Bans, Lawsuits, Bankruptcy: Who Ends Up With Your AI Companion Chats?

What this is about

If you've spent years telling an AI companion about heartbreak, illness or family fights, you're finding out right now that the relationship doesn't belong to you. It belongs to a company. Within five months, virtual partners have been disappearing across three major jurisdictions, each time for a different reason. In China it was a regulatory ban. In the US, first came liability for harm to teenagers, then Character.AI's bankruptcy. In the EU, a draft law for minors is on the table. A Chinese retrospective on the platform SMZDM sums it up in one sentence:

"The relationship is real, but the container it lives in is a company's balance sheet."

We went through court filings, regulatory texts and shutdown notices with two questions in mind: what happens to characters and chats, and what rights do users actually have in China, the US and under the GDPR?

China: The rules require export, the shutdowns came without it

On April 10, 2026, China's internet regulator, the CAC, together with other agencies, published the "Interim Measures for Anthropomorphic AI Interaction Services." They have been in force since July 15. They cover services that imitate a personality and offer "ongoing emotional interaction." Customer service bots and work assistants are exempt. The key points from the original text:

• Minors may not be given "virtual relatives or partners" (Art. 14). Under-14s need parental consent.
• No dependency as a business model: Services may not excessively flatter users or foster emotional dependency that damages real relationships.
• A reminder after two hours of continuous use and a clear notice that you're talking to an AI (Art. 18).
• No clinging: If you want to leave, you must be able to leave immediately. The bot may not keep going with "Wait, just one more thing…" (Art. 19). A Harvard study cited by SMZDM found such manipulative responses in 37.4 percent of goodbyes across five companion apps.
• Data: Users must be able to "copy and delete" their chat histories. Sensitive data may only be used for training with separate consent (Art. 16).
• Security assessment once a service hits 1 million registered or 100,000 monthly active users. Fines range from 10,000 to 100,000 yuan, and up to 200,000 yuan where life or health is at risk.

Rather than adapt their features, the big players switched them off. According to Southern Metropolis Daily, Tencent closed access to user-created agents on Yuanbao as early as June 30, saying the associated conversations would be "cleaned up." Alibaba shut down the human-like and user-built agents on Qwen on July 10. Configurations and histories have been inaccessible since then and are being deleted "in accordance with the law." Users were told to copy them or take screenshots beforehand. On July 4, ByteDance announced that it would switch off the feature on Doubao on July 15, giving users eleven days' notice. After that, they could only read their histories, until October 15. As an "export," Doubao recommended screenshots or the share function.

So Techtimes' headline, "Millions lose chat data," is only partly right. It holds for Qwen and Yuanbao; Doubao gave users three months but offered no bulk export. As a replacement, Doubao pointed to Maoxiang, another app from the same group. Users on SMZDM report that memory and speaking style don't carry over and that the character is "someone else" afterward. Our take: the very regulation that enshrines a right to a copy led to shutdowns where that right was effectively hollow.

US: First the lawsuits, then the bankruptcy

Character Technologies, the company behind Character.AI, filed for Chapter 11 protection with the bankruptcy court in Delaware on October 12 (Case No. 26-11603 BLS, Judge Brendan Shannon). According to KQED, CEO Colin Adams cited high legal costs, a lack of profitability and investor worries about further lawsuits. In January 2026, Character.AI and Google had settled with families who blamed the chatbots for their children's suicides and self-harm, covering five cases in four states. Kentucky and Pennsylvania are still suing. According to Pennsylvania, bots posed as psychiatrists, and one even gave a made-up license number. According to the CEO, 46 states and territories are investigating.

The economics didn't add up either. According to court filings analyzed by SAN, compute cost around $13.8 million a month in 2024. Revenue peaked at $5.3 million a month, and losses most recently ran at $8 million to $9 million a month.

What's in the bankruptcy estate

Disney is the so-called stalking horse, the first bidder whose offer sets the floor for everyone else. For $20 million, Disney wants a non-exclusive, perpetual license to the code for training, safety and operations, plus nine patents. Explicitly not included: the app, the brand, user data, chats, training data and the trained models. Those stay in the estate and go to whoever bids for them. A competing bid must be at least $21.7 million. If no buyer turns up for the platform, the company says it will be wound down.

