On October 14, 2026, New York Attorney General Letitia James announced that she and 39 other bipartisan state attorneys general have filed a notice of interest in Character.AI's bankruptcy case, seeking protective conditions on consumer data sold through the proceeding, and disclosed a multi-year, multi-state investigation into the company. It is the first time US state enforcers have collectively asserted data oversight over an AI companion firm's bankruptcy estate, meaning hundreds of
According to The Verge, James is one of 40 bipartisan attorneys general who filed a notice of interest in Character.AI's bankruptcy case. Such filings do not block a sale directly; they place regulators' positions on the docket so that prospective buyers, creditor committees and the trustee learn about the regulatory exposure before bidding and negotiation begin. In her statement, James also disclosed that her office's inquiry into Character.AI did not begin with the bankruptcy and has been running for years with multiple states participating.
The core demand concerns data assets. The attorneys general want protective conditions attached before Character.AI sells or transfers consumer data in bankruptcy. Because the platform has long skewed toward teenage users, minors' conversation data is flagged as a priority concern. In a standard liquidation, user data is booked as an intangible
Why now? Three preconditions matured at the same time. First, the core asset of an AI companion product is not model weights but accumulated conversation history, character personas and user profiles. Model architectures can be copied through open weights or licensing, but a two-year record of late-night conversations cannot be rebuilt — it is both the source of stickiness and the line item most easily packaged for sale in a liquidation.
Second, US bankruptcy law has long defaulted toward monetizing consumer data rather than protecting it. The RadioShack and Toys R Us bankruptcies both sold customer lists and email addresses, drawing consumer-group protests but never producing a uniform
Liu Gong's take: within the next six months, any buyer of Character.AI's user data will almost certainly have to accept data-isolation, purpose-limitation and deletion covenants in the purchase agreement, or the deal will struggle to clear state-level review. This is not a routine bankruptcy dispute; it is the first time an AI companion product has been forced to answer in court whether user data counts as a liquidatable asset.
From a technical standpoint, conversation data is far less divisible than outsiders assume. Dialogue history, character-card settings, affinity scores and memory vectors are entangled: character settings alone cannot reproduce a personalized experience, while raw transcripts need re-anonymization and re-labeling before they can be used for training. Liu Gong has seen the same problem in industrial visual inspection — when defect-image datasets move between factories, re-labeling and de-identification often cost more than the data is worth, and buyers end up collecting fresh samples instead. Consumer companion data carries an extra burden: it is laced with self-disclosures from minors, so a single leak triggers state enforcement. The gap between book value and post-compliance net value is therefore large, and liquidation models typically do not price that gap.
On the business side, the asset-splitting logic of the estate changes completely. In a conventional liquidation, a user list is among the easiest assets to monetize. After a 40-state notice, the buyer of a data package assumes not just the purchase price but remediation costs, class-action exposure and reputational spillover. Rational bidders will shift their offers toward model weights, trademarks and app distribution — the clean assets — and price the data package as a liability rather than an asset. Liu Gong predicts that within 12 months, other AI companion companies will voluntarily add bankruptcy data-disposition clauses to their terms of service and financing documents, and M&A due-diligence checklists will gain a dedicated line for companion-data provenance.
For the industry, the real spillover is that this route bypasses legislative gridlock. Congress has failed for years to pass federal privacy law, so state attorneys general are effectively writing rules inside a bankruptcy courtroom. Following the 23andMe precedent, the likely path is that the bankruptcy court approves a conditional data disposition: advance notice and opt-out rights for users, separate handling rules for minors' data, and compliance commitments from the buyer. Once that template hardens, it becomes the reference standard for other AI insolvencies. For smaller AI vendors, compliance stops being a pre-launch review cost and becomes a pricing factor running through financing, M&A and exit — which matters far more than whether Character.AI itself can repay its creditors.
- Google's non-exclusive technology licensing deal with Character.AI in August 2024 was valued at roughly $2.7 billion, with founder Noam Shazeer and several core researchers returning to Google; the bankruptcy shows that licensing did not resolve the company's standalone commercial viability.
- The signatories are 40 bipartisan state attorneys general, covering 80% of the 50 US states; earlier regulatory action came mostly from single states, such as the Texas attorney general's 2025 investigation.
- Character.AI publicly reported roughly 20 million monthly active users in 2024, with average daily usage exceeding two hours; the cost of migrating and de-identifying companion data scales
For the industry, the near-term shift (within three months) is deal structure: the previously bundled package of user data, model and brand may have to be broken up, with the data portion subject to separate buyer qualifications and compliance commitments, lowering recovery rates. Over 12 months, that conditional-disposition template will be reused in other AI insolvencies, and data compliance will move from a legal footnote to a core input in M&A pricing, spawning dedicated data due-diligence and compliance insurance products.
For ordinary users, the near-term visible change is a clear notice window: the trustee will have to inform users before any data transfer and offer opt-out or
Among competitors, OpenAI, Meta and several companion-AI vendors declined to comment on specific terms, though some have emphasized their own data-hosting and minor-protection safeguards in investor communications to distance themselves from Character.AI. Some prospective bidders said privately that if the data package carries long-term compliance covenants, their offers would be cut sharply, and they would rather buy only model weights and brand.
Experts are split. Privacy scholars generally read the 40-state notice as using bankruptcy court as a de facto legislature — effective but unpredictable. Bankruptcy practitioners worry that piling on data conditions depresses overall liquidation
Questions worth exploring:
- How user data is recognized and written down in bankruptcy: booking conversation history, character personas and user profiles, and provisioning for compliance costs
- Comparing 23andMe and RadioShack: how consumer data has been disposed of in US bankruptcies, and what threshold triggers state intervention
- From SB 243 to a 40-state coalition: whether companion-AI compliance costs become the final push driving smaller vendors out
Further reading: 2026 Frontier LLM Comparison: A Practical Guide — a durable, evergreen reference. Bookmark it.
This item comes from The Verge AI (published 2026-10-14 18:47:54). Wenfeng Radar provides a Chinese/English summary and commentary on overseas AI developments; the original content is owned by its author.
Daily aggregation of overseas AI developments with in-depth analysis. Bookmark us to never miss an important signal; back to homepage.