Is ChatGPT confidential? Millions of people now confide in AI chatbots about their health, their marriages, their finances and sometimes the very conduct a prosecutor is investigating. In February 2026 a federal court in New York gave the first clear answer to what happens when those conversations become relevant to a case: they are not privileged, and the other side can read them. This analysis explains that ruling, the courts that have since taken a different line, what the providers' own terms say, and how Turkish law answers the same questions under the Attorneys' Act, the criminal and civil procedure codes and the KVKK.

Is ChatGPT confidential? The short answer

Not in the legal sense that matters most. A chatbot conversation is private in the everyday sense that other users cannot see it. It is not confidential in the way the law protects a conversation with your lawyer, doctor or therapist. The provider stores it, its staff may review it in defined circumstances, its terms allow disclosure when the law requires, and courts can order it preserved or produced.

OpenAI's chief executive said as much in July 2025. On a podcast, Sam Altman noted that people use ChatGPT as a therapist or life coach, and that while a conversation with a therapist, lawyer or doctor attracts legal privilege, nothing equivalent yet exists for ChatGPT, as TechCrunch reported. The warning has since become case law.

What did the court decide in United States v. Heppner?

United States v. Heppner, No. 25 Cr. 503 (JSR), is a securities and wire fraud prosecution in the Southern District of New York. Agents searching the defendant's home seized about thirty-one documents recording his exchanges with Claude, the consumer AI assistant operated by Anthropic. According to his lawyers, he wrote them in 2025, after receiving a grand jury subpoena and knowing he was the target, and on his own initiative used the tool to prepare reports outlining his defence strategy. He claimed attorney-client privilege and work-product protection.

Judge Jed S. Rakoff rejected both claims from the bench on 10 February 2026 and gave his reasons in a written memorandum filed on 17 February 2026, describing the question as apparently one of first impression nationwide. The reasoning follows the three elements of the privilege.

  • No lawyer. Claude is not an attorney. Recognised privileges, the court said, rest on a trusting human relationship with a licensed professional who owes fiduciary duties and can be disciplined; no such relationship exists with an AI platform.
  • No confidentiality. Anthropic's privacy policy said that inputs and outputs are collected, may be used for training and may be disclosed to third parties, including governmental regulatory authorities. On those terms there was no reasonable expectation of confidentiality. The court also cited a January 2026 ruling in the OpenAI copyright litigation that users have no substantial privacy interest in chats voluntarily given to a platform that retains them in the normal course of business.
  • Not for legal advice. The defendant acted of his own volition, not at counsel's direction, and the tool disclaims giving legal advice. The court left one door ajar: had counsel directed the use, the tool might arguably have functioned like a lawyer's agent, as accountants and other experts working for counsel can.

Two further points matter most in practice. Sending the chats to his lawyers afterwards changed nothing: non-privileged communications are not, in the court's words, "alchemically changed into privileged ones" by being shared with counsel. And in a footnote the court added that even if information the defendant had learned from his lawyers and typed into the tool had been privileged, he waived the privilege by sharing it with Claude and Anthropic, just as if he had shared it with any other third party.

The work-product claim failed in parallel: the doctrine shields material prepared by or at the behest of counsel, and these documents were the defendant's alone.

Not every court agrees: the work-product split

On the same day, a magistrate judge in Michigan reached a different result on different facts. In Warner v. Gilbarco, Inc., an employment case brought by a self-represented plaintiff, the defendants sought everything about her use of AI tools in the lawsuit. Magistrate Judge Anthony P. Patti refused in an order of 10 February 2026. The motion came after the close of discovery, but the order also gave substantive reasons: the material was not discoverable, it was in any event work product, and using ChatGPT did not waive that protection, because waiver requires disclosure to an adversary or in a way likely to reach one. ChatGPT and similar programs, the order says, are tools, not persons.

