AI Attorney Client Privilege After Heppner: Where the Split Stands This Fall

TL;DR:Courts are still split on whether AI chats are protected, but the split now has a shape. Since February, when a federal court in New York refused protection to one person's own AI-generated documents, courts in Michigan, Colorado, Texas, and New York have protected AI-assisted litigation prep as work product. Protection is turning on who directed the AI, which tool held the material, and what the protective order says.

The AI attorney client privilege question is not settled, but it is no longer a coin flip. Courts are protecting AI-assisted litigation prep that looks like ordinary work product, and they are policing what goes into consumer tools. The February ruling in United States v. Heppner is still the warning for anyone who uses a consumer tool on their own. I run a multi-agent AI team applied to legal work, business projects, and teaching others how to use AI. Here is how I read where the cases stand.

What did Heppner decide, and why does it still matter?

In February 2026, Judge Jed Rakoff of the Southern District of New York held that documents an individual created on his own with a consumer AI tool were neither privileged nor protected work product. The case is a federal fraud case, United States v. Heppner. The court ruled from the bench on February 10 and issued a written opinion on February 17. The core of the reasoning was simple: an AI tool is not a lawyer, so a conversation with it is not a privileged communication. Heppner matters because it is the cleanest statement of the risk. Material created on a consumer tool, outside counsel's direction, may be discoverable.

The same day Rakoff ruled from the bench, a federal court in Michigan went the other way. In Warner v. Gilbarco, Magistrate Judge Patti treated a self-represented litigant's ChatGPT queries and the responses as work product, and treated the AI as a tool rather than a third party, so using it did not waive protection. I covered both rulings when they came out in my earlier post on Heppner and AI privilege. What has happened since is more interesting than the original split.

Which courts have protected AI-assisted litigation prep since then?

Three courts, each adding a piece. In Morgan v. V2X (D. Colo., March 30, 2026), the court protected the AI-assisted analysis but required the party to disclose which AI platform received confidential information. It also amended the protective order to keep confidential material out of consumer AI tools unless the provider is contractually barred from training on it or disclosing it. In Tate Group Automotive v. Legacy Automotive Capital (Texas Business Court, June 3, 2026), a represented company's principal, who is not a lawyer, had ChatGPT conversations the court treated as work product. Using ChatGPT did not waive the protection. In Assini v. Hayward (N.Y. Sup. Ct., Nassau County, June 4, 2026), the court quashed a subpoena to OpenAI that sought a self-represented party's prompts, inputs, and outputs.

None of these courts handed out a blank check. Tate still ordered a few pages that were not work product produced, and it made the party identify the discovery documents it had shared with ChatGPT. Assini relied on Morgan and Warner, treated Heppner as not binding, and then pointed the self-represented party to New York's new AI rule, 22 NYCRR Part 161, with a warning that failing to comply could bring sanctions. The protection came with conditions attached.

What is the split really turning on?

Posture and tool choice. Posture means who directed the work and why. A self-represented party preparing their own case, or a represented company preparing for litigation, has the strongest claim. A represented party who works on a consumer tool on their own, without counsel directing it, has the weakest. That was Heppner. Tool choice means where the material went and what the provider is allowed to do with it. Morgan made that explicit by writing it into a protective order. The law on legal artificial intelligence tools is being built one trial court at a time, and those two questions keep coming up.

Nothing here turns on the brand. The reported cases mostly involve ChatGPT, but the questions are the same for Claude for lawyers or any other model: who directed it, what went into it, and what the provider's terms say about training and disclosure.

Can opposing counsel subpoena ChatGPT?

They can try, and in at least one case it failed. In Assini, the other side subpoenaed OpenAI directly for a self-represented party's prompts, inputs, and outputs. The court quashed the subpoena and the amended subpoena, treating the AI-assisted preparation as protected litigation material. That is one New York trial court, and it does not bind any other court. But it answers a question many people assumed would go the other way. Going around the party to the AI provider did not get around the protection. The more common fight, as Tate and Morgan show, is over what the party itself must disclose about its AI use.

Most of the "ChatGPT lawyer" headlines this year were about fake citations. Privilege is the quieter issue, and it may matter more over time, because it reaches every case where AI touched the prep work.

What should anyone using AI on a case set up now?

The cases point to four things worth settling before AI touches case material. This is general information about where courts are heading, not advice about any particular case.

One. Decide who directs the AI work, and write it down. The protection is strongest where the work is clearly litigation prep done by the party or at counsel's direction. Heppner shows what happens when nobody can say that.

Two. Pick tools by their terms, not their features. Morgan drew the line at providers that are contractually barred from training on or disclosing the material. Consumer accounts often do not offer that.

Three. Read the protective order for AI language, and raise it early if there is none. After Morgan, it is reasonable to expect the other side, or the court, to ask.

Four. Keep a record of what went into which tool. Tate made a party identify the discovery documents it had shared with ChatGPT. That is easy to answer with a log and hard to answer without one.

I went deeper on the tool side of this in a piece on confidentiality across a small office's AI stack. The short version holds here too: the setup decisions you make before the first prompt are the ones that matter later.

Where does this leave the AI attorney client privilege question?

Unsettled, with a direction. As of mid-October 2026, I could not find an appellate ruling in Heppner or Assini, and the trial courts that have ruled do not bind each other. The pattern across them is consistent enough to plan around. Courts treat AI as a tool, protect work that looks like litigation prep, and police where confidential material goes. Heppner marks the edge: a represented party's own work on a consumer tool, outside counsel's direction, gets the least protection. No court has said an artificial intelligence lawyer exists. Every one of them has asked who was actually directing the work.

Until next week, the team keeps running.