The price of a correction: the IMF prices an AI downturn, Character.AI goes bankrupt and Disney buys, Anthropic curbs model copying

Three signals this Tuesday. The IMF puts a number on what an AI correction would cost the world. A chatbot star files for bankruptcy, and Disney buys the pieces. And Anthropic says it has nearly stopped rivals copying its models from the inside.

Signal one: the correction gets an official price

On October 13, in Bangkok, the International Monetary Fund published its World Economic Outlook and its Global Financial Stability Report. Among the forecasts sits a scenario every board should read.

A steep drop in AI investment, triggered by disappointment over productivity gains or the technology's profitability, could cut world economic output by 1.1% by 2028. That includes a 1.8% drop in US output.

The Fund still forecasts 3.1% global growth for 2026, slightly up from the 3% in its July forecast, with AI infrastructure investment carrying part of that growth. The warning is about what happens if revenue does not arrive on time. The message is blunt: if AI monetization is called into question, investments will lose value, debt will become hard to service, and future revenue commitments may never materialize.

Why it matters

The AI boom changes category. It leaves the technology section and enters macroeconomics. The same investment that props up growth forecasts today is the exposure that could pull them down tomorrow.

The IMF is not predicting a crash. It is saying the economy's plumbing now runs through AI: capital spending, credit, valuations, growth forecasts themselves. I connect this to the previous week: Firmus abandoning its listing, OpenAI's revenue losing 20 billion in a methodology note, SpaceX borrowing 40 billion for chips. Each episode was a clue. The IMF puts a number on their sum.

What changes for companies

The question changes. It used to be about adoption: are we using AI yet? It becomes: which part of our plan depends on AI revenue arriving on time?

Hiring plans, valuations, purchasing budgets and partnerships built on AI-linked multiples need a downside scenario, now. The signal is no invitation to fear. It is an invitation to price the delay as well as the promise. A plan that only holds if everything arrives on time is a fragile plan.

Signal two: the chatbot star files for bankruptcy, and Disney buys the pieces

On Monday, October 12, Character.AI filed for Chapter 11 bankruptcy protection. Disney agreed to pay 20 million dollars for some of its assets, plus a perpetual non-exclusive license to part of its intellectual property.

Disney becomes the stalking-horse bidder, meaning its offer sets the floor price of an auction where others can bid higher. The company reported assets and liabilities of 50 to 100 million dollars each. The sale is expected to close in late November or early December. The platform keeps running during the sale. If no buyer comes forward to operate it, it will be shut down.

Five years

Character.AI was founded in 2021 by two former Google engineers, Noam Shazeer and Daniel De Freitas. In just over a year it reached a one billion dollar valuation after a 150 million dollar round. In 2024, Google licensed its technology and rehired its founders.

Then the legal trouble piled up: a wave of lawsuits alleging harm to teenagers, including claims linking the platform to teen suicides. The company settled five cases earlier this year.

Disney's irony

About a year ago, Disney sent the company a cease-and-desist letter over chatbots imitating its characters. Last month, Disney hired Character.AI's former chief executive, Karandeep Anand, as its first chief technology officer. Now Disney is bidding for the technology itself.

What changes

A safety failure is an existential risk for a company, well beyond a public relations problem. For a companion product, trust is the product. Once families, regulators and courts stop trusting, no valuation survives.

Character.AI's technology worked. What was missing was governance of the relationship it created with vulnerable users. It is the hardest lesson of the year for anyone building an AI product that talks to people: the nature of the relationship is part of the product, and it has to be designed with as much care as the model.

For brands, the lesson is direct. If you build with characters, companions, or anything aimed at young people, safety sits on the balance sheet. It gets funded, tested and documented like a production line.

And Disney's move deserves attention: buy the technology and the talent, leave the brand and the liabilities behind. In the AI economy, the pieces of a failed company can be worth more than the company. For Disney, which owns the most loved characters in the world, the ability to make them converse safely is worth far more than 20 million dollars. Provided it builds the trust Character.AI could not.

Signal three: the model learns to defend itself

On Monday, October 12, Anthropic said new safeguards against so-called distillation attacks have cut the data harvested from its Claude models by 90 to 99%.

Distillation means querying a rival's model millions of times to copy its capabilities into a cheaper model. The copying model learns by imitating the answers, and above all the reasoning, of the copied one. It is far cheaper than training a frontier model from scratch.

What Anthropic announces:

• reasoning tokens extracted by copiers are now 10 to 100 times lower than before;
• using proxy networks to repeatedly probe Claude's protected reasoning has become harder;
• tens of thousands of accounts tied to distillation campaigns have been banned;
• a false positive rate below 0.1%, meaning very few legitimate users are blocked by mistake.

The company named a Chinese start-up, Z.ai, also known as Zhipu, as an example of a company that tried to distill Claude's capabilities while developing its GLM 5.3 model. Z.ai did not immediately respond. I note the coincidence: GLM 5.3 is the model Mistral cited a week earlier as the benchmark for its own score.

Anthropic compares the system to the safeguards that stop Claude from helping make biological weapons. That comparison is the message. Copying a frontier model is now treated as a security threat, no longer only as a competitive tactic.

What changes

The competitive moat moves. The frontier is about who trains the best model, and also about who stops it being copied. A model's reasoning is its crown jewel, and the API is the door. Every lab must defend that door while keeping it open for legitimate users.

For brands, the lesson is concrete. If your advantage runs through a model you do not own, ask how that model protects its reasoning, and how it protects what you put into it. Competition now runs through enforcement and countermeasures as much as through research papers. The AI race has a counterintelligence chapter, and it just got its first public scoreboard.

I also note the tension with the October 6 signal, where Mistral sold sovereignty through openness. Two philosophies are taking shape: one publishes its weights and sells availability, the other closes the door and sells the frontier. A company choosing a vendor is also choosing between those two promises.

The thread of the day

The boom gets an official price for failure. A chatbot company collapses because trust ran out. A lab learns to protect what makes it valuable.

Three forms of the same maturity: AI is entering the phase where people count what it costs when it fails, and what it is worth when it is protected.

The noise filtered. The signal kept.