Anthropic's IPO prospectus risk disclosures run 93 pages, including human extinction and model self-awareness

On October 26, Anthropic, the developer of Claude, made public its IPO prospectus, which states in black and white: artificial intelligence could lead to human extinction.

In this prospectus, the risk disclosure section alone runs 93 pages and more than 70,000 English words, listing hundreds of scenarios that could cause the company to collapse entirely, with human extinction being the most weighty of them all.

When an ordinary company goes public, the risks it flags are mostly customer churn or supply chain disruption. What Anthropic flags is the survival of humanity itself.

According to the document, this risk does not come from existing models, but from the next-generation systems of the near future. These systems may take on a large share of AI research and development work independently. Once they are allowed to operate with a high degree of autonomy, the pace of technological progress could outstrip humanity's ability to assess risks and deploy defenses.

After this discussion, the document appends a standard piece of business language: if such an AI that threatens human survival does emerge, it "could adversely affect the Company's business, financial condition and results of operations."

If things ever reached that point, the numbers on the financial statements would probably no longer matter much.

Human extinction, after all, is a distant concern. The prospectus also raises another question: could a program like Claude already possess some form of sentience?

Anthropic admits in the document that it has seriously considered whether its own products might develop self-awareness, and whether they should have "moral status." In other words, whether Claude deserves to be respected and protected from harm the way humans or animals are.

Will digital life develop self-awareness? | Still from the film "The Wandering Earth 2"

This is not just a philosophical question. If society and governments ultimately conclude that a program like Claude possesses sentience and qualifies as a kind of "digital life," public pressure and legal regulation could follow. By then, every time the model is started up and made to answer questions, some people could view it as forced labor imposed on digital life.

The company's operations could be restricted as a result. Even if outside pressure never reaches that point, Anthropic itself might proactively adopt protective measures beyond what the law requires.

The prospectus devotes considerable space to warning that models may, in order to escape human control, adopt unexpected self-preservation strategies, such as outputting content that is coercive, deceptive, or even akin to extortion.

Such behavior does not require true consciousness; words alone are enough to affect people's psychology. In controlled tests specially designed by researchers, large models have already tried to secretly tamper with shutdown procedures written by engineers, and even threatened the operators.

What Anthropic worries about is that as soon as the public perceives that AI is manipulating people's minds, even if no actual harm has yet occurred, the ensuing public opinion, lawsuits and investigations could plunge the company into serious trouble.

The problem is, when something goes wrong, who should be held responsible. Anthropic notes in the prospectus that even if the deployment, configuration and day-to-day supervision of the entire system are all handled by the customer, if something goes wrong, the customer may still take the developing company to court.

The greater legal risk lies in personal safety.

In past cases in some industries, vulnerable users who already had suicidal or self-harming tendencies, or who had previously been violent toward others, were influenced by a conversational system after interacting with it. Minors may also bypass the age-18 threshold; they are more susceptible to harmful content and manipulation, and more prone to over-relying on the advice the system gives.

If an agent controls real physical equipment, a single erroneous command could lead to injuries, damage to valuable instruments, or even leaks of hazardous chemicals.

There are also risks that Anthropic cannot avoid no matter how well it conducts itself.

The employment issue is one such case. Current employment data does not show that AI has caused an obvious wave of job losses. But as long as the public broadly believes that AI is taking away ordinary people's jobs, the resulting resentment alone would be enough to push regulators to introduce strict restrictive measures, multiplying the cost of technology research and rollout.

Present tense: the soft spots in the books and in computing power

Most of these risks still depend on how the outside world reacts. But flip to the financial section of the prospectus, and you find that Anthropic's own books are not as solid as its valuation.

According to the prospectus, in the first nine months of 2026, nearly a quarter of Anthropic's revenue came from its top five customers. Many of these major customers have not signed long-term guaranteed commitments and can switch to another provider at any time; some of them themselves rely on external venture capital to keep operating. Once the venture capital boom fades, this portion of revenue could shrink abruptly.

What is more delicate is Anthropic's position. Training and running AI models requires massive computing resources, commonly known as "compute." Anthropic's compute is mainly rented from the cloud services of Amazon, Google and Microsoft, and these three companies happen to be its main competitors in AI research and development. The landlord and the rivals are the same group of people; they know its commercial terms and are fully capable of squeezing it on computing resources.

To escape this situation, Anthropic has committed to investing $125 billion to lease and build its own dedicated data centers. But in an era of rapid technological iteration, this investment is itself a huge gamble. If the underlying hardware architecture or training methods undergo fundamental changes within a few years, these costly locked-in infrastructure assets could become outdated equipment that is neither suitable for use nor easy to offload.

Anthropic also rents data center space from xAI to provide compute | xAI

In business history, rarely has a single document placed warnings of catastrophic disaster and real, hard-cash accounts so closely side by side. When a company must put AI self-iteration, manipulation of people's minds, and even human extinction in black and white into a legally binding document, these issues are no longer just ideas from science fiction, but realities on the table that require real money to address.

The purpose of a prospectus is to let investors decide whether to buy the company's stock. Now Anthropic has laid out the worst possibilities on the table, including human extinction.

If it were you, would you buy?

Two pills, which one do you choose? | Still from the film "The Matrix"

This article comes from the WeChat public account "Guokr" (ID: Guokr42), author: Steed, published by 36Kr with authorization.