What ChatGPT knows about your divorce. And who can subpoena it.

Why this is now a routine discovery question

The Heppner ruling is the watershed for criminal matters, the same logic we covered in your ChatGPT and Claude conversations are not protected. Family courts have moved in parallel. By mid-2026, the standard discovery questionnaire in family law cases in roughly half of US states includes an explicit question about AI tool use. Where it is not yet on the questionnaire, sophisticated counsel raise it during the first deposition.

The reason it is asked is grounded in observed practice, not theory. People in the middle of a divorce talk to a chatbot about things they would never write in an email: their honest assessment of the marriage, what they think their spouse will demand, the financial scenarios they are running, what they would settle for, what they intend to hide, what they fear the other side will discover. The chatbot is a private journal that happens to be on someone else’s server. Once the question is asked in deposition, the conversations are fair game.

What the chatbot actually has on you

The export of a few months of ChatGPT or Claude history, in a divorce context, typically reads as follows.

Direct factual admissions: how much money you actually have, what assets you have not yet disclosed, the income from the side business you forgot to mention. Strategic admissions: which custody arrangement you would accept rather than fight, what you think the other parent’s weaknesses are. Emotional admissions: your honest feelings about the children, your spouse, the new partner, the in-laws. Each of these has been produced in actual cases through 2025 and 2026. None of them help you when read in court.

The exports also contain context that sophisticated opposing counsel finds useful: the timestamps of when you searched for “how to hide money from spouse” or “is my prenup enforceable in California,” which establish patterns of intent independent of any specific message you sent. The metadata is sometimes more damaging than the content, which is the same pattern we covered in metadata vs content in court.

The protocol if you have already used a chatbot

Three steps, in order. The order matters.

1. Stop, and tell your lawyer before you delete

Stop using the chatbot for anything related to the divorce. Address the conversation with counsel before deleting anything. The line between routine data hygiene and evidence spoliation is drawn at the point where legal proceedings become reasonably foreseeable. By the time someone is reading an article like this one, that line is almost certainly already crossed. Deletion at this point can constitute spoliation, which is treated worse than the underlying conversations would have been. The framework is the same one in deleting evidence vs protecting yourself.

The lawyer’s job at this point is to evaluate what is in the conversations, decide whether to disclose voluntarily, and structure the production if necessary. Voluntary early disclosure to opposing counsel often produces a better outcome than waiting for a discovery request to surface the same conversations later.

2. Export everything before anything else

Export the full conversation history from each platform you have used. ChatGPT: Settings, Data Controls, Export Data. Claude: Settings, Privacy, Export. Gemini: takeout.google.com. Copilot: through your Microsoft account settings. The exports arrive within twenty-four to seventy-two hours. They are JSON archives. Save them in a location your lawyer can access; do not save them in a shared cloud account that your spouse might have access to.

Read what is in them. The export is the same one opposing counsel can request through subpoena. Knowing what is there before they do is a meaningful tactical advantage.

3. Move sensitive thinking offline

For the rest of the case, the rule is simple. Anything that touches the divorce, the children, the finances, the strategy: not in a chatbot. Not in a cloud note. Not in a shared device. The thinking has to happen on paper that you destroy, in conversations with your lawyer that are privileged, or in your own head. The principle is the compartmentation we cover in operational identity separation between the chatbot account and the divorce file: data that exists outside your control is data that can be subpoenaed.

If your spouse used ChatGPT

The mirror question, less often considered. If your spouse has been using a chatbot to think through the divorce, those conversations are also producible. Your lawyer can ask for them through standard discovery. The same question that targets you targets them.

The signals to look for: shared devices that show ChatGPT or Claude in browser history, shared accounts that show app installations on phones, references in conversation that suggest the spouse has been “researching” the divorce online. None of these confirm AI tool use, but they are reasons to ask the discovery question. If the answer is yes, the conversations are produced. If the answer is no and the conversations later surface, the false statement is its own problem for the spouse.

The Enterprise tier question

If you used your employer’s ChatGPT Enterprise account or Microsoft Copilot for Business to think through your divorce, the analysis is more complicated and worse. The conversations are now also producible from your employer, who may produce them in response to a subpoena without notifying you. They may also surface in employer monitoring of personal use of company tools, which can become a separate workplace issue. The right move is to keep work-account AI tools out of personal matters entirely. If that line has already been crossed, the next step is the same as for consumer tools: tell counsel before taking any action on the conversations.

Frequently asked questions

If I use a paid plan with no-training, is the conversation private?

No. The training opt-out is about whether your data is used to improve the model. It does not affect retention for safety, compliance, or legal-hold purposes. ChatGPT Plus retains your conversations under the same default thirty-day window plus any active legal holds. Claude Pro and Gemini Advanced are similar. The paid tier reduces some flows. It does not produce privilege.

The framework is the one in your law firm uses ChatGPT: the contractual posture matters for some categories of risk, but no consumer or even prosumer tier creates the kind of confidentiality that an attorney-client relationship creates.

Will my spouse’s lawyer actually subpoena ChatGPT?

Increasingly, yes. The cost has dropped: most law firms now have template subpoena language for major AI providers, and the providers have streamlined response processes. The technical question (whether the data exists and can be produced) is settled. The legal question (whether a court will compel production over privilege objections) was settled by Heppner for criminal matters and is being settled in family courts case by case. Most family court judges, asked to compel production, have ordered it.

What if I used the chatbot before the relationship deteriorated?

Older conversations have less obvious relevance, but the discovery request will typically cover the entire history. Conversations from before the marriage problems began may be requested anyway, and the producing party has to make a relevance argument to limit the scope. The most relevant question is what is in the conversations, not when they happened. Read the export. If everything in it predates anything that could be relevant, your lawyer has a stronger argument for limiting the production.

What about disappearing-message apps for the same purpose?

Signal with disappearing messages enabled, between you and your lawyer, is generally protected because the lawyer-client privilege covers it independently of what the app does. Signal between you and a friend you are processing the divorce with is not the same: that conversation is not privileged, and the friend is a fact witness whose phone is also subject to subpoena. The disappearing-message feature reduces but does not eliminate exposure. Talk to a privileged person about privileged things; do not assume an app feature creates the legal protection.

---

There’s no perfect setup. Anyone selling you perfect is selling fear. The goal is simple: make yourself a harder target than the person next to you.