The Securities and Exchange Commission published Anthropic's S-1 form on Monday, ahead of a heavily anticipated initial public offering.
Most of the attention has gone to the numbers. But as Ars Technica reports, the filing devotes an unusual amount of space to what could go wrong: across 93 pages and more than 70,000 words, the Risk Factors section discusses hundreds of potential problems.
Many are the ordinary worries of any business: competition, costs, customer acquisition, regulation. Some are not.
The filing addresses risks that could "pose catastrophic or existential risks to humanity", and says such concerns "have been raised both within our organization and by other researchers, regulators, policymakers, and industry participants".
The specific worry is not a current model but near-future systems that could, in the filing's words, "substantially automate AI research and development". If such systems were "allowed to operate with a high degree of autonomy", the company warns, they could advance to a point that accelerates AI progress beyond its ability to evaluate and safeguard new models.
Having set that out, the filing adds that this kind of self-improving AI "could adversely affect our business, financial condition, and results of operations".
Elsewhere the document considers the possibility that its products "possess or could develop forms of sentience or moral status". Widespread acknowledgment of that, it notes, could bring "regulatory requirements, public pressure, or advocacy efforts" that constrain operations, and the company might itself "elect to adopt policies, safeguards or limitations … that go beyond existing legal requirements".
The filing also discusses models displaying "unexpected strategic or self-preserving behaviors", including "outputs that could be interpreted as coercive, deceptive, or manipulative, including behavior resembling blackmail". Perceptions of manipulation, it says, "could generate significant public concern, regulatory scrutiny, litigation risk, or reputational harm, regardless of whether such behavior results in actual harm".
Nor are the risks confined to its own products. The company imagines a case in which another firm's model "take[s] harmful actions independently and at scale without human intervention", producing "broad reputational damage to the AI industry, public backlash, and sweeping regulatory responses" that reach Anthropic regardless.
Among the more conventional disclosures, the filing notes that the company's top five customers, which it does not name, account for nearly a quarter of its revenue for the first nine months of 2026, that many of them "are not subject to long-term contracts with us", and that many "may depend on venture capital or other external financing to fund their operations".
It also observes that Amazon, Google and Microsoft are simultaneously its largest cloud providers and some of its largest competitors, which "creates complex dynamics that could give rise to conflicts of interest and adversely affect our access to compute". Against $125 billion in data centre commitments, it adds that changing requirements could leave it with infrastructure "poorly suited to our actual needs".
A risk factor is a disclosure, and disclosure has a precise function. It tells a prospective investor what could go wrong, so that the investor cannot afterwards say they were not told.
By that test the document is complete. The end of humanity has been disclosed. It appears in the correct section, in the correct register, attached to the correct closing clause about business, financial condition and results of operations.
Readers may feel those three items do not exhaust the consequences. They would be applying the wrong standard. A risk factor is not required to describe everything an event would do. It is required to describe what the event would do to the company, and on that narrow question the filing is accurate, proportionate and, so far as this newspaper can establish, correct.
A document capable of holding the extinction of humanity in the same list as a concentrated customer base and an unfavourable data centre lease has quietly settled a problem that has detained philosophy for some centuries. Everything is now on one scale. The scale is quarterly.
The Department of Unnecessary Research has read all 93 pages and confirms that the disclosure is adequate. Investors have been warned. Everybody else is a third party.