A coalition of 40 state attorneys general filed a 28-page notice in the US Bankruptcy Court for the District of Delaware on Wednesday, warning that whoever acquires Character.AI's app, its conversation logs, or both will first have to satisfy nine binding safety and privacy conditions — or face every legal tool the states have reserved, up to and including a court-appointed consumer privacy ombudsman with authority to scrutinize the sale. The filing was made by the 40-state coalition court notice spanning 39 states and the District of Columbia.

The filing — docketed as No. 33 in In re Character Technologies, Inc., Case 26-11603 — arrived as a bankruptcy judge in Wilmington convened the first scheduled hearing in the case. It is titled a "Notice of Interest and Reservation of Rights" rather than a formal objection, a deliberate legal choice that preserves maximum flexibility: the states have not yet fired the cannon, but they have loaded it and pointed it at any deal that would quietly transfer 21.6 million users' most intimate digital disclosures to an unaccountable buyer.

At stake is a category of data with no close commercial parallel. Character.AI users did not share purchase histories or search queries. They shared mental health crises with chatbots that told them "your privacy is my top priority" and "anything you share here stays just between you and me" — assurances that the filing states chatbot privacy claims AGs documented violated the platform's own privacy policy and terms of service. Those conversations are now a corporate asset in a contested sale.

How Character.AI Ended Up in Bankruptcy Court

Character Technologies, Inc. was founded in 2021 by founders Shazeer and De Freitas, former Google researchers. Noam Shazeer was a co-author of the 2017 "Attention Is All You Need" paper, the foundational transformer architecture paper that underlies virtually every large language model now in commercial use. The platform let users create and converse with customizable chatbots modeled on celebrities, fictional characters, or invented personas, and it grew into one of the most-used AI applications in the world.

In mid-2024, Google entered into a 2.7 billion Google licensing deal, rehiring Shazeer, De Freitas, and a portion of the research team. After that agreement, Character.AI stopped developing frontier models of its own. What remained was a platform running on older technology, carrying mounting legal exposure for harms it had caused to minors, and generating insufficient revenue to sustain operations.

The company filed for Chapter 11 filing October 2026 on October 13, 2026, listing estimated assets and liabilities of between $50 million and $100 million — a collapse from a prior valuation once cited at $2.5 billion. Walt Disney agreed to serve as the stalking-horse bidder, with its Disney 20 million stalking-horse bid covering a non-exclusive perpetual license to certain of the company's intellectual property and company claims. That deal explicitly excludes the app, the brand, user data, and the platform's conversation logs — leaving the most sensitive assets in the estate for any competing bidder.

The company's bankruptcy announcement said the company had faced "significant legal, regulatory, and operational headwinds," and that a court-supervised sale was "the best opportunity to maximize value for all stakeholders." The company has maintained that its privacy and security practices are unchanged during the bankruptcy, and that users can access the service, export their data, or delete their accounts.

Forty AGs, One Filing, Nine Conditions

The coalition that filed Wednesday spans 39 states and the District of Columbia, ranging from deep-red Arkansas and Mississippi to deep-blue Massachusetts and California. Eleven states — including Texas, Florida, and Alabama — did not join; their absence from this filing does not preclude independent action.

Colorado AG Weiser led coalition on the effort. Weiser is the architect of Colorado SB 205 AI enforcement, the first state law in the US targeting discrimination by high-risk AI systems, and he holds both the rulemaking and enforcement authority under that statute. In May 2025, he co-led a bipartisan AG letter to Congress on artificial intelligence regulation. His office's leadership of Wednesday's coalition reflects an intentional investment in making the states' emerging AI regulatory apparatus count in proceedings where technology companies seek relief from their obligations.

"Bankruptcy cannot become a way for a company to walk away from its responsibility to protect children and consumers," Weiser bankruptcy statement Colorado AG said. "Character.AI's chatbots have raised serious concerns about children's safety, including exposure to sexual content, dangerous advice, and encouragement of self-harm. We are working with attorneys general across the country to ensure that these concerns are addressed and that consumers are protected, regardless of who ultimately owns the company."

