In January 2026, a federal judge forced OpenAI to hand over 20 million ChatGPT conversations. Nobody asked the users. Nobody told them. So can your ChatGPT conversations be used in court? Yes. They are treated as ordinary electronic records: discoverable in a lawsuit, subpoenable by investigators, and admissible as evidence, with no special privilege protecting them. A conversation with a lawyer, doctor, or therapist is shielded by legal privilege. A conversation with a chatbot is not.
A chat window is closer to an email than to a confession booth. Treat it that way and most of the legal risk takes care of itself.
This is the calm read between scary headlines and dense law-firm memos: what is actually discoverable, what deletion really does at the provider level, how privilege works and where it stops, and a short list of things you should not type into a general-purpose chatbot. The one thing to hold onto is that the input box is a record, not a private space.
Why AI chats are not legally privileged
AI chats are not privileged because privilege only attaches to specific protected relationships, and a chatbot is none of them. Attorney-client privilege, doctor-patient confidentiality, and the therapist privilege exist because the law has decided those relationships need protected candor. There is no equivalent doctrine for talking to an AI. When you type a question to ChatGPT, Claude, or Gemini, you are disclosing information to a company, and that company can be compelled to produce it.
OpenAI's own CEO has said this out loud. On Theo Von's podcast in July 2025, Sam Altman acknowledged that people use ChatGPT like a therapist or life coach, yet there is no legal confidentiality for those conversations the way there is with a real therapist or lawyer. He confirmed that if a lawsuit demanded those conversations, OpenAI would be legally required to produce them, and he argued that AI chats should carry the same kind of privacy protection people get with a therapist. That protection does not exist today.
Half the people seeking legal advice from ChatGPT had no idea it could be subpoenaed. A Kolmogorov Law survey of 1,000 U.S. adults who use AI chatbots found 56% have sought legal advice from a tool like ChatGPT, half did not know those conversations could be subpoenaed, and 67% believed their chats should be privileged like a talk with a real lawyer. The belief is widespread. The protection is not.
The courts are still sorting out the edges
One narrow nuance: privilege can sometimes survive if you use a chatbot as a tool inside an existing legal process. In Warner v. Gilbarco, a Michigan federal court treated a self-represented plaintiff's ChatGPT materials as protected work product, reasoning that generative AI programs are "tools, not persons," so feeding them notes did not waive protection. But the opposite result is just as real. In United States v. Heppner, a New York federal court found that a defendant's Claude messages were not privileged, because they were not communications with counsel and were not confidential under the provider's data policy.
The takeaway from those split rulings is not that chats are sometimes safe. It is that the only reliable protection comes from the underlying relationship, not the software. Routine AI use for business or personal questions is unlikely to be viewed as privileged by any court.
The CEO who built the case against himself
Here is the part most explainers skip. They tell you chats could be used against you someday. A Delaware court has already done it. On March 16, 2026, in Fortis Advisors, LLC v. Krafton, Inc., the Delaware Court of Chancery resolved a fight over a $250 million earnout tied to Krafton's acquisition of Unknown Worlds, the studio behind Subnautica. Vice Chancellor Lori Will found that Krafton CEO Changhan Kim, fearing he had signed a "pushover" contract, asked ChatGPT to help him contrive a corporate "takeover" strategy to avoid the payout, then followed most of the chatbot's recommendations over the following month.
A CEO asked ChatGPT how to dodge a $250 million payment. The chatbot answered. The judge read the whole exchange and quoted it in the opinion.
The opinion block-quoted Kim's actual prompts and the AI's strategy back to him, turning the chat log into what one analysis called a virtual witness to his intent. The court rejected Krafton's defenses, found the company breached the deal, reinstated an ousted Unknown Worlds executive, and extended the earnout window. The chats were not a side detail. They were the evidence of state of mind that decided the case. That is the difference between a hypothetical risk and a docketed one.
What is actually discoverable
Everything you send and receive is discoverable. Courts classify AI prompts and outputs as electronically stored information, the same bucket as emails, texts, and Slack messages. That classification matters because the bar for discovery is low. In civil litigation, a party can demand any nonprivileged material that is relevant to the dispute and reasonably calculated to lead to admissible evidence, subject to proportionality. The information need not be admissible itself, only likely to point toward something useful.
In a civil case the usual route is not a subpoena to OpenAI, which runs into Stored Communications Act limits, but a discovery demand for your own chat exports. Either way the records surface. Because chats are ESI, they can be used the way any other record is used: to impeach testimony, to show what you knew and when, or to demonstrate your state of mind, exactly as Kim's prompts did. A message describing what really happened the night of an accident is the kind of evidence opposing counsel looks for. Stored AI memory, the persistent profile some tools build from your past chats, sits in the same category and is just as reachable.
- Prompts you type, including the messy first draft you immediately rephrase.
- The model's responses, which can quote or paraphrase what you disclosed.
- Persistent memory and saved context that carries facts across sessions.
- Account metadata: timestamps, which device, how often, which workspace.
- Files and images you upload for the AI to analyze.
What deleting a chat actually does
Deleting a chat does not reliably erase it. Removing a conversation clears it from your view, but it does not guarantee the provider has erased it. Two different forces govern this: the company's retention schedule and any court preservation order layered on top. OpenAI's standard policy deletes data within about 30 days of a delete request, but a litigation hold can override that schedule entirely and freeze data in place, deleted or not.
