On July 19, 2024, at 1:09 AM Eastern time, CrowdStrike started distributing a software update to 8.5 million Windows computers that were using the company’s Falcon cybersecurity sensor. By early