On 30 March 2026, in an employment case against V2X, Inc., Magistrate Judge Maritza Dominguez Braswell in Colorado distinguished Heppner as a criminal case in which the defendant acted apart from his lawyer, whereas a self-represented civil litigant is both party and advocate. Routing information through a third-party system, she reasoned, does not forfeit all privacy, and chatbots designed to invite candour arguably strengthen the case for privacy.

The same order shows the other side. The plaintiff had to name any AI tool used with the defendant's confidential documents, and the protective order now bars putting confidential material into an AI platform unless the provider is contractually prohibited from training on inputs and from disclosing them to third parties, and allows deletion. The judge accepted that this rules out most mainstream low-cost tools for now.

The split is narrower than it looks. None of these courts held a chatbot conversation protected by attorney-client privilege; the disagreement concerns work product, which in US civil procedure can cover a party's own preparation. Protection depends on who created the material, for whom and on what terms, and the default answer is no.

Can your ChatGPT conversations be used in court?

Yes, and the largest example involves no user's own case. In the news publishers' copyright litigation against OpenAI, consolidated in New York as In re OpenAI, Inc. Copyright Infringement Litigation, the court in May 2025 ordered OpenAI to preserve output log data that would otherwise be deleted. According to OpenAI's own account, the order reached ChatGPT Free, Plus, Pro and Team users and API customers without a zero-data-retention agreement, including deleted chats, but not Enterprise or Edu. OpenAI says the obligation ended on 26 September 2025, that it again deletes deleted chats and Temporary Chats within 30 days, and that it still holds a limited set of April to September 2025 data the plaintiffs continue to seek.

The fight then moved to production. In late 2025 Magistrate Judge Ona T. Wang ordered OpenAI to produce 20 million de-identified ChatGPT logs, and on 5 January 2026 District Judge Sidney H. Stein upheld the order, according to Bloomberg Law, finding privacy adequately weighed and noting that users had voluntarily submitted their conversations. This is the January 2026 order that the Heppner memorandum cites.

The lesson: deletion is a company policy that a court can suspend, and users who are not parties to a lawsuit can still have their conversations swept into it.

What the providers' own terms say

Courts read the terms closely, so users should too. OpenAI's privacy policy, updated on 30 July 2026, names OpenAI OpCo, LLC in San Francisco as controller for users outside the European Economic Area and Switzerland, which includes users in Türkiye. Deleted conversations and Temporary Chats are removed within 30 days unless they must be kept longer for safety or legal reasons, such as a lawful subpoena. Personal data may be shared with government authorities where a legal obligation requires it or OpenAI believes in good faith that this is necessary to comply with one, and on further grounds such as preventing fraud or protecting the safety of users or the public.

There is also a safety channel: in an August 2025 post, OpenAI said conversations suggesting plans to harm others go to human reviewers, and an imminent threat of serious physical harm to others may be referred to law enforcement.

Business products differ: OpenAI says it does not train on business data by default, and Enterprise and Edu administrators control retention. The employer, not the employee, decides how long conversations are kept.

Does Turkish law protect your ChatGPT conversations?

Turkish law has no doctrine of "attorney-client privilege" in the Anglo-American sense. It reaches a similar result through rules that all attach to the lawyer.

  • Article 36 of the Attorneys' Act (Law No. 1136) forbids lawyers to disclose what is entrusted to them or learned through their work; they may testify about it only with the client's consent and may refuse even then without liability.
  • Article 46(1)(a) of the Code of Criminal Procedure (CMK) lets lawyers refuse to testify about what they learn in that capacity, and Article 249 of the Code of Civil Procedure (HMK) does the same for legally protected secrets, preserving the Attorneys' Act.
  • Article 154(1) CMK provides that a suspect's or accused's correspondence with defence counsel may not be inspected.

Each protection presupposes a lawyer. A chatbot conversation is neither a communication with counsel nor information entrusted to one, so a user's own AI chat sits outside these rules, like a diary. The Heppner court tested the facts against the elements of a privilege; a Turkish court would reach the same result more directly, because there is no general privilege to test, only professional duties and procedural shields built around the lawyer.

