Your Conversation with Claude is now Exhibit B

It’s 11:47 p.m. on a Tuesday. The kids are asleep. The email from your ex came in at dinner — something about pickup, something about the school. You open Claude. You paste the email. You ask what to write back.

You close the laptop. You go to bed.

That conversation is now a document on someone else’s server, governed by terms of service that explicitly contemplate disclosure. It is not a memory. It is not a private thought. In the wrong custody case, in the wrong jurisdiction, on the wrong day, it is Exhibit B.

What two federal courts decided in eight weeks

In the first eight weeks of 2026, federal judges in two different districts gave the country its first real answers on whether AI chat logs are protected from discovery. Neither answer is reassuring.

United States v. Heppner. On February 10, Judge Jed Rakoff of the Southern District of New York issued an oral ruling — followed by a written opinion on February 17 — in the criminal case of Bradley Heppner, a Dallas financial executive charged with securities fraud. Heppner had used Anthropic’s Claude to generate roughly 31 documents on his defense strategy: anticipated charges, possible arguments, what the government might try. The FBI seized the documents during a search of his home. His defense team asserted attorney-client privilege and work-product protection. Rakoff rejected both, on three independent grounds:

Claude is not a lawyer. Privilege requires a confidential communication with a licensed attorney for the purpose of obtaining legal advice. The first element fails on its face.
The communications were not confidential. Anthropic’s terms contemplate disclosure to governmental and regulatory authorities and use of inputs for model training. The court treated those terms as dispositive of any reasonable expectation of confidentiality.
Heppner used Claude on his own initiative. His own counsel admitted at the hearing that they had not directed him to use the tool. Work-product protection requires preparation in anticipation of litigation by or at the direction of counsel.

This is the first ruling of its kind anywhere in the country. It will be cited in every adjacent case for years.

NYT v. OpenAI. A few weeks earlier, in January 2026, the same court — different judge — handed down something arguably more consequential. Magistrate Judge Ona Wang and District Judge Sidney Stein ordered OpenAI to produce 20 million ChatGPT conversation logs to The New York Times and other publishers in their copyright suit. OpenAI offered a curated subset. The court refused and required the full sample. The users whose conversations sit in that production received no notification and had no opportunity to object. They are not parties. Their chats are now in the discovery record.

A May 2025 preservation order in the same case forced OpenAI to retain conversations users had already deleted. Temporary chats too.

Warner v. Gilbarco. On the same day Heppner was decided, a magistrate judge in the Eastern District of Michigan went the other way. Judge Anthony Patti held that a pro se employment-discrimination plaintiff’s exchanges with ChatGPT were protected by the work-product doctrine, on the theory that the AI was a tool — not a third person — and the chats represented a litigant’s internal mental impressions reformatted through software.

The contrast between Heppner and Warner is the doctrinal fault line. Two judges, same day, opposite conclusions, on facts that look superficially similar at the keyboard. Parents do not get to pick which framework a future judge applies.

The free-versus-paid myth

The single most common misconception about all of this is that paying twenty dollars a month for ChatGPT Plus or Claude Pro creates a private space. It does not.

Heppner specifically involved the free tier of Claude. But the legal analysis from the major firms reading the ruling has been consistent: the distinction matters less than most consumers assume. Anthropic and OpenAI both use conversations from free and individual paid plans — Claude Free, Pro, and Max; ChatGPT Free, Plus, and Pro — for model training by default. Users can opt out of training. Opting out does not remove the platform’s contractual right to disclose data in response to subpoenas, court orders, or regulatory inquiries.

Only enterprise-tier agreements — ChatGPT Enterprise and Business; Claude’s commercial and government plans — exclude user data from training by default and provide contractual confidentiality protections. A twenty-dollar subscription does not buy privilege.

The NYT preservation order is the cleanest illustration. It applies to ChatGPT Free, Plus, Pro, and Business, including chats users believed were deleted. Only Enterprise, Edu, and API customers with zero-data-retention agreements were exempt. These are not products consumers buy. They are products buying departments at law firms and Fortune 500 companies negotiate.

Why this hits family law harder than any other domain

Custody and divorce cases run on three things courts read carefully: communications, finances, and credibility. AI chat logs now sit in that pile.

What gets typed into a chat window during a custody dispute is rarely benign:

“Can he take her out of state without telling me?”
“What’s the wording for an emergency motion if she missed pickup again?”
“How do I write a response that doesn’t make me look unstable?”
“What happens if I record him without telling him?”
“How do I make sure she pays the medical bills the parenting plan says she owes?”

Each of these, in opposing counsel’s hands, supplies a portrait of strategy, intent, or worry the parent never meant to publish. Discovery doesn’t require the parent’s permission. No privilege protects the conversations because the chatbot is not a licensed attorney. Deleting conversations from visible history does not remove them from the platform’s servers, where they may be retained — and, under a preservation order, must be retained — far longer than the consumer-facing UI suggests.

This is the moment to acknowledge what’s hard about it. Parents in high-conflict custody cases are exhausted. They are looking for a way to think clearly under conditions that make clear thinking nearly impossible. The chat window feels like it. It answers immediately. It does not charge by the hour. It does not require an appointment. The trade is invisible until it isn’t.

The category error

Parents are using consumer LLMs for two distinct things at once, and the difference is doing real legal work in the background.

General education. “What does temporary custody mean in Illinois?” “What’s the difference between legal custody and physical custody?” “What is a guardian ad litem?” Low-risk. Discoverable, but not strategically damaging. The information is the same information available on any state bar website.

Case-specific strategy. Pasting in actual emails. Naming actual people. Working through the actual response, with the actual context, in the actual matter. High-risk. The exact material a competent opposing counsel will subpoena.

The chat window doesn’t distinguish between the two. The parent doesn’t, usually. The court will. The Heppner court drew exactly this line: a parent who asks Claude or ChatGPT to help draft a substantive response in their case is not researching the law. They are doing case work, in a venue that gives them no protection for it.

What documentation should look like

The point of this piece is not that parents should stop documenting their custody cases. The opposite. Parents in contested custody cases should document everything. The question is where, and into whose terms of service.

Truly Noted is built for this. It is a documentation tool and case-prep layer for parents in active custody and divorce matters. It is not a substitute for an attorney, and it does not pretend to be. Legal interpretation belongs to the attorney. The product organizes what the parent already has — text messages, emails, OurFamilyWizard exports, court filings, therapy records — into a timeline and a brief the attorney validates, item by item.

A short list of architecture decisions that distinguish Truly Noted from a chat window:

Names and identifying details are anonymized client-side before any data touches an LLM. The model never sees the parent’s name, the child’s name, or the other parent’s name. This is the load-bearing privacy decision. Everything else follows from it.
Paid LLM contracts with training-data carve-outs. Free-tier LLM use is excluded from the architecture entirely. The contractual posture parents cannot get on their own, Truly Noted gets on their behalf.
Encryption in transit and at rest, with cheap, durable encryption suitable for indefinite retention with integrity preserved.
Two-factor authentication for parent accounts. Tokenized email links for attorney access — twenty-four-hour expiration, recipient confirmation, a “sent in error” path that kills the token and notifies the parent.
Discrete, logged consent events for terms, not-legal-advice acknowledgment, data processing, and attorney designation. Each event is its own record.
The validated brief is non-shareable from the parent’s view. The attorney’s marked-up version cannot be downloaded or forwarded by the parent.
The work product flows into the attorney relationship, where the Heppner test starts to bend back toward privilege: counsel-directed use, on a confidentiality-protected platform, for the purpose of obtaining legal advice.

This last point matters more than any other. The Heppner court did not categorically exclude AI from privilege. It excluded a particular pattern of use: a represented client, on a free public chatbot, working alone, without his attorney’s knowledge or direction. The court left explicit room for a different outcome where AI is used at counsel’s direction on a platform with proper confidentiality protections — what legal commentators have begun calling the third-party consultant framework.

Truly Noted is built to be that kind of system. Not an open chat window where parents talk to a model. A documentation layer where parents organize records they already have, anonymization happens before any LLM sees a single name, and the work product is delivered into a workflow their attorney controls.

We will not promise that any output of any tool is categorically privileged. That is a question for the courts, the attorney, and the facts of the case. The honest claim is structural: Truly Noted is designed to keep parents out of the trap Heppner warns about. The chat window puts the parent in the worst possible factual posture for any future privilege analysis. Truly Noted is built to put them in the best one available.

What parents should do this week

Stop pasting case-specific material into ChatGPT, Claude, Gemini, or any consumer AI. This includes the email you’re about to respond to. This includes screenshots. This includes anything that names a person involved in your case.
Tell your attorney what you’ve already typed in. Volunteering it now lets counsel assess exposure before opposing counsel finds it. Spoliation — destroying records once litigation is reasonably anticipated — carries its own sanctions. Do not delete first and ask later.
Use AI only for general legal education. Definitions, statute names, vocabulary. Stop the moment a fact about your case appears in the prompt.
Document the actual record in a tool designed for it. Note it. Don’t chat about it.

Close

The chat window felt private because the rest of the internet trained us to think of it that way. The court file is going to disagree.

Sources

United States v. Heppner, S.D.N.Y. (Feb. 17, 2026) (Rakoff, J.)
In re OpenAI, Inc. Copyright Litig., S.D.N.Y. (January 2026) (Stein, J.; Wang, M.J.)
Warner v. Gilbarco, Inc., E.D. Mich. (Feb. 10, 2026) (Patti, M.J.)
Jones Walker LLP, “Your AI Conversations Are Not Privileged: What a New SDNY Ruling Means for Every Lawyer and Client” (February 2026)
New York State Bar Association, “Loose AI Prompts Sink Ships: How Heppner Shook the Legal Community” (March 2026)
Harvard Law Review Blog, “United States v. Heppner” (March 2026)
Lawfare, “AI and Privilege After United States v. Heppner” (March 2026)