from the privacy-matters dept

What is covered by freedom of thought in the AI age is being decided by courts that are hearing cases about something entirely different. The Chapter 11 filing of Character.ai on October 12 is the second such case in a year, and it shows that the record of how a person thought can be produced as evidence, auctioned as an asset or inherited by a buyer, depending on which dispute reaches that record first.

The picture that fits is a diary written on pages that belong to somebody else. The user types into a window that works as a place to think, and the law treats the window as the property of the company that supplied the pages. Nobody asks whether the stationery supplier owes a duty of confidence before protecting a diary, because the diary is protected for what it contains, namely people’s inner thoughts. A chatbot user has no such luck: when the company fails, its creditors can claim whatever the estate holds, and the trustee inventories the notebooks like any other asset.

A U.S. insolvency estate now holds the confidences of millions of users

Character.ai filed for Chapter 11 protection in Delaware and says it will keep operating while a court-supervised process sells the business, with a close expected in late November or early December 2026. Disney is the stalking horse bidder, offering $20 million for a non-exclusive, perpetual license. According to a notice filed by forty U.S. State attorneys general, that license excludes the app, the user data, and the chat content, so the most intimate part of the company stays in the estate unless another bidder wants it. The company’s lawyer told the court that no buyer has yet made an offer for the user data or the chatbot business.

The privacy policy anticipated this moment. It allows chat transcripts, voice recordings, and images to be disclosed in connection with a sale, and it does not require a buyer to follow the original terms and conditions a user signed up for. In its filing, Pennsylvania claims the company had more than 20 million monthly active users, and the Kentucky complaint, cited in the multistate notice, counts more than 2.1 million U.S. accounts with a declared age under 18 out of 21.6 million as of March 2025. The notice furthermore argues that many minors were recorded as adults, as they did not give their actual birth date when creating a profile.

The attorneys general appeared before Judge Brendan Shannon on October 14 to ask that any buyer protect those chat histories. New York Attorney General Letitia James said the company had collected “some of the most private thoughts children have ever shared.” The states reserved the right to seek a consumer privacy ombudsman, an approach that was used in the 23andMe case.

The OpenAI litigation reached the same result through copyright

This is not a phenomenon limited to bankruptcy. Similarly, nobody in the New York Times v OpenAI copyright case launched in 2023 turned freedom of thought into a key element, even if it was raised at the fringes of the debate. In May 2025, the court ordered OpenAI to preserve output logs that would otherwise be deleted, and Sam Altman answered by calling for an “AI privilege,” arguing that talking to an AI should be like talking to a lawyer or a doctor. The duty to preserve new logs ended on September 26, 2025, and excluded logs from requests originating in the European Economic Area, Switzerland and the United Kingdom from continued preservation. On January 5, 2026, Judge Stein affirmed an order to produce 20 million de-identified conversations, finding that the magistrate judge had adequately balanced users’ privacy against the relevance of the documents. In other words, the chat conversations were just business as usual under the logic of ordinary operation of discovery in a U.S. court.

The individual scale followed in February. In United States v. Heppner, Judge Rakoff held that 31 documents a defendant generated with Claude were not privileged, because Claude is not an attorney and the consumer terms defeated any expectation of confidentiality.

Sam Altman and Judge Rakoff disagree on the result but look at the issue from the same perspective: to them, protection depends on what the machine is. Privileges protect relationships between humans and professionals who owe enforceable duties, so a chatbot cannot qualify. In the present bankruptcy case, the point is moot, as privilege only governs what a court may hear. It says nothing about what a company may keep, sell, or hand to a buyer, and that is the question that should be at the core of Character.ai bankruptcy.

Protection has to attach to the act of thinking, not to the file it leaves behind

International human rights law offers two zones. The forum internum, the inner realm of thought, is absolutely protected. The forum externum, the realm of expression, is subject to limitation. The binary was drawn when thinking left no trace and was limited to the inner workings of our brain. In a paper published in July 2026, I propose the creation of a third zone, the forum interstitiale, for external acts that perform the function of thinking, such as a prompt, an abandoned draft and a confession made to a machine. 

Applied to the Character.ai estate, such an approach would guarantee predictability, as the protection would exist before the user types, not after a bankruptcy judge balances interests case by case. 

Bankruptcy procedure handles personal data as a line in a customer list. A diary has no place on that list and a privacy policy cannot supply sufficient safeguards. 

Europe has the same gap and a second way to widen it

Europe should not feel smug about Delaware. The e-Evidence Regulation has applied since August 18, 2026, and providers now have 10 days to answer a European Production Order and 8 hours in an emergency. The rules cover content data and reach services that store or process user data as a defining part of what they offer, a description that includes AI chatbots. A year of conversations with a chatbot is content data. The negotiators of that instrument were not thinking about chatbots, which is the pattern that Delaware and the Southern District of New York have already shown.

The one moment in the OpenAI saga when European residents fared better was the exclusion of requests originating in the European Economic Area. That exemption is, however, triggered by where a request came from and has nothing to say about what the user was doing.

So what now?

Four rules would stop the diary from becoming an auction lot:

1. Bankruptcy courts should appoint a consumer privacy ombudsman in the Character.ai case before any bid for the chat logs is considered, and should decline to treat boilerplate sale clauses as consent to transfer the records of users, and especially children’s private thoughts.
2. Legislators should put one screening question to every AI instrument: what happens to user conversations when the company fails, when a court asks for them and when a regulator pushes the provider to filter them.
3. Providers should retain less, keep identity data separate from content and offer modes in which nothing readable exists to hand over, so that they can say so truthfully to a court, a trustee or a buyer.
4. Litigators should argue the functional frame, which asks what the person was doing, and abandon the relational frame, which asks what the machine is.

None of this means that chatbot conversations should be beyond the reach of law. Evidence of wrongdoing may need to be preserved. Allegations of harm to children must be investigated. Legitimate competing interests do not disappear because someone used AI.

But neither should those interests automatically reduce the record of a person’s thinking to an ordinary corporate asset or an unrestricted source of evidence.

The law, in most cases, already understands that a diary deserves protection because of what it contains, not because of who manufactured the notebook. AI conversations require a similar change of perspective.

Caroline De Cock is a communications and policy expert, author, and entrepreneur.

Filed Under: bankruptcy, data, freedom of thought, privacy
Companies: character.ai, openai