Two qualifications. Article 38 of the Constitution protects against compelled self-incrimination, which limits what an accused can be forced to hand over but not what a lawful search can find. And Turkish law does not expressly address notes a client prepares alone to brief counsel; an AI chat history would be a weak candidate for protection, because it has already been shared with a third party.

Criminal investigations: law offices, devices and the Constitutional Court

The strongest Turkish protection concerns the lawyer's office. Under Article 130 CMK, a law office may be searched only by court order, only for the matter specified, under a prosecutor's supervision and with the bar president or a representative present. If the lawyer or the bar's representative objects that a seized item belongs to the lawyer-client relationship, it is sealed and a judge decides within twenty-four hours; if the objection is upheld, the item is returned at once and the seizure records are destroyed.

That covers material in the lawyer's hands. A chatbot history on the client's phone, laptop or cloud account falls under Article 134 CMK: where there is strong suspicion based on concrete evidence and no other way to obtain the evidence, a judge, or in urgent cases a prosecutor subject to judicial approval within twenty-four hours, may order a search of the suspect's computers and records and the copying of data. A seized phone or laptop is often still logged in to its accounts, so an open chatbot app can be part of what a lawful search reaches.

Article 134 is itself in transition. In a decision of 12 February 2026 (E. 2023/128, K. 2026/36), published in the Official Gazette on 25 May 2026, the Constitutional Court annulled the article. By a majority, it struck down the core rules allowing a judge to order the search and copying of a suspect's computers and the seizure of devices whose encryption cannot be broken, finding them contrary to Articles 13 and 20 of the Constitution: the law says nothing about how data obtained under them are to be kept or deleted once the judgment becomes final, for how long and by what procedure, what rights the person concerned has, or which authority decides. It then annulled the rest of the article, which could no longer be applied, under Article 43(4) of Law No. 6216. The annulment takes effect on 25 February 2027; until then the article applies, which gives the legislature until that date to enact a replacement.

Chats held by a US provider are a separate route, which in practice runs through international judicial assistance.

Civil disputes: can a Turkish court order you to produce your AI chats?

Turkish civil procedure has nothing like American discovery. It does have targeted production duties. Article 199 HMK treats electronic data as documents. Under Article 219, parties must produce documents in their hands on which either side relies, electronic ones as printouts and, on request, in examinable electronic form. Under Article 220, where a requested document is necessary to prove the alleged fact and its existence is established, the court sets a final deadline; if the party fails without acceptable excuse, the court may accept the other side's account of its content. Article 221 extends production to third parties, who may refuse on the grounds available to witnesses.

A Turkish court is therefore unlikely to entertain a request for "all your conversations with ChatGPT". A specific, identified conversation that is necessary to prove a fact is different: its production can be ordered, and refusal can cost the case. Article 189(2) HMK bars unlawfully obtained evidence, so a chat taken by breaking into someone's account cannot be used; one obtained lawfully, for example from a company-administered workspace in line with data protection rules, can.

Data protection: KVKK, transfers abroad and the professional rules

For individuals chatting about their own lives, the KVKK has limited bite: Article 28(1)(a) excludes processing by a natural person in activities relating solely to themselves or their household, provided the data are not given to third parties. The position changes when the user is a business or a professional and the chat contains other people's data.

Under Article 9 of the KVKK, as amended in 2024, personal data may go abroad on an adequacy decision, on appropriate safeguards such as standard contracts notified to the Authority within five business days, or, only incidentally, on narrow exceptional grounds. The Authority's guide for lawyers, published on 22 September 2026, says that uploading client files to tools such as ChatGPT, Claude or Gemini may be a transfer abroad under Article 9, and recommends masking, minimisation and avoiding special-category data. We covered it in our report on KVKK's warning that pasting a case file into a chatbot may be a transfer abroad, and the transfer routes are set out in our guide to cross-border transfers under the KVKK. The Authority's June 2025 information note on chatbots had already warned of users over-sharing through lack of awareness.

The bar says the same in professional terms. The Union of Turkish Bar Associations' advisory guide on AI use by lawyers, discussed in our analysis of the AI hallucination cases, states that a lawyer may use AI but may never delegate professional secrecy to it, and that unanonymised client information, special-category data, case-file contents, evidence, trade secrets and negotiation strategy should not go into public tools or tools with insufficient contractual safeguards. Behind both guides stand the Penal Code offences of unlawfully disclosing personal data (Article 136) and disclosing client or commercial secrets learned through one's profession (Article 239).

The Heppner footnote adds the client side, which neither guide addresses: a client who pastes a lawyer's advice into a consumer chatbot may, at least before a US court, waive the protection it had.

Korea's new attorney-client privilege and why it does not reach chatbots

South Korea has just legislated a move from a duty-based model like Türkiye's, built on the confidentiality duty in Article 26 of the Attorney-at-Law Act and Articles 112 and 149 of the Criminal Procedure Act, to an express privilege. On 29 January 2026 the National Assembly passed an amendment, announced by the Ministry of Justice, inserting Article 26-2. Promulgated on 19 February 2026 as Law No. 21357, it enters into force one year after promulgation and also covers earlier communications.

It lets a lawyer and a client, or prospective client, decline to disclose confidential communications made to give or receive legal assistance, and documents and materials, including electronic ones, prepared for litigation, investigation or inquiry in a matter the lawyer has accepted. Exceptions include client consent, serious public interest such as the lawyer's involvement in a crime, the lawyer's own dispute with the client, and other statutes.

The first limb, like Heppner, protects communications between lawyer and client, which a chat with an AI service is not. The second protects documents prepared for the case without, on its face, requiring counsel's direction; whether it reaches a client's AI-assisted preparation is untested. Korea's regulator has already treated chatbot prompts as personal data sent abroad, as in its corrective recommendation to DeepSeek.

Practical guidance for individuals, companies and lawyers

Individuals

  • Write to a chatbot as if the conversation could one day be read in a courtroom.
  • Do not use a consumer chatbot to rehearse your account of a dispute, plan a defence or summarise your lawyer's advice. Under Heppner, the first is unprotected and the last can waive protection.
  • Temporary Chat, deletion and turning off training reduce exposure but do not eliminate it: deletion takes up to 30 days, and a legal hold can suspend it.
  • Take particular care with health, family and other people's data. Using a chatbot as a confidant is lawful, but not confidential.

Companies

  • Set a written rule on which tools may be used, for what and with which data; our AI use policy guide sets out the elements.
  • Prefer business tiers whose contracts exclude training, limit retention and allow deletion.
  • Treat prompts and outputs as business records that may need preserving when litigation, especially in the US, is foreseeable.
  • Monitoring employees' AI use is itself regulated; the limits KVKK set for workplace e-mail, discussed in our analysis of its e-mail monitoring principles, are the closest guide.
  • Where personal data go to a foreign provider, put the Article 9 transfer basis in place before use.

Lawyers

  • Keep client files out of consumer tools and use closed systems with contractual safeguards for sensitive work, as the TBB and KVKK guides expect.
  • Tell clients early that their own chatbot conversations about the matter are unprotected and that pasting advice into a chatbot can put it at risk. California has put some of these duties into statute, as noted in our report on SB 574.
  • Verify every output before it reaches a court; Mata v. Avianca remains the standard cautionary tale.

Three things are worth watching: whether a US appellate court resolves the work-product split; how the Turkish legislature replaces Article 134 CMK before 25 February 2027; and how Korean courts read the second limb of Article 26-2 once it applies.

Frequently asked questions

Are ChatGPT conversations private?

They are private from other users, but not confidential in a legal sense. The provider stores them, may review them in defined safety cases and may disclose them when the law requires. US courts have ordered chat logs to be preserved and produced, and a federal court held in February 2026 that a user's own AI chats were not protected by attorney-client privilege.