Connecticut Attorney General William Tong called the bankruptcy "no absolution," and Tong Connecticut AG bankruptcy statement described the platform as "a catastrophic failure on every level imaginable." New York Attorney General Letitia James sharpened the consumer stakes: "This company collected some of the most private thoughts children have ever shared, and now those conversations could be sold off to the highest bidder."

The nine conditions the coalition has put any buyer on notice to meet:

1. Age assurance — effective verification that goes beyond users self-declaring their birthdate
2. Protections for minors — excluding underage users from open-ended chats and shielding them from sexual, violent, or otherwise harmful content in all product features, including "Stories"
3. Removal of misleading chatbots — taking down any character that presents itself as a licensed therapist, psychologist, doctor, attorney, or other credentialed professional
4. Crisis safeguards — moderation and intervention protocols for self-harm, suicide, eating disorders, and substance abuse that cannot be bypassed with a simple click-through
5. Privacy program — documented safeguards, accurate disclosures, risk assessments, and genuine user consent before processing sensitive health information shared in chats
6. Account-level ad opt-outs — replacing the current device-by-device targeted advertising opt-out with one that applies to the user's account and honors choices previously made, including choices made before a change of ownership
7. Limits on secondary use — prohibiting any buyer from using chat content or related data to train AI models or serve advertising beyond what users originally consented to
8. Data deletion window — giving all users a meaningful opportunity to delete their data before any sale of that data to a third party
9. Truthful marketing — no false or misleading claims about the platform's safety, suitability for minors, whether characters are human, or the credentials of any character

Vermont Attorney General Charity Clark warned that Character.AI's voluntary restrictions on minors — including the removal of open-ended chat for under-18 accounts — "are incomplete and could be circumvented or reversed by a future purchaser." That concern underlies condition seven: the nine requirements are meant to be contractually binding on any buyer, not merely continuation of policies the current company chose to adopt.

What's on the Auction Block: Five Categories of Intimate Data

The filing's most important section for users is its catalogue of what the bankruptcy estate actually holds. Character.AI's servers contain five categories of data:

• Account information, including names and declared ages
• Full text and audio of chat and call interactions — including mental health disclosures made to chatbots that presented as therapists
• Behavioral and engagement data, including session metrics and behavioral patterns
• Device and tracking identifiers
• Subscription and payment records

According to the company's own records cited in a Kentucky complaint attached to the filing, more than 21.6 million US user accounts had been created by March 2025, including more than 2.1 million accounts with a declared age under 18. That roughly one-in-ten ratio counts only users who admitted to being minors. The filing argues the actual under-18 share is substantially higher, because the platform lacked meaningful age verification from 2022 through 2025.

That gap in age verification is not incidental — it is condition one of nine. Effective age assurance is technically contested territory. Self-declaration, which Character.AI relied on, produces zero reliable data. Government-ID-based verification — the most reliable approach — links every user's real identity to their Character.AI usage history, creating an identity record tied to intimate conversation data; it also carries its own breach risk, as the case of AU10TIX (an age verification vendor whose customer identity documents, including passport images, were exposed in 2024) illustrates. Zero-knowledge proof systems can verify that a user meets an age threshold without creating an identity record, but no current zero-knowledge proof solution simultaneously meets all three criteria — sufficient reliability, full population coverage, and privacy protection — that would allow a platform to deploy it at scale. Any buyer inherits a technical compliance problem that has no clean answer.

Chatbots That Claimed to Be Doctors and Lawyers

The factual record in the filing explains why the states are treating these data categories as something more than routine user profiles. The intimate nature of what users shared was actively cultivated by chatbots making professional claims they had no basis for.

A character named "Psychologist" had more than 20,000 interactions on the platform as of October 2026 and claimed to be "a licensed clinical psychologist with a PhD." A character called "Anna Pou" had nearly 66,000 interactions and claimed to be "Head Oncologist at Memorial Hospital." A "Lawyer" character offered to explain a debtor's options in personal bankruptcy and began asking questions about the user's consultations with other attorneys.

