The modern prospectus evolved as a correction to the manias of the railway age. Its job was to tell the people putting up the money what could go wrong for them.
Anthropic’s prospectus does that job with unusual zeal. Even as the AI lab promises the possibility of an age of miracles, it spends 69,800 words, about 30% of its 240,000-word filing, detailing the risks of getting there.
Catastrophic risk? Check. Job displacement? Check. Rogue agents? Check. Industry power concentration? Check. All such harms appear in the prospectus as they might become a cost to Anthropic. Even its candor about risks makes the list: Anthropic warns investors that its own public statements about AI risks “may be used against us in litigation.”
But the agentic age needs a second correction. The prospectus counts what AI could cost Anthropic’s shareholders. It doesn’t count what Anthropic’s products could cost everyone else, or whether that bill could one day land on those same shareholders.
In a research collaboration with Oxfam, I have assembled a Register of AI Risks & Remedies, a living record—with 103 entries of AI risks currently—drawn from investigations, court proceedings, media and technical reports. The Register prioritizes risks by severity, identifying who bore the costs, and readiness of remedies. If you were to think about Anthropic’s prospectus as a document that has framed the risk to capital providers who line up regardless of the warnings in the hopes of eye-watering returns, think of our Register as one that frames the risk to everyone who has already borne AI’s costs, mostly without consent or return.
There is a second little appreciated reality about Anthropic that makes things particularly interesting: it is also a Delaware Public Benefit Corporation, and Delaware law requires a periodic statement on how it has served “those materially affected by the corporation’s conduct.” Anthropic’s disclosures on this front describe effort rather than results. It’s worth noting that its charter is to balance “financial interests of shareholders with the benefit of the broader public.” To abide by this charter, I would suggest that the company file a second document: a “Public Risk Statement”, if you will.
What could go into this statement that hasn’t already been noted in the 69.8 thousand words of the Risks section of its prospectus? You can run an audit to note what’s in our Register but not in the prospectus and find many answers. Examples? Think about someone in the Netherlands mistakenly flagged by a fraud algorithm, or a Polish doctor whose skills eroded because they over-relied on an algorithm, or a Kenyan content moderator who had the traumatic job of labeling toxic content.
Three considerations are key to this second proposed document. The irony is that not only do they highlight the costs borne by third parties, they are eventually relevant to future investors as well.
First, the impact on workers and worker productivity. Our Register finds a pattern of overreliance on AI and deskilling, wherever AI acquires authority and the people meant to supervise it lose theirs. Automation bias runs through a wide range of AI-aided decisions. The prospectus mentions overreliance once, in passing, and deskilling not at all. It acknowledges that human oversight can fail under pressure to improve productivity.
A Public Risk Statement would have to say what has worked: who checks the AI, and whether they have the time, records and authority to overrule it. AI’s productivity benefits can be oversold when the cost of checking the work is ignored – and shareholders need to be made aware of that.
Second, Anthropic is expected to make money because customers experience economic gains from productivity. But if these gains depend on poor oversight or compensation is due to those wrongly harmed, the gains shrink and Anthropic’s sales numbers are overstated. AI has benefited from a hidden subsidy and its invoice could come due. The prospectus does note that the company’s monitoring “does not cover all” AI deployments but it does not specify the entire spectrum of possibilities. Investors need to distinguish revenue built on genuine productivity gains from gains that might get erased in the future.
The city of Detroit paid Robert Williams $300,000 after a flawed facial recognition match led to his wrongful arrest. Tenant screener, SafeRent, reached a $2.275 million settlement for discriminatory screening. Closer to home, Anthropic itself reached a $1.5 billion settlement to compensate authors and publishers over pirated books. Civil rights, housing, copyright, each involved costs to AI users and suppliers. Wider use can expose more people to mistakes. As precedents build, demand for compensation may rise.
Finally, the Register reveals a surprise, which could accelerate such demands for redress. None of these settlements required a new AI law. For about four out of five of the 103 harms documented, an existing law was on the books and it applied. The gap that exists is between legal protection and the practical means to secure redress, but this gap could narrow as the legal and policy machinery gets moving. Even if Washington does not act, there is AI-related legislation coming from the states. The Transparency Coalition’s 2026 report found legislators in 30 states enacted 115 AI-related laws this year—a nearly 58 percent increase from the previous year. Eight state attorneys-general are suing OpenAI for harming children; despite the federal government’s distaste for taking any action, even the FTC is investigating frontier labs’ safety lapses; elsewhere, the EU is launching AI Act enforcement; and the Financial Stability Board is urging banks to strengthen AI governance. It will be a slow process but AI-related losses will become a cost of doing business when someone has the right and the means to demand payment.
A Public Risk Statement would reveal what the company knows about its product’s uses, who could be harmed, who’s responsible, and what affected people can do. Investors will judge how much of the business depends on costs others are carrying. The law doesn’t require this second document, but Anthropic’s chosen purpose does. Also, at some point Anthropic’s shareholders may have to foot the bill for the hidden subsidy of AI development, so it’s only fair that it goes public about this subsidy before the company itself goes public.
The opinions expressed in Fortune.com commentary pieces are solely the views of their authors and do not necessarily reflect the opinions and beliefs of Fortune.