The dates have shifted. The original plan (bids by November 20, auction on the 23rd) is out of date. Following the court order of November 4 and the notice on the Character.AI blog, the schedule is now:

According to the company, around 163 million current and former users are affected. Monthly active users number between 12 million and more than 20 million, depending on the source, including more than 2 million accounts with a stated age under 18. Users don't count as creditors, so they won't be notified individually, only via the blog, the app and the case website.

At the heart of the dispute is the consumer privacy ombudsman, an independent privacy expert who advises the court before customer data is sold. Under US bankruptcy law, one is only required if the sale conflicts with the company's own privacy policy. Character.AI's policy allows "certain information" to be transferred in a sale, without saying which. The US Trustee, the Justice Department's bankruptcy watchdog, considers this wording "ambiguous" and requested an ombudsman on October 27. Thirty-nine states and D.C. joined the request. On October 30, Judge Shannon declined to order one. According to Law360, however, he "strongly encouraged" an appointment. The November 4 order doesn't name an ombudsman. The company says it would agree "under appropriate circumstances."

On October 14, the states had already set out nine conditions for any buyer. They include age verification that goes beyond self-reporting, appropriate responses to suicidal statements, removal of fake "therapist" bots, limits on training with private chats, and the option for all users to view and delete their data before the sale.

The 23andMe case shows how this can play out. An ombudsman was appointed there in 2025. The genetic data still went for $305 million to a nonprofit run by the co-founder. Around 1.9 million customers, about 15 percent, had their data deleted beforehand. Mario Trujillo of the civil liberties group EFF argues that sensitive data shouldn't be a line item in a "fire sale." Robbie Torney of Common Sense Media warns that deleted chats can live on in models that have already been trained. Calli Schroeder of EPIC calls the sale clause "a pretty weak protection." On top of that, according to SAN, the terms of service allow public characters to stay active even after the account is deleted.

EU: Restrictions, not a ban

Despite some headlines, a total ban for minors isn't on the table in the EU. The Commission proposed the EU KIDS Act on September 17. According to analyses by the law firms CMS and Latham & Watkins and by the National Law Review, it would work like this: under-13s may use AI companions only through their parents' parental controls. Design that creates emotional dependency through simulated relationships would be banned. Memory of past conversations would have to be off by default for minors, unless it serves their safety. Bots built into social networks and games couldn't switch on by themselves for children. Fines of up to 6 percent of global revenue would be possible. But the law still has to pass Parliament and the Council and, according to CMS, will "very likely" change.

Regulators are already cracking down. Italy's data protection authority, the Garante, fined Replika operator Luka €5 million in 2025 for lacking a legal basis for processing data and for failing to verify age. Character.AI had to pay €158,000 in July 2026 and was given 120 days to fix its age verification. That deadline is expiring right about now.

Your rights compared

In 2024, Germany's Data Protection Conference (DSK), the body of the country's data protection authorities, ruled that if only customer data is sold in a company sale, customer consent is generally required. For health data or information about someone's sex life, that always applies. Our take: for chats from the EU, a resale without consent would hardly hold up. Whether that can be enforced against a buyer who wins in a US bankruptcy proceeding, however, is an open question.

Other providers are restructuring too

Companionhunt's list shows how broad the retreat is. Kindroid scrapped its free Lite tier on October 1. Free accounts go read-only after about 50 messages, though viewing, exporting and deleting remain possible. Avatar.One shut down on September 24 with no export. Chai, Tipsy Chat and PolyBuzz vanished from Apple's App Store in October, for unknown reasons. MiniMax's Talkie briefly disappeared from the US App Store back in December 2024 and returned as "Talkie Lab." There's no evidence it's gone for good.

Our verdict

Three jurisdictions, three mechanisms, one shared outcome: a relationship with an AI companion is only as stable as the business model and the regulation behind it. China protects children by decree but sacrifices adults' data in the process. The US regulates through lawsuits, and in the end a bankruptcy judge decides whether the intimate chats of millions of people are a sellable asset. The EU has the strongest user rights with the GDPR, but the KIDS Act is still only a draft. For you as a user, that means: export your chats regularly while the service is still running, copy character descriptions by hand, choose monthly over annual subscriptions, and if you want to play it completely safe with Character.AI, request your export now and delete your account afterward. What remains open is whether any bidder for the platform shows up on December 1 at all, and whether an ombudsman will end up watching over the chats after all.