On May 13, 2025, Magistrate Judge Ona T. Wang ordered OpenAI to "preserve and segregate all output log data that would otherwise be deleted," which swept in Free, Plus, Pro, Team, and standard API users, even chats deleted by the user or sent in temporary chat mode. ChatGPT Enterprise and Edu accounts, and API customers on zero-data-retention endpoints, were carved out. For everyone else, hitting delete did nothing to remove the data the court wanted.
That broad order did not last forever. On October 9, 2025, Judge Wang lifted the blanket preservation requirement, so OpenAI could resume normal deletion for most users. But a narrower obligation survived: logs tied to accounts the plaintiffs specifically flagged still had to be kept, and anything already preserved stayed accessible to them. The lesson is structural, not tied to one case. Deletion at the interface is a request, and a court can suspend it whenever your data becomes relevant to a dispute.
You would hear about a subpoena for your therapy notes. You will not for your chatbot logs. Here is how a chatbot conversation compares to a privileged one across five questions:
| Question | Talking to a lawyer or therapist | Talking to a general AI chatbot |
|---|---|---|
| Legal privilege | Protected by attorney-client or therapist privilege | No privilege; treated as a third-party disclosure |
| Subpoenable in a lawsuit | Generally shielded unless privilege is waived | Yes; classified as discoverable ESI |
| What 'delete' does | Records governed by professional retention rules | Hides from your view; provider may still hold it |
| Court can force production | Rarely, and only around privilege limits | Yes, as the 20M-log order showed |
| User notified before disclosure | Usually, through the privilege process | Often not; affected users got no notice |
The 20 million figure has a backstory worth knowing. News plaintiffs first asked for 120 million logs out of the tens of billions OpenAI had preserved. OpenAI offered 20 million de-identified logs, about 0.5% of its data, calling that more than enough, then tried to narrow it further to only conversations matching plaintiffs' keywords. The magistrate rejected that narrowing, and in January 2026 Judge Sidney Stein affirmed the full 20-million-log production. The users behind those conversations were not asked and did not get to object.
What not to type into a chatbot
The practical rule is simple. Do not put anything into a general-purpose chatbot that you would not want read aloud in a deposition. That does not mean stop using AI. It means treat the input box like a work email that could be forwarded, not a diary. The categories below are the ones lawyers flag most often as future evidence.
One line to remember: treat the input box like a work email that could be forwarded, not a private journal. If you would not want it read aloud in a deposition, do not type it.
- Admissions about an accident, injury, or incident you may be sued over.
- Hiring, firing, or HR decisions, especially anything that hints at a protected-class motive.
- Confidential client or employer information that triggers a third-party disclosure.
- Detailed legal strategy you would normally route through your attorney.
- Medical, financial, or relationship details you assume are private therapy-style talk.
- Anything involving a current or anticipated dispute, where a preservation duty may already apply.
If a matter is genuinely sensitive, route it through the privileged relationship first. Ask your lawyer, then use AI as a drafting tool under their direction. Enterprise tiers with zero data retention and a signed agreement reduce exposure, but they do not create privilege that the law does not recognize.
Privacy you control versus privacy you assume
Most of the risk above comes from one structural fact: your conversations live on someone else's servers, governed by their retention schedule and reachable by their legal obligations. The fix is being deliberate about where your personal context is stored and who can be compelled to produce it. Privilege is a legal status no consumer tool can grant you, but architecture is something you can choose.
This is the gap MemX is built to address. MemX (memx.app) is an external AI memory layer, a personal memory app that holds your context across the AI tools you use, designed to be private by architecture: per-user isolation, encryption at rest, and on-device options so your memory is not pooled into one provider's shared logs. That reduces how much sensitive context sits inside any single chatbot's retained data. To be clear, MemX does not claim end-to-end encryption or zero-knowledge, and no consumer tool can grant legal privilege. The point is control over where your context lives, not a courtroom shield.
Assumed privacy is the comfortable belief that hitting delete makes a chat disappear. Real privacy is knowing exactly which system holds your data, on what terms, and how isolated it is. The legal picture will keep shifting as courts and lawmakers debate whether AI conversations deserve a new privilege. Until that exists, the safe assumption is the one the 20-million-log order made unavoidable: a chatbot conversation is a record, and records can be produced.
01Can ChatGPT conversations be used in court as evidence?
Yes. Courts treat AI chats as electronically stored information, the same category as emails and texts. They are discoverable in lawsuits, can be subpoenaed, and are admissible as evidence. There is no chatbot privilege protecting them from production.
02Is ChatGPT confidential like talking to a therapist or lawyer?
No. Unlike talks with a lawyer, doctor, or therapist, AI conversations carry no privilege. OpenAI's CEO has confirmed there is no legal confidentiality for ChatGPT chats and that the company would produce them if subpoenaed.
03Can my ChatGPT chats be subpoenaed, and does deleting them help?
They can be subpoenaed. Deleting removes a chat from your view, but the provider may retain it on a roughly 30-day schedule, and a court preservation order can freeze it indefinitely, including chats you deleted or sent in temporary mode.
04Was OpenAI actually forced to hand over chat logs?
Yes. In January 2026, federal Judge Sidney Stein affirmed an order requiring OpenAI to produce 20 million de-identified ChatGPT logs in the copyright case brought by The New York Times and other publishers. The affected users were not notified.
05What should you not tell AI chatbots?
Avoid admissions about disputes, hiring or firing decisions, confidential client or employer data, detailed legal strategy, and sensitive medical or financial details. Treat the input box like a forwardable work email, not a private journal.
