(By You Yunting) Recently, news that Character.AI (“Character”), a U.S. AI AI companion company, has entered bankrupt and that its data assets may be sold out at auction has drawn wide attention. What is striking, however, is not the bankruptcy but the proposed sale itself. Character’s principal business is an AI character-based companion chat platform, and its platform data includes a large volume of emotional chat records between users and AI—content that is not only private but often pornographic or violent, and at times revealing the darker side of human nature. Yet prospective buyers in the United States are seeking to acquire them. This article examines the legal issues raised by auctioning such records.
1. Why Did Character File for Bankruptcy?
Character’s two co-founders both came from Google, and one of them was a co-inventor of the Transformer architecture that underpins large language models. Google accordingly moved to secure their expertise. In August 2024, Google paid approximately US$2.7 billion to Character: part was a license fee securing a non-exclusive right to use Character’s foundation-model technology, and the remainder funded a buyback of shares from the founders and early investors. After the transaction, several core researchers, including the two co-founders, were recruited by Google DeepMind.
Google took the team and a license to the technology but acquired no equity interest, so Character remained a separate legal entity. Its core research team, however, had been hollowed out, and it faced a raft of high-profile lawsuits. With its team depleted and litigation mounting, Character ran out of cash and commenced bankruptcy proceedings.
2. Why Are Character’s User Chat Records Legally Sensitive?
Character provides AI companion services that create a sense of real human presence and foster user trust. It is precisely this trust that leads users to disclose in their chats with Character’s AI agents a great deal of legally sensitive data:
First, privacy. Users treat Character’s AI personas as trusted confidants and disclose secrets they would not share with another person. Such disclosures are highly sensitive personal information.
Second, personal information. The platform collects users’ names, email addresses, phone numbers, location data, device IDs, session logs and other identifying data. Therefore, every chat session can be linked to an identified or identifiable natural person.
Third, sensitive personal information. Chat content frequently touches on health data, sexual orientation, religious or political beliefs, and minors’ information—all categories of sensitive personal information.
Finally, harmful content. In the course of casework, the author has reviewed a sample of user chat data from emotional-companion AI services and found that much of it was wholly at odds with socialist spiritual civilization. This is because many users feed into the AI negative emotions, dark thoughts or desires that they cannot confide in real life, resulting in chat records rife with material that is pornographic, violent, self-harm-related, abusive or apt to foster excessive dependence, among other things.
3. On What Basis Could Character’s User Chat Records Be Auctioned?
Character’s privacy policy does contemplate the collection, use and transfer of user chat data in connection with a bankruptcy sale. It provides that Character may collect personal information including names, email addresses, phone numbers, and dates of birth; geolocation information; interests and hobbies; financial information; and chat records including audio, video, and images, among other categories of personal information. It may use such information to provide services, analyze user interests, advertise, improve services, train models, process payments, and other purposes.
Character’s privacy policy also contains a further clause covering the negotiation or consummation of any merger, sale of company assets, financing, or acquisition of all or part of its business by another company, under which Character may disclose certain information. Character treats a bankruptcy sale of data as falling within the sale of company assets.
4. What Is the Current Status of the Proposed Auction of Character’s User Chat Records?
That is why, once Character entered bankruptcy proceedings, its user chat data became a saleable asset. What would a buyer do with it? A secondary market for distressed-company data has emerged in the United States, driven largely by AI training. After Spirit Airlines ceased operations in May 2026, Google acquired its 100 million emails and 500 million Microsoft Teams messages at auction for US$10 million; the dataset was then de-identified and used to train AI agents on real-world workflows.
After Character entered bankruptcy, its data—including user chat data—was offered for sale. Concerned about this, the attorneys general of 40 U.S. states, led by the New York Attorney General, filed a joint notice with the court on October 14, 2026, setting out nine conditions for any sale involving user data:
effective age verification; restrictions on minors’ access to chat services; removal of AI characters presented as therapists or physicians; crisis-response protocols; a comprehensive privacy program; notice of any data sale and an opt-out right; no use of private chat content for AI training without valid consent; a right for users to access and delete their data before any sale; and a prohibition on misleading representations
On October 30, the court held a public hearing on the data auction. Before the hearing, the state attorneys general made an additional joint request, asking the court to appoint a consumer privacy ombudsman under the United States bankruptcy code to review the proposed sale of data relating to approximately 12 million users, including minors. The court permitted the asset-sale process to proceed but did not approve the sale of user chat records. The judge also declined to appoint a privacy ombudsman, strongly urging instead that the company retain privacy experts to further assess child-safety and data-protection issues and address the states’ concerns.
5. Why Would Chinese Law Bar the Auction of These User Chat Records?
China has put in place a dense regulatory framework for emerging technologies such as AI. The collection, processing and transfer of user–AI chat data are governed by a suite of laws, regulations and regulatory requirements. Against that backdrop, were Character a Chinese company, an auction of its user chat data for AI training after insolvency would be highly unlikely.
The core regulatory instrument is the Interim Measures for the Management of Anthropomorphic AI Interactive Services (the “Measures”), issued by the Cyberspace Administration of China (CAC) and four other ministries. Character’s AI companion service falls within “anthropomorphic interactive services” as defined in Article 2 of the Measures—i.e. sustained emotional interaction simulating human personality traits, thought patterns and communication styles. Under Article 16(2) of the Measures, unless otherwise provided by law or with the explicit consent of the individual concerned, providers shall not provide user interaction data to any third party.
Only two exceptions could in principle permit such a transfer. In the first, a statutory exception applies—typically where government or judicial authorities require the data for the performance of their duties, in which case Character, as a personal information processor, is obliged to provide it. In the second, the individual concerned (the user) gives explicit consent. Thus, use of the data by a third party for AI training would require the user’s separate consent.
Although Character’s privacy policy does notify users that certain information, including user chat data, may be disclosed in connection with a sale of company assets. But that blanket notice does not satisfy the Personal Information Protection Law(the “PIPL”), which confines collection to the minimum scope necessary for the processing purpose and requires fresh consent where a recipient changes the original purpose or method of processing after bankruptcy.
Under the above provisions, Character may collect chat data only insofar as necessary to deliver its own AI companion services. A buyer at auction, by contrast, would acquire it for AI training. That change of purpose triggers a fresh notice-and-consent obligation before the data can lawfully be transferred. Yet, as noted above, the data is so sensitive that users are unlikely to consent to AI training.
The PIPL also excludes anonymized information from the definition of personal information. Could Character therefore anonymize the chat data and auction the resulting dataset? That is a gray area. The uncertainty is factual: if technical processing genuinely achieves anonymization—so that the speaker can no longer be re-identified, even in theory—the data falls outside the PIPL and most consent requirements fall away.
Even successful anonymization, however, does not resolve the separate question of whether the data may be used as training corpus. Records rife with pornographic, violent, self-harm-related, abusive or dependency-inducing material would, if fed into a large model, likely yield comparable outputs.
Article 4 of the Interim Measures for the Management of Generative AI Services requires providers to uphold core socialist values and prohibits generating content that incites violence, is obscene or pornographic, is false or harmful, or is otherwise prohibited by laws or administrative regulations. Article 11 of the Measures likewise requires that data used for pre-training or fine-tuning have a lawful source and comply with laws, administrative regulations and core socialist values.
Providers do conduct data cleaning before training, and such filtering removes much problematic material. But the chat records reviewed in casework contain a substantial amount of content inconsistent with core socialist values, and it is doubtful that cleaning captures all of it. Ultimately, private chats with AI tend to elicit the darker side of human nature; under the principles embodied in these rules, such material is ill-suited as AI training data.
Finally, users confide in AI because they believe the conversation is private. An AI company that commercializes that trust should bear corresponding data-protection duties, and insolvency should not extinguish them. Users who do not wish their chat records to be treated as saleable assets in insolvency may exercise their right to request deletion and require the company to erase their data.
Disclaimer of Bridge IP Law Commentary
Short Link: