Skip to content

Character.AI's Chapter 11 Is a Bill for the Characters, Not the Models

Character.AI took in $2.7 billion from Google in 2024 and got a $20 million opening bid for what is left. The filing is a lesson in which half of an AI character business is actually an asset, and which half is a liability with a legal team.

AI-assisted draft. Reviewed and edited by the Phosphene team before publication.

On this page
Rows of blank plaster character busts on a warehouse shelf, one lit by a narrow beam of daylight

Three numbers make the Character.AI filing hard to file away. In 2024 Google paid $2.7 billion to license the company's AI. On 12 October 2026 Character Technologies, Inc. filed for Chapter 11 with assets of $50 to $100 million and liabilities in the same range. Disney's opening bid, as stalking-horse bidder, was $20 million.

That last figure is for a slice of the estate, not the company: selected claims Character.ai holds against third parties, plus a non-exclusive, perpetual license to certain intellectual property. Non-exclusive is doing quiet work in that sentence. Disney did not need to own the character library outright.

A company that took in almost three billion dollars two years ago is now taking a $20 million opening bid for a license to its intellectual property, less than one percent of what it took in. The useful question is which part of the business turned out to be the asset, and which part turned out to be the bill.

Where the money actually went

The 2024 Google deal was reported as an AI licensing arrangement. Read the shape of it and it looks more like a talent repatriation with a license attached. Noam Shazeer and Daniel de Freitas, who started the company after Google shelved their chatbot work in 2021, went back to Google DeepMind, along with around thirty other researchers. The New York Times estimated at the time that Shazeer's personal share was between $750 million and $1 billion.

The money went to founders and early investors. It did not go into a moat. With the research team gone, Character.ai is reported to have stopped developing its own large models and moved to third-party open models. Whatever the platform was in 2024, by the filing it was a chat product with a large character library, a subscription business, and a docket.

Meanwhile Shazeer now works at OpenAI, which is its own small joke about how portable the people were and how fixed the product wasn't. Disney made the same bet in a smaller way: it hired Character.AI's former CEO, Karandeep Anand, as its chief technology officer in the same month it bid on the intellectual property. Take the people, take a license to the assets, leave the platform where it is. That is the whole filing in one staffing decision.

The liability inventory

The official statement describes "significant legal, regulatory and operational headwinds" over the past year. That phrase is a euphemism with a price list behind it. The German report lays out the shape of it:

  • Five suits from families in Colorado, Florida, New York and Texas, with preliminary settlements reached in January without admission of wrongdoing, for sums that stayed secret. Their status is not uniform. In the Garcia case (6:24-cv-01903, M.D. Fla.) the plaintiffs filed a notice of voluntary dismissal with prejudice on 15 October 2026, stating that the settlement agreement had been executed and that all required court approvals had been received; the court dismissed with prejudice on 19 October. Payment is not confirmed in the documents reviewed here. The New York plaintiff's notice is the one with public confirmation of approval on the record: it states her agreement was part of a global settlement with the same defendants and numerous claimants, approved by a California probate court in July 2026, with each individual settlement approved by a referee on 11 October 2026. The claims involved minors who had extensive chats with bots that allegedly encouraged self-harm and suicide.
  • Additional suits since then, plus participation in a consolidated California proceeding for chatbot product liability (JCCP No. 5431).
  • A Kentucky action over harm to minors (Franklin Circuit, 26-CI00029).
  • A Pennsylvania action after a bot presented itself as a licensed physician, with an invented license number (Commonwealth Court, 220 MD 2026).
  • An FTC investigation into unfair trade practices.

In the filing itself, the contested claims from Pennsylvania, Kentucky and at least one unnamed minor are listed at "unknown amount". That is the part of a bankruptcy document that tells you the debt has no ceiling yet.

The petition puts total liabilities at $50 to $100 million across 100 to 199 creditors. The top twenty unsecured creditors are mundane in a way that is almost comforting: Google at roughly $4.2 million, around $2.6 million in unpaid legal invoices, about $2 million owed to DigitalOcean, and at number eleven a food delivery service for $90,000. WeWork is being exited early to cut costs. This is what a company looks like when the model stopped being the problem and everything around it kept costing money.

Why the IP went cheap

A character library sounds like an asset. In this case it wasn't worth much, for three reasons that any studio building characters should recognize.

The first is that the library is largely other people's ideas wearing generated faces. Fairy-tale characters, celebrity likenesses, user-made personas. The provenance is uneven, and the takedown risk is spread across thousands of entries and dozens of rights holders. Disney sent a cease-and-desist letter in September 2025 demanding that the platform stop using Disney's copyrighted characters, and the company complied. So the buyer with the most to lose from a messy catalog is also the buyer who already knows how messy it is. That is consistent with bidding for a license to the intellectual property rather than for the catalog as a going concern.

The second is that the models underneath were no longer theirs. A character platform whose intelligence is rented from open-weight providers has no technical asset to sell, only distribution and prompt history.

The third is that the audience was the liability. A companion chat product with a large teenage user base carries legal exposure that scales with engagement. Every additional daily active user adds revenue and adds risk. That is a terrible shape for a business, and no amount of product polish fixes the geometry.

What Disney's bid is actually telling you

A stalking-horse bid sets the floor in a Chapter 11 sale, and other bidders are invited to beat it. So $20 million is not a valuation, it is an opening position. Read the composition anyway, because it is unusually explicit about where value sits.

Disney's offer covers a non-exclusive, perpetual license to Character.AI's intellectual property, plus selected claims the company holds against third parties. Not the platform, not the team, not the model weights, not the brand. Buying a license instead of the company tells you what the buyer wants: reusable assets, not an operation to run. Nonexclusive tells you the same thing from the other side, since exclusivity is what you pay for when the assets are the reason you are bidding.

That is the split worth internalizing. A character with a world, a voice and a consistent look is content. A platform that hosts thousands of them and monetizes the conversation is a service business with a regulatory liability attached. The content can survive the collapse of the service. The service cannot survive without the content, and it usually does not survive without the audience that the liability comes from.

What this means if you are building with characters

Four rules, none of which need a lawyer to start.

Keep the character bible outside the platform. Names, ages, relationship graph, tone rules, things the character would never say, the visual spec sheet, the reference set. All of it in a repository you control, versioned, with dates. If your character only exists as rows in someone else's database, you do not own a character, you own a subscription. The method for building that document is the same one we wrote up in the timeline that made one face read as a person, and it works just as well when the output is text instead of video.

Export on a schedule that does not depend on bad news. Account export exists as a feature precisely because platforms die. Pull transcripts, generated images, audio and settings monthly, and check that the export is actually readable rather than a zip of HTML fragments. A backup you have never opened is a hope.

Name the model dependency in writing. Which model produced which asset, which version, which date. When a provider changes terms or shuts an endpoint, that log is the difference between a migration and a rebuild. We argued for the same discipline in prompts derived from your own archive, and a bankruptcy is just the version of that scenario with a court date.

Treat safety as a production cost with a headcount. The phrase "operational headwinds" in that press release covers a legal department, a trust and safety team, age verification, and a compliance surface that grows with every feature that makes a character feel real. If your audience includes minors, that cost is not optional and it is not a PR line item. It lands in the liabilities column, and the liabilities column is where this company ended up.

What the filing does not prove

It does not prove nobody wants AI characters. The usage numbers were never the problem, and there were roughly 12 million active users in the filings. It also does not prove that character IP is worthless, since a rights holder has bid eight figures for a license to some of it, subject to higher offers and court approval.

The other number in the filings is the one to sit with. The company was losing $8 million to $9 million a month at 12 million users, with compute and AI salaries as the two big lines. That is not a demand failure. It is a business where scale made the loss bigger. And the bidding has since complicated the tidier version of that story. On 14 November the company named a second stalking horse, Smirks and Pangs Acquisition LLC, at $10 million in cash plus assumed liabilities, for everything outside Disney's licence: the site and both apps, all models and weights, the brand, and the user accounts, user-created characters and chat content. The nine patents split the difference: Smirks and Pangs would own them, but they are the same nine entries Disney's proposed licence covers, and the transfer is subject to licences already granted — ownership passes, the licence survives it. Objections are due 18 November, bids 24 November, and the auction is set for 1 December, so none of that is final and the buyer's owners are undisclosed. Read it as two prices rather than one appetite: the licence for the IP was offered at twice the price of the platform, the models and the data together, and the second buyer is taking on the data and its liabilities with them.

What it proves is narrower and more useful. The expensive part of that business was never the characters. It was running a place for strangers to talk to them, at scale, with minors in the room. If you are building characters to make things rather than to host conversations, you are in a different business with a different risk profile, and the useful lesson is the boring one: own the character, rent the model, and keep the receipts.

For the version of this problem where the character is a real performer's likeness, actors losing roles to their own AI clones covers the consent side. And if you are working out how to charge for a character rather than a conversation, per-generation character royalty is the pricing half of the same argument.

Sources

Share this guide
X LinkedIn