Anthropic's IPO filing details risks to investors but ignores societal costs

The scope of the prospectus

Anthropic is preparing for an initial public offering (IPO) by releasing a massive prospectus to potential investors. As reported by Fortune, the document reaches 240,000 words in total length. A significant portion of this filing is dedicated to explaining what could go wrong for the company's owners.

The company spent 69,800 words—roughly 30% of the entire document—detailing various risks. These include catastrophic risks, the displacement of jobs, the rise of rogue agents, and the concentration of power within the industry. These points are framed primarily as potential costs to Anthropic's shareholders rather than the public.

The filing even includes a warning about the company's own transparency. Anthropic notes that its public statements regarding AI risks might be used against the firm in future litigation. This level of candor serves to inform capital providers about the legal hurdles they might face as the company scales.

The missing societal costs

While the prospectus focuses on financial risks, critics argue it ignores the costs borne by the general public. A research collaboration with Oxfam created a Register of AI Risks & Remedies to track these impacts. This register includes 103 entries of harms that the current prospectus does not fully account for.

These external costs include incidents like a Polish doctor whose skills eroded due to over-reliance on an algorithm. Other examples include a person in the Netherlands being wrongly flagged by a fraud algorithm and a Kenyan content moderator dealing with traumatic work. These are real-world consequences that the prospectus treats as secondary to investor interests.

The author of the Fortune piece suggests Anthropic should file a second document called a Public Risk Statement. This would go beyond financial hazards to describe who is being harmed by the technology. It would clarify what the company knows about its products and who is responsible when things go wrong.

Productivity and deskilling

One major issue identified in the Register is the pattern of over-reliance on AI and the resulting deskilling of human workers. This happens when people meant to supervise AI systems lose their authority or ability to do the work. The prospectus mentions over-reliance only once in passing and does not mention deskilling at all.

The author argues that AI productivity benefits can be overstated if the cost of checking the work is ignored. If human oversight fails under pressure to meet speed goals, the resulting errors could create long-term problems. Shareholders need to know if the company's growth is built on genuine gains or on a lack of proper supervision.

If these gains depend on poor oversight, the revenue Anthropic generates might be overstated. The author suggests that if compensation is eventually due to those wrongly harmed by the technology, the company's sales numbers could shrink. This creates a hidden subsidy where the public carries the risk while the company collects the profit.

Legal precedents and liability

The Register of AI Risks shows that many harms do not require new laws to be addressed. For about four out of five of the 103 documented harms, existing laws were already on the books and applied to the situation. This means that legal protection exists, but the practical means to get paid for damages may still be developing.

Recent examples of these costs include a $300,000 payment to Robert Williams after a flawed facial recognition match led to his arrest. Another settlement reached $2.275 million for discriminatory tenant screening. Anthropic itself reached a $1.5 billion settlement to compensate authors and publishers over pirated books.

As more precedents are built, the demand for compensation is expected to rise. Even without new federal action, state-level legislation is increasing, with 30 states enacting 115 AI-related laws this year. These losses will eventually become a standard cost of doing business as the legal machinery moves forward.

The public benefit mandate

Anthropic is unique because it is a Delaware Public Benefit Corporation. This status requires the company to provide a periodic statement on how it serves those materially affected by its conduct. Currently, the company's disclosures on this topic describe its efforts rather than its actual results.

The company's charter specifically aims to balance the financial interests of shareholders with the benefit of the broader public. Because of this specific purpose, the author argues the company should be more transparent about the risks it poses to others. This would allow investors to see how much of the business depends on costs that others are currently carrying.

The author concludes that it is only fair for the company to be public about these subsidies before it goes public. By revealing what it knows about product uses and potential harms, Anthropic could provide a clearer picture. This would ensure that shareholders are not surprised when the bill for these hidden costs eventually comes due.