Can a company entrust a task to AI if no one can cover the consequences of a mistake? Anthropic’s stock market filing brings this question to the fore. But there are several levels of risk between a wrong answer and a catastrophe beyond the capacity of insurance. Distinguishing between them changes how we read the warning from Claude’s creator.
What Anthropic actually says
In its S-1/A form filed with the SEC on 2 November 2026, Anthropic discusses the potential financial consequences of unexpected behaviour by its models. The company warns that its insurance coverage could prove inadequate, particularly in the face of disproportionate damage caused by one or more incidents.
It adds that a catastrophic AI-related event could exceed the capacity or expertise of an insurer, or even the sector as a whole. The conditional matters: this passage describes a possible exposure. It provides neither a measured frequency nor a quantified probability that would support announcing an imminent catastrophe.
This warning appears in a document intended for investors. Anthropic had announced the initial filing on 26 October, for a proposed stock market listing under the symbol ANT on Nasdaq. The amended document consulted remains preliminary: this filing should not be confused with evidence that a listing is already in effect.
AI insurance does exist, however
The warning does not mean that no use of artificial intelligence is insurable. In its aiSure documentation, Munich Re presents cover for defined risks involving AI systems, including generative AI. The insurer mentions production errors and certain liabilities associated with hallucinations or intellectual property, in particular.
Its FAQ explains that the assessment covers how the model works, its data, the stability of its performance and its monitoring. Updates and changes in behaviour are also included in the evaluation. The aim, then, is not to insure “intelligence” in general, but to examine a system and how it is used.
This commercial documentation demonstrates that an offering exists, not that all applications are automatically accepted. Nor does it allow us to conclude that a policy would cover unlimited damage. Targeted cover and an inability to absorb an extreme scenario can perfectly well coexist.
The decisive leap: from answering to acting
To understand what is at stake, let us take a fictional example. An AI summarises a document and gets a date wrong. If an employee checks the summary before sending it, the mistake may have no consequences. If the same summary automatically triggers an order, the questions change: who controls the order, what amount can be committed, and how can it be cancelled?
This scenario is not an incident attributed to Claude. It helps separate three elements that are too often conflated: the quality of the text produced, the permissions granted to the system and the consequences of a bad decision. A convincing answer, by itself, tells us nothing about the latter two.
We can extend the example to a tool designed to respond to customers. Suggesting a reply to an adviser and directly sending a commercial commitment are two distinct tasks. It may be the same model and the same sentence, but there is not the same opportunity to correct it before the customer receives it.
The effects of scale deserve as much attention as an isolated error
Another hypothetical situation: an incorrect rule is applied to one file, then to ten thousand. The initial mistake may be identical; the scale of the problem is no longer the same. For an organisation, the right question is therefore not only “does the model make mistakes?” but “how far can the same mistake spread before it is detected?”.
This invites us to examine the entire chain: input data, response, decision, action and the possibility of reversal. A check placed only at the end may come too late if an irreversible action has already been carried out. Conversely, approval before a payment or the deletion of data intervenes precisely where the consequences change in nature.
These examples do not quantify a risk or prove the effectiveness of any particular safeguard. They offer a practical framework for avoiding the reduction of a service’s reliability to the ease of its interface.
Useful questions to ask before delegating a task
For a team considering an AI tool, a few questions can provide a starting point: what exact task will be entrusted to it? Which mistake would be acceptable, and which would not? Who checks the result? Can the system only make suggestions, or can it also act? Is there an auditable record and a way to correct the effects of a bad action?
Insurance comes afterwards to complement this thinking. The presence of an offering on the market tells us nothing about a particular company’s policy. Conversely, a provider’s warning about extreme scenarios does not mean that every everyday use entails damage on that scale.
The value of the Anthropic filing lies in bringing the conversation back to a very practical responsibility: when a system makes a mistake, someone must be able to contain its effects. Assessing this scope task by task is more useful than setting supposedly infallible AI against a technology declared uninsurable across the board.
Archive photo: Dario Amodei, on the right, at TechCrunch Disrupt in September 2023.