Can ChatGPT chats be used as evidence in court?

Yes. In Türkiye, electronic data count as documents under the Code of Civil Procedure, and a court can order a specific, necessary conversation to be produced. In criminal cases, a lawful search of a device can reach chatbot histories. Chats obtained unlawfully, for example by accessing someone's account without authority, cannot be relied on in civil proceedings.

Does deleting a chat or using Temporary Chat make it confidential?

It reduces exposure but does not create confidentiality. OpenAI says deleted chats and Temporary Chats are removed within 30 days unless they must be kept for safety or legal reasons. In 2025 a US court order required OpenAI to preserve deleted chats for several months. Copies already on a device or shared elsewhere are unaffected.

Is it safe to use ChatGPT as a therapist or to talk about personal problems?

It is lawful, and many people find it useful, but it carries none of the confidentiality that protects a session with a doctor, psychologist or lawyer. Avoid details you would not want a third party to read, particularly about health, family members or conduct that could become the subject of a dispute.

If I share my ChatGPT chat with my lawyer, does it become privileged?

No. In Heppner the court held that non-privileged material does not become privileged because it is later sent to counsel, and that typing a lawyer's advice into a chatbot can waive protection over that advice. In Türkiye, the lawyer's secrecy duty covers what the client entrusts to the lawyer, not the separate chat history the client keeps.

Can my employer see my ChatGPT conversations?

On a company business account, possibly: enterprise administrators control retention and may have access. Under Turkish law, any monitoring must still meet the KVKK principles of transparency, purpose limitation and proportionality. A personal account on a personal device is different, although using it for company data may breach company rules.

Burhan Doğuş Ayparlar's View

This section sets out my personal assessment as an attorney at law in Türkiye.

I think Heppner is right about privilege, but most commentary describes it slightly wrongly. The decisive point is not that Claude is a machine, or that its privacy policy permits disclosure. It is that the defendant was not talking to his lawyer. Privilege protects a relationship that carries obligations: the lawyer owes loyalty and secrecy, is regulated and can be disciplined. A chatbot owes nothing of the kind. Extending privilege to it would protect the user's words without anyone being responsible for protecting them.

Where I part company with the decision is its reliance on the privacy policy to defeat confidentiality. If confidentiality turns on the small print, every email account and cloud drive is in the same position, and we do not treat them that way. But that is an argument about privacy and searches, not privilege, and the two should be kept apart.

For Türkiye the point is structural. Our protections are built around the lawyer: secrecy under Article 36 of the Attorneys' Act, the right to refuse testimony, the law-office search rules in Article 130 CMK and the inviolability of correspondence with defence counsel. They do not reach what clients now generate themselves, on platforms abroad, often before engaging counsel. That gap is no reason to invent an "AI privilege". It is a reason for the legislature, which has until 25 February 2027 to replace Article 134 CMK, to build safeguards that fit cloud-held data: a precise search scope, separation of irrelevant material and deletion deadlines.

Finally, for my own profession: the Heppner defendant was trying to prepare for his lawyers, and the preparation itself turned out to be the risk. That explanation belongs at the start of a matter, not after the search.

This article is for information only and does not constitute legal advice. It is based on the Heppner memorandum of 17 February 2026, the Michigan order of 10 February 2026, the Colorado order of 30 March 2026, OpenAI's published statements and privacy policy, press reports of the January 2026 order in the OpenAI copyright litigation, the Turkish statutes cited as published on mevzuat.gov.tr, the Constitutional Court decision published in the Official Gazette on 25 May 2026 and in the Court's decisions database, the KVKK guide of 22 September 2026, the 2026 TBB guide and the Korean Attorney-at-Law Act as published on law.go.kr, all accessed on 20 October 2026. The US rulings are trial-level decisions binding on no other court, and the Korean amendment is not yet in force. Statements about Turkish, Korean and US law are general; specific cases require individual assessment. The analysis and assessments are the author's own.