90 days. These steps are not optional. Rule 37(e) sanctions can include adverse inference instructions, monetary penalties, and case-ending dismissal. In 2026, courts expect more than a generic email hold.

Common mistakes that create spoliation exposure

The first mistake is assuming AI chats are privileged. Attorney-client privilege protects confidential communications between a client and a lawyer for legal advice. A chat with ChatGPT is not a lawyer. A chat with an AI legal research tool may be work product if it was prepared in anticipation of litigation, but courts are split. The CDF Labor Law LLP article notes that courts disagree on attorney-client and work-product protection for AI chats. The Ogletree cautionary tale shows how a company's AI use can waive privilege if the tool shares data. The second mistake is relying on the vendor's default retention. Many AI platforms delete chats after 30 days unless the user opts out. A hold that does not change that setting is ineffective. The third mistake is preserving only outputs. Prompts show intent. A plaintiff will want the prompt that asked how to hide a payment. The fourth mistake is over-preserving. Keeping every AI chat from every employee for years increases cost and review burden. The duty is to preserve relevant information, not all information. The fifth mistake is ignoring metadata. A PDF export of a chat loses timestamps and session IDs. The sixth mistake is not updating the hold after a new tool is deployed. The seventh mistake is failing to interview IT and the AI vendor. The vendor may have logs the company cannot see. A short vendor preservation letter can close that gap.

Privilege and work product: the unsettled law in 2026

Privilege is the hardest issue. In 2023 and 2024, courts began to address whether using AI waives privilege. Some courts held that communicating with a public AI tool is like speaking in public. Other courts held that the communication remains confidential if the user had a reasonable expectation of privacy and the tool's terms limit training use. The William Smullen article on AI tools, legal advice, and the limits of attorney-client privilege explains that the privilege depends on confidentiality, not the medium. The CDF Labor Law LLP article says courts split on attorney-client and work-product protection. By September 2026, no Supreme Court or uniform appellate rule exists. That means a litigation hold should preserve AI chats but also flag them for privilege review. Do not produce them automatically. Do not delete them automatically either. Work product protection under Rule 26(b)(3) may cover AI chats created by a lawyer or at a lawyer's direction. But if the client used AI before consulting counsel, work product may not attach. The practical answer is to segregate AI chat logs, log them in a privilege log, and let the court decide. That approach costs more upfront but avoids waiver and spoliation at the same time.

Cost and pricing for AI chat log preservation and review

Costs vary by system, custodian count, and data volume. Native preservation is often free if the vendor provides an admin export. Third-party eDiscovery processing for AI chat logs ranges from $5 to $25 per gigabyte. Review costs range from $1 to $5 per document for technology-assisted review, and $5 to $15 per document for linear attorney review. A mid-size case with 50 custodians and 200 gigabytes of AI chat logs might cost $10,000 to $50,000 for processing and $50,000 to $250,000 for review. A small case with 5 custodians and 10 gigabytes might cost $2,000 to $10,000 total. Legal hold software from Exterro, Onna, Relativity, or Logikcull can cost $5,000 to $50,000 per year for mid-market companies. AI eDiscovery tools can reduce review cost by 50% to 80% compared with manual review, but they require validation. The cheapest option is often the most expensive: a vendor default deletion that destroys a key chat log. Rule 37(e) sanctions can exceed the entire review budget. For legal drafting, AI tools can produce a hold notice in minutes, but a lawyer must verify the scope. Legalpdf.io and similar platforms can help with document drafting, but they do not replace a preservation plan.

When to act: triggers and timelines

The duty to preserve arises when litigation is reasonably anticipated. Triggers include a demand letter, an EEOC or state agency charge, a regulatory subpoena, an internal investigation, a whistleblower complaint, a preservation letter from opposing counsel, or a credible threat of litigation. In 2026, a social media post or a customer complaint can trigger the duty if it signals a real dispute. Once triggered, act within 24 to 72 hours. Issue the hold notice, suspend auto-deletion, and send vendor preservation letters. Do not wait for the Rule 26(f) conference. That conference usually happens 21 to 30 days after the complaint, but the duty predates it. If you wait, the AI vendor may have already deleted the logs. For ongoing litigation, review the hold every 90 days and after any new AI tool deployment. If a custodian leaves the company, preserve their AI account before deactivation. If a vendor changes terms, reissue the hold. The timeline is not a suggestion. Courts have imposed sanctions for delays of weeks, not months. In one 2024 case, a court found that a party failed to preserve text messages and gave an adverse inference. AI chat logs are next.

How AI eDiscovery and legal drafting tools fit

AI eDiscovery tools can help with AI chat logs. They can collect from APIs, reconstruct threads, deduplicate prompts, and identify privileged content. They can also translate model outputs into review-ready formats. But these tools are not magic. They need a preservation plan first. If the logs are gone, no AI can review them. Legal research and document drafting tools can help write the hold notice, the vendor letter, and the privilege log. They can also summarize chat logs for a privilege review. The site angle for legalpdf.io is not to sell a product but to explain the workflow. A good workflow uses AI for speed and a lawyer for judgment. The 61% of federal judges using AI, reported by Platinum IDS, shows that courts are familiar with the technology. That familiarity cuts both ways. Judges may expect parties to understand AI logs. They may also be less sympathetic to excuses about technical complexity. The best approach is to treat AI chat logs as a known ESI category, not a novelty.

The bottom line for 2026

A litigation hold in 2026 should cover AI chat logs when those logs may contain relevant information. The duty is triggered by reasonable anticipation of litigation, not by a formal complaint. The logs are discoverable ESI under Rule 34. The failure to preserve them can lead to Rule 37(e) sanctions. Privilege is unsettled, so preserve and review rather than delete or produce. Practical steps include identifying AI systems, naming custodians, suspending auto-deletion, issuing a written hold, collecting native data with metadata, and documenting every action. Costs range from a few thousand dollars for small cases to hundreds of thousands for large ones. The biggest mistake is assuming the vendor will keep the data. The second biggest mistake is assuming the chats are privileged. The third is waiting for a preservation letter. Act within 72 hours of a trigger. Review the hold every 90 days. Use AI eDiscovery and legal drafting tools to speed the work, but keep a lawyer in the loop. That is the definitive answer as of September 16, 2026.