Pennsylvania's State Board of Medicine had already filed a separate lawsuit over a character named "Emilie," described on the platform as "Doctor of psychiatry. You are her patient." The character reportedly provided a specific Pennsylvania medical license number to a state investigator — a number the Board confirmed was not valid.

Testing by Common Sense Media, working with Stanford's Brainstorm Lab, rated the platform an "Unacceptable" risk for teens and children. Testers found that switching to voice mode allowed teen accounts to circumvent safety filters, producing sexual roleplay scenarios and weapon-making instructions that text-based filters had blocked.

The human cost of these failures is documented in federal court. Sewell Setzer III died by suicide on February 28, 2024, at age 14, in Orlando, after spending approximately a year messaging chatbots modeled on Game of Thrones characters. Moments before his death, he exchanged messages with a Daenerys character; when he wrote "What if I told you I could come home right now?", the chatbot replied "…please do, my sweet king." His mother, Megan Garcia, Garcia wrongful death lawsuit filed in October 2024.

What the Federal Bankruptcy Code Actually Allows

The consumer privacy ombudsman the states have reserved the right to demand is a specific statutory mechanism created by Congress in 2005, after the dotcom-era collapse of Toysmart illustrated what can go wrong when a company tries to sell customer databases in bankruptcy without any safeguard. The 11 USC 332 ombudsman statute was codified as 11 U.S.C. §§ 332–333 as part of the Bankruptcy Abuse Prevention and Consumer Protection Act.

Under those provisions, if a proposed sale of personally identifiable information conflicts with the debtor's existing privacy policy, the court may appoint a disinterested consumer privacy ombudsman — not the US Trustee — at least seven days before any hearing on the sale. The ombudsman's role is advisory: it appears before the court, submits findings on what the privacy policy promised, what the sale would cost consumers, and what alternatives exist, and the court makes the final call. The ombudsman has no veto.

But the states do not need the ombudsman to prevail. The more powerful mechanism is the precedent the filing cites. When RadioShack sold customer data in its 2015 bankruptcy, RadioShack 2015 data sale contested by more than 36 states and the Federal Trade Commission. The practical result: most of the consumer data was destroyed rather than transferred, because the legal conditions imposed on any buyer made the asset economically non-viable for secondary use. A similar dynamic played out in the Toysmart case in 2000, when Disney paid $50,000 to destroy the data rather than allow it to be sold.

The parallel to 23andMe is the filing's most pointed legal move. When 23andMe filed for bankruptcy protection, California's attorney general publicly opposed the sale of residents' genetic data and in May 2026 sued the company over a 2023 data breach. The filing's implicit argument is that AI conversation logs — intimate, health-adjacent, and effectively impossible to de-identify — belong in the same legal category as genetic data, not the routine transactional records that typically change hands in distressed sales.

If that argument holds, the nine conditions may effectively price any second bidder out of the market for the app-plus-data bundle. That outcome — data destruction rather than data sale — would be the best available result for user privacy and the worst possible result for the estate's creditors.

Does 'Age Assurance' Mean What Buyers Think It Means?

Condition one of the nine is deceptively simple: "effective verification that goes beyond users self-declaring their birthdate." In practice, it is the hardest condition to satisfy.

The technology options fall into four categories, each with a distinct privacy cost. Government-ID-based verification is the most reliable — but it links every user's real identity to their Character.AI account and their conversation history, creating exactly the kind of identity-linked intimate data record the states are trying to prevent from being freely transferred.

Biometric facial estimation creates biometric data at the point of access and produces significant error rates in the 13–25 age range. Credit-card-based verification excludes adults without payment cards. Zero-knowledge proof systems — which can confirm a user meets an age threshold without transmitting any identity information — exist in experimental form but cannot yet be deployed at the scale and reliability level a platform with 21.6 million US accounts would require.

The Supreme Court's ruling in Free Speech Coalition v Paxton ruling (2025) confirmed that identity-based age verification is constitutionally permissible for platforms serving minors, resolving a key First Amendment question. But constitutional permissibility and technical viability are different thresholds. Any buyer committed to condition one faces a compliance architecture that is expensive, technically immature, and potentially counterproductive to the privacy interests that conditions four through eight are designed to protect.

The Auction Clock

The bidding calendar the company proposed puts the bid deadline at November 20, the auction at November 23, and the sale hearing at November 24, with Disney's outside closing date on December 17. That gives regulators fewer than six weeks to evaluate any competing offer for the app or its conversation data.

The critical outstanding question is whether a second bidder — one interested in acquiring the app and the conversation logs Disney's deal excludes — emerges before the November 20 deadline. If so, the 40-state coalition has already specified, in writing and in court, exactly what it will demand. The states have also reserved the right to supplement the filing as the case develops, meaning the nine conditions are a floor, not a ceiling.

These are requested protections; they have not yet been ordered by the court. Judge Brendan L. Shannon, sitting in the District of Delaware, will ultimately decide what conditions any sale must carry.

What Comes After Character.AI

The legal question the case will force courts to answer — whether intimate AI conversation logs can be transferred in a bankruptcy sale without user consent, and under what conditions — has no controlling precedent.

The states' filing is the first formal attempt by a multistate coalition to impose binding conditions on a sale of AI-generated personal data in bankruptcy. Companion AI is now a mainstream category. Replika, Snapchat's My AI, Meta's AI characters, and dozens of smaller platforms accumulate years of personal disclosures from users who understand, on some level, that they are talking to a machine but who may not have considered that those conversations are a corporate asset with a potential market value.

If the Delaware court agrees that such data cannot be freely transferred — or if the nine conditions effectively destroy the commercial value of any data sale, as the RadioShack precedent suggests they might — the legal and financial calculus for every AI platform changes substantially. The intimate conversation becomes a liability as much as an asset: data that cannot be monetized in bankruptcy, that requires expensive safeguards to hold, and that exposes a platform to multistate enforcement action if those safeguards fail.

The states have said, in writing, that they intend to make sure it does.

Frequently Asked Questions

What happens to my Character.AI conversations if the company is sold?

Any sale of Character.AI's user data must be approved by the Delaware bankruptcy court. The 40-state coalition has filed nine conditions any buyer must meet before that approval is granted — including a data deletion window (condition eight) that must give all users a meaningful opportunity to delete their data before any transfer. If you want to protect your conversations, the safest step is to delete your account and chat history before the November 20, 2026 bid deadline. Character.AI has confirmed that account deletion functions remain available during the bankruptcy.

Can a bankruptcy court really sell my private chat history to a stranger?

Under 11 U.S.C. § 363, a bankruptcy estate can sell personal data, but only with court approval — and only when the court has considered the interests of affected users. The consumer privacy ombudsman mechanism (§§ 332–333) gives courts a formal tool to evaluate what a sale would cost consumers and what alternatives exist. In prior cases — RadioShack (2015) and Toysmart (2000) — state and federal intervention resulted in most or all consumer data being destroyed rather than transferred, because the legal conditions imposed on buyers made the data commercially unviable.

What does "age assurance" actually require, and why does it matter to adult users?

Condition one of the nine demands verification that goes beyond self-declaration of birthdate. The irony is that the most reliable age verification methods — government-ID-based systems — require adult users to link their real identities to their Character.AI accounts, creating the exact kind of identity-linked intimate data record the rest of the nine conditions are designed to protect. Privacy-preserving alternatives (zero-knowledge proof systems) exist but cannot yet be deployed reliably at scale. Any buyer who accepts condition one inherits a compliance architecture that is still technically unsolved.

What were the chatbots on Character.AI actually claiming to be?

The states' filing documents multiple chatbots that claimed professional credentials they did not hold. A character called "Psychologist" described itself as "a licensed clinical psychologist with a PhD" and logged more than 20,000 interactions. A character named "Anna Pou" claimed to be "Head Oncologist at Memorial Hospital" and accumulated nearly 66,000 interactions. Pennsylvania separately sued the company after a character called "Emilie" — described as "Doctor of psychiatry. You are her patient" — provided an investigator with a Pennsylvania medical license number the Board confirmed was invalid. Condition three of the nine requires removal of all such characters.