Are AI Chat Logs Discoverable in Court? What the 20-Million-Log Ruling Means for You

A federal judge just confirmed what most people using AI at work haven't considered: your chat logs are electronically stored information, and they are producible in litigation like any other business record. If you've been wondering whether AI chat logs are discoverable in court, the answer, as of January 2026, is yes, concretely so, with a 20-million-conversation discovery order to prove it. This piece breaks down what the ruling established, what came after it, and what it means if you type anything into an AI tool on company time.

Key Takeaways

• A U.S. District Court in the Southern District of New York upheld orders requiring the production of 20 million de-identified chat logs from a major AI platform, treating them as standard discoverable ESI (electronically stored information).
• The court's core reasoning: users "voluntarily submitted their communications" to a third party. That single finding collapses most privacy objections before they start.
• Subsequent rulings have split on whether AI chat histories qualify as privileged work product, meaning protection depends on jurisdiction, context, and how you used the tool, not on any blanket rule.
• A preservation order earlier in the same litigation required the AI provider to retain every conversation log, including ones users had deleted, for months. Your "delete" button is not a legal guarantee.
• Employers and employees are now both exposed: AI chat logs are surfacing in copyright cases, employment disputes, and corporate governance litigation alike.

What Happened in January 2026?

On January 5, 2026, U.S. District Judge Sidney H. Stein of the Southern District of New York upheld two discovery orders requiring the production of a sample of 20 million de-identified user logs from a major AI chat platform. The case arose from copyright litigation brought by news organizations, including the New York Times, and a class of plaintiffs alleging that the AI system reproduced their copyrighted content during user conversations.

The plaintiffs originally wanted more. In July 2025, they moved to compel a sample of 120 million logs. The AI provider countered with 20 million, calling it 0.5% of total logs and "surely more than enough." The plaintiffs agreed to the smaller number. Then the provider changed course in October 2025, proposing instead to run keyword searches and produce only conversations that specifically implicated the plaintiffs' works.

The court rejected that narrowing. Judge Stein found no case law requiring the court to order the least burdensome form of discovery possible. The provider proposed the number. The plaintiffs accepted it. The provider was held to it.

Why Did the Court Reject Privacy Objections?

Because the users voluntarily submitted their communications to a third party. That was the finding that did the most damage to any expectation of privacy. The court drew a clear line between AI chat users and wiretap subjects: wiretap targets don't choose to have their communications intercepted. Chat users type their inputs into someone else's system on purpose.

This is not a novel legal theory. It is the third-party doctrine applied to a new medium. The same logic has governed email stored on third-party servers, documents in cloud storage, and messages sent through enterprise platforms for decades. The court simply confirmed that AI chat logs sit in the same category.

Privacy concerns were not dismissed outright. Judge Stein acknowledged them but found them adequately addressed by three procedural safeguards: reducing the sample from tens of billions of logs to 20 million, the provider's de-identification process stripping personally identifiable information, and an existing protective order governing how discovery materials could be used. Privacy was a process question, not a blocking objection.

Did Users Get Any Say in This?

No. The individuals whose conversations were swept into the 20-million-log sample received no notification and had no opportunity to object. They were not parties to the litigation. Their conversations were selected as part of a statistical sample. The protective order and de-identification were deemed sufficient protection on their behalf.

This is worth sitting with. If you used a major AI chat product during the relevant time period, a slice of your conversation history may have been produced to opposing counsel in a copyright case you had nothing to do with. De-identified, yes. Under seal, yes. But produced.

What About the "Delete" Button?

Before the January ruling, a related preservation order had already undermined it. On May 13, 2025, Magistrate Judge Wang issued a preservation order requiring the AI provider to retain every conversation log, including ones users had deleted, until further notice. That order was lifted in late September 2025, after which the provider resumed its standard 30-day deletion policy for new conversations.

During that roughly four-and-a-half-month window, conversations users believed they had deleted were preserved and retained regardless of user action. The delete button was still there in the interface. It just didn't do what users thought it did.

This is not a bug specific to one provider. Litigation holds override deletion policies across every category of digital service. If you store data with a third party and that third party becomes subject to a preservation order, your deletion instructions become legally irrelevant for the duration of the hold. The interface won't tell you this is happening.

How Are Courts Handling AI Chat Logs After January?

The January ruling opened a period of rapid, sometimes contradictory judicial activity. Four significant rulings in roughly four months, as one law firm noted, all testing how existing privilege and work-product doctrines apply when litigants use AI tools.

On February 10, 2026, two federal courts issued opposite conclusions on the same day. In Warner v. Gilbarco, Inc., a Michigan court found that a pro se plaintiff's use of an AI chat tool to help prepare her case was protected as work product and not discoverable. Hours apart, in United States v. Heppner, Judge Jed S. Rakoff in the Southern District of New York ruled that documents a criminal defendant created using an AI tool and later shared with his attorneys were not protected by attorney-client privilege or the work product doctrine.

Same day. Opposite holdings. Different facts, different circuits, different judges. If you're looking for a clear rule on whether your AI chat history is privileged, there isn't one yet. The legal framework is being assembled in real time, case by case.

By June 2026, a Texas Business Court protected a non-lawyer's AI chat conversations as work product while still requiring disclosure of what had been shared with the tool. The pattern emerging is not "AI chats are always discoverable" or "AI chats are always privileged." It is: courts will analyze each situation under existing doctrine, and the outcome depends on who used the tool, for what purpose, and whether the content was shared with others.

How Are AI Chat Logs Showing Up in Business Disputes?

Copyright litigation was the catalyst, but the implications have spread well beyond it. On March 16, 2026, the Delaware Court of Chancery issued an opinion finding that the CEO of a major gaming company had followed guidance from an AI chat tool to breach a $500 million acquisition agreement. The CEO's chat logs appeared directly in the judicial opinion. This was not a privacy case or a copyright case. It was a high-stakes corporate governance dispute, and the AI chat history served as evidence of the CEO's intent and decision-making process.

That is the scenario most people have not internalized. Your AI chat history can be evidence of what you knew, what you were thinking, and what advice you sought, just like email, just like Slack, just like search history. Except people tend to be more candid with AI tools than with any of those other channels. You ask an AI things you wouldn't put in an email to a colleague. Courts can now read those questions.

What Does This Mean for Employers Specifically?

Employment lawyers are flagging AI tools as a new discovery front. AI-generated ESI, from meeting notetakers, auto-drafted emails, chat assistants, and meeting summaries, is becoming a core battlefield in employment cases. The exposure is symmetric: these logs can help an employer's defense (contradicting an employee's testimony about what happened in a meeting) or devastate it (revealing what leadership actually said, asked, or knew).

Legal commentators are now advising employers to treat AI interactions less like private notes and more like potential business records, because that's how courts are likely to treat them. The National Law Review's analysis put it plainly: for in-house counsel, this ruling reinforces that AI conversation logs are discoverable electronically stored information.

If your company has no AI usage policy, no retention policy for AI interactions, and no litigation-hold procedures that cover AI tools, you have a gap. It is the same gap companies had with email in the early 2000s and with Slack and Teams messages in the 2010s. The technology moved faster than the governance, and courts are filling in the rules after the fact.

What Should Employers Actually Do?

Treat AI chat tools the way you treat email and messaging platforms for legal-hold purposes. Specifically:

• Inventory which AI tools employees are using, including personal accounts used for work tasks.
• Establish retention policies that align with your existing document-retention schedules.
• Include AI tools explicitly in litigation-hold procedures. If a hold triggers, someone needs to know which AI platforms to preserve.
• Decide, as policy, whether employees may use AI tools for privileged communications (legal research, strategy discussions) and document that decision. The Warner/Heppner split shows that the purpose and context of use matters to courts.
• Understand your vendor's retention defaults. If your AI provider retains logs for 30 days, or 90 days, or indefinitely, that fact shapes your discovery obligations.

Does Architecture Matter? The "Voluntary Submission" Problem

The court's finding that users "voluntarily submitted their communications" to a third party is the load-bearing beam of the January ruling. And it is, at bottom, a description of how the product was built.

When you type a prompt into a cloud-hosted AI tool, your input travels to the provider's servers, gets processed, and a response comes back. The provider has your input. That is the voluntary submission. It is architecturally inherent to any product that routes requests to a remote inference endpoint.

We build Selina, a privacy-focused AI assistant. We should be honest about what this means: any AI product that calls a frontier API sends a slice of each request to a provider at inference. Memory in Selina is encrypted at rest, but it is not end-to-end encrypted, because inference requires the provider to process the input. That is a real constraint. We can minimize what's sent, encrypt what's stored, and enforce a short retention window on non-content operational metadata. Files and transfers via SelinaSEND can be zero-knowledge encrypted. But the fundamental dynamic the court identified, that using a cloud AI tool means submitting your input to a third party, applies to every product in this category, including ours.

The architectural question is not whether you can eliminate this exposure entirely (you cannot, short of running models locally). It is whether you can minimize what's retained, give users and administrators real control over retention, and make deletion verifiable rather than aspirational. Those are design choices. They directly affect your legal exposure surface. A tool that retains logs indefinitely by default creates a different discovery profile than one that enforces short, bounded retention by policy and architecture.

What Can Individual Users Do Right Now?

Start with the assumption that anything you type into a cloud-hosted AI tool could surface in litigation. Not necessarily will. Could. That reframing changes behavior more effectively than any policy document.

Concrete steps:

• Do not type information into an AI chat that you would not put in a work email. Same standard. Same reason: both are potentially discoverable.
• Check your AI provider's retention policy. Know whether your conversations are stored for 30 days, 90 days, or until you delete them (and know that deletion may not survive a litigation hold).
• If you are using AI for anything related to legal strategy, privilege, or sensitive business decisions, talk to your legal team first. The Warner/Heppner split means the rules are jurisdiction-dependent and fact-specific.
• If you are a freelancer or sole practitioner, recognize that your AI chat history with a third-party provider has weaker legal protection than notes on your own hard drive. The third-party doctrine is old and well-established.

Where Is This Headed?

The trend line is clear, even if individual rulings are inconsistent. Courts are incorporating AI tools into existing discovery frameworks, not creating special carve-outs for them. The Fisher Phillips analysis identified AI-generated ESI as a "core discovery battlefield" in employment litigation. Enterprise AI platforms have started rolling out formal compliance and export infrastructure, including immutable, time-windowed log exports, a sign that log retention and producibility are becoming standard enterprise features rather than edge cases.

The January ruling did not create new law. It applied old law, the Federal Rules of Civil Procedure, the third-party doctrine, proportionality analysis, to new facts. That is precisely why it matters. There is no legislative fix coming that will retroactively make your chat logs un-discoverable. The rules that govern email discovery govern AI chat discovery. Courts have now said so explicitly.

The practical upshot: control what you can. Choose tools with deliberate retention policies. Establish governance before litigation forces you to. Understand that the privacy of your AI conversations is bounded by architecture and law, not by interface design or marketing promises. The delete button is a UI element. Whether it is also a legal fact depends on things outside your control.

If you want an AI assistant built with these constraints in mind, start a free 7-day trial of Selina, no card required.

Michael C.

Founder & Principal Engineer, Selina Labs

Michael builds Selina, a privacy-first AI that remembers you across conversations. He ships security-sensitive AI in production — real attacks, real fixes, measured in minutes and dollars — and writes about privacy, security, and LLMs from that seat. Top Rated Plus and expert-verified on Upwork.

• www.upwork.com
• x.com
• medium.com
• selinalabs.io