Anthropic's IPO filing prices in risks to investors, not society

By Olivia Hart

Anthropic’s IPO prospectus devotes 69,800 words, about 30% of its 240,000-word filing, to risks facing the company and its shareholders. Hart frames the document against the original purpose of modern prospectuses: to correct railway-age manias and tell capital providers what could go wrong with their money.

Even as Anthropic describes the prospect of an age of miracles, it catalogs catastrophic risk, job displacement, rogue agents and industry power concentration. Anthropic also warns that its own public statements about AI risks “may be used against us in litigation.”

These are risks to Anthropic. Hart’s central criticism is that the filing does not systematically count what Anthropic’s products could cost everyone else, including whether shareholders may eventually pay that bill.

What does Anthropic disclose?

Anthropic is a Delaware Public Benefit Corporation. Delaware law requires a periodic statement on how it serves people materially affected by its conduct, and Hart cites Anthropic’s public-benefit report and Los Angeles Times reporting in arguing that its disclosures describe effort rather than results.

The company’s charter requires it to balance the “financial interests of shareholders with the benefit of the broader public,” Hart notes. Its prospectus, however, concentrates on harms that could become Anthropic’s own costs.

Alongside Oxfam, Hart maintains the Register of AI Risks & Remedies, a living record with 103 entries drawn from investigations, court proceedings, media and technical reports. It ranks risks by severity, identifies who bears the cost and assesses whether remedies are ready.

The Register focuses on people already bearing AI’s costs, mostly without consent or return. Hart points to Amnesty International’s account of a person mistakenly flagged by a fraud algorithm in the Netherlands, a study indexed by PubMed of a Polish doctor whose skills eroded through algorithmic overreliance, and TIME reporting on a Kenyan content moderator labeling toxic content.

Anthropic discloses investor risk in detail but does not systematically identify external victims.

How can weak oversight distort the business?

Workers and productivity come first. The Register finds a recurring pattern of overreliance and deskilling wherever AI gains authority and human supervisors lose it, a risk CSET examines as automation bias. Anthropic’s prospectus mentions overreliance once in passing, does not mention deskilling and acknowledges that human oversight can fail under pressure to improve productivity, Hart says.

A Public Risk Statement would need to identify who checks the AI and whether those people have the time, records and authority to overrule it. Hart argues that productivity gains are oversold when the cost of checking the work is ignored.

Anthropic expects customers to deliver economic gains that support its revenue. If those gains depend on poor oversight, or wrongly harmed people are owed compensation, the gains shrink and Anthropic’s sales figures overstate durable value, Hart says.

Anthropic acknowledges that its monitoring does not cover every deployment, but the prospectus does not set out the full range of possibilities. Hart calls the subsidy hidden and warns that its invoice could arrive later; investors would need to separate real productivity gains from gains that future claims could erase.

AI revenue appears more durable when Anthropic counts customer productivity while ignoring the labor and liability behind it.

Who is already paying for AI failures?

Three settlements quantify external AI costs.

• The ACLU reports that Detroit paid Robert Williams $300,000 after a flawed facial recognition match led to his wrongful arrest.
• The published SafeRent settlement terms show that tenant screener SafeRent reached a $2.275 million settlement for discriminatory screening.
• Anthropic’s settlement documents show that it reached a $1.5 billion settlement to compensate authors and publishers over pirated books.

These cases put civil-rights, housing and copyright costs on AI users and suppliers. As wider use exposes more people to mistakes, Hart argues, precedents can increase demands for compensation.

External AI harms already carry financial consequences, even when Anthropic’s filing does not assign them to the company.

Can existing law make those costs stick?

None of the three settlements required a new AI law. For about four out of five of the 103 harms in the Register, an existing law already applied, Hart says, identifying a gap between legal protection and the practical means to obtain redress.

The gap can narrow without federal action. The Transparency Coalition’s 2026 State AI Legislation Report says lawmakers in 30 states enacted 115 AI-related laws in 2026, a nearly 58% increase from the previous year.

Yahoo reported that eight state attorneys general are suing OpenAI over harm to children. The Federal Trade Commission is investigating frontier laboratories’ safety lapses, TechRepublic reported, while the European Union is launching AI Act enforcement, Euronews reported.

The Financial Stability Board is also urging banks to strengthen AI governance. Hart expects enforcement to be slow, but says AI-related losses will become a cost of doing business when affected people have both the right and the means to demand payment.

Existing law can already reach AI harms, provided affected people can use it.

What should Anthropic disclose before going public?

A Public Risk Statement would reveal what Anthropic knows about its products’ uses, who could be harmed, who is responsible and what affected people can do. That would let investors judge how much of the business depends on costs carried by others.

Hart says law does not require the document, although Anthropic’s chosen public-benefit purpose does. She also argues that shareholders may eventually foot the hidden subsidy of AI development.

The forward test is whether Anthropic discloses those external liabilities before its public offering, rather than allowing courts, regulators or compensation claimants to assign the price afterward. Until then, investors have detailed disclosure of Anthropic’s risks but little systematic disclosure of AI’s wider costs.