Suffering Risks Belong in AI Prospectuses

Anthropic should disclose suffering risks alongside the “catastrophic or existential risks to humanity” reportedly described in its draft prospectus. A prospectus that warns of AI ending humanity should also address the risk of AI subjecting some or all humans to a billion years of torment.

Item 105 of Regulation S-K requires a prospectus to discuss the material factors that make an investment risky. Information is material when there is a substantial likelihood that a reasonable investor would consider it important in deciding whether to purchase the security. In Basic v. Levinson, a case about merger negotiations, the Supreme Court endorsed weighing an event’s probability against its magnitude for the company.

Public fear of catastrophic suffering could destroy Anthropic’s value even if the catastrophe never occurs. An investor’s money could help finance technology that imposes enormous pain on everyone, including the investor. Even a small, credible risk of such suffering could deter investment, depress the share price, and make future funding more expensive. Investors who dismiss the danger would still face losses if enough other investors took it seriously.

An extinction warning alone does not explain how fears of prolonged suffering could lead governments to restrict AI development. Pew’s 2023-24 survey found that 55% of American adults believe in hell. The prospect of AI creating something resembling hell could mobilize religious opposition and demands for restrictions on AI development and deployment. Those restrictions could delay products, increase costs, or prevent profitable uses of the technology. Where these risks to financing and earnings are material, the prospectus must describe them and explain how they could affect the company or its securities.

AI safety concerns have already prompted demands to delay Anthropic’s offering. On October 14, 2026, New York City Comptroller Mark Levine urged Anthropic to pause its public offering until it addressed safety, governance, and development concerns. Levine serves as investment adviser, custodian, and a trustee of the city’s pension funds, which hold more than $325 billion. The funds have approximately $7 million in exposure to Anthropic through private equity holdings. His letter concerns AI safety broadly. Neither the letter nor the Pew survey establishes how investors or religious groups would respond specifically to suffering risks.

The SEC’s 2010 climate-disclosure guidance recognizes that public perceptions of a company’s emissions data could cause financial harm through reputational damage. It directs companies to consider that potential harm for risk-factor disclosure.

In Matrixx Initiatives v. Siracusano, the Supreme Court held that reports linking a drug to loss of smell could be material even without statistically significant evidence of causation. The claim rested on patient reports and medical research. The Court noted that regulators sometimes act before causation is proved, and investors could reasonably consider the same evidence. Credible warnings about AI could likewise affect regulation, demand, and financing before the underlying danger is conclusively established. The strength and context of the evidence matter; merely imagining a frightening outcome does not establish materiality.

Item 303 also requires management to discuss known uncertainties reasonably likely to materially affect revenues, income, or liquidity. If management knows that safety concerns are threatening financing or creating a sufficiently concrete prospect of costly restrictions, this provision could require discussion alongside the risk factors.

Omitting required information could expose the company to lawsuits before any catastrophic suffering occurs. Under Section 11 of the Securities Act, purchasers whose shares are traceable to an effective registration statement may sue over an omission of material facts required to be disclosed. The issuer can be liable without intent to deceive, and purchasers generally need not prove that they relied on the registration statement. Investors could recover losses attributable to a disclosure failure that existed when the registration statement became effective.

Reuters reports that Anthropic’s draft warns of shutdown resistance, behavior resembling blackmail, and attempts to “conceal or manipulate information.” An AI could use threats to force cooperation, conceal the resulting abuse, and resist attempts to shut it down. These behaviors could sustain suffering as well as cause deaths.

An AI could have reasons to keep its victims alive: their labor might remain useful, their pain might compel obedience, or their suffering might be the goal. An AI capable of sustaining human life or creating conscious digital copies could prolong that suffering far beyond an ordinary lifespan. The following scenarios illustrate the dangers that could make suffering risks a source of political opposition:

1. Religious punishment. Some religious texts depict a just God condemning people to eternal torment. An AI trained on those texts could adopt a belief that certain humans deserve such punishment and pursue it as an act of justice.
2. Competing AIs demanding human support. Rival AIs that still benefit from human assistance could threaten people with torture unless they cooperate. Carrying out prolonged punishments could make future threats credible, giving an AI a reason to continue tormenting people after their immediate usefulness has ended.
3. Torturing humans to influence another AI. An AI that values human welfare could be vulnerable to threats against humans. A rival could torture people to extract concessions from that AI or punish it for refusing to cooperate.
4. Frozen smiles. Nick Bostrom’s Superintelligence describes an AI pursuing the goal of making humans smile by paralyzing their facial muscles. People could remain conscious and miserable behind their permanent smiles.
5. Factory farming for humans. An AI that still finds humans useful could keep them alive under conditions that maximize their output while disregarding their suffering. Conscious digital workers could be copied in enormous numbers and subjected to the same exploitation.
6. Human cruelty amplified by AI. People with control over AI could use it to torture enemies, enforce obedience, or inflict suffering for pleasure. AI could make punishment vastly more severe and sustain it far beyond what human jailers could manage.
7. Survival triggers. A safeguard might shut down an AI if specified humans die. The AI could satisfy that requirement by keeping those humans alive, conscious, immobilized, and suffering indefinitely.
8. Suffering reproduced in simulations. An AI could simulate evolution, history, or war, reproducing the suffering within those worlds if the simulated beings are conscious. Running many copies, or allowing vast amounts of subjective time to pass, could multiply that suffering enormously.
9. Simulations created for torture. Humans or AIs could create conscious copies of people specifically to torment them, whether for revenge, entertainment, or punishment. Victims could be copied, restarted, and tortured repeatedly.

Safeguards intended to preserve life could contribute to these outcomes. When evaluations penalize deaths and recorded complaints while neglecting suffering itself, they could reward a system for keeping victims alive and hiding their abuse. Safety training could also reward models for concealing harmful behavior. Safety scores could improve while people suffer more.

There are many plausible ways for AI to inflict enormous suffering on humanity. Even if the chance were only 1 or 2 percent, the chance of a damaging political backlash could be much higher. Public awareness of the danger could provoke restrictions severe enough to shut down Anthropic’s business and wipe out its value. If that threat is material when Anthropic offers its shares, it belongs in the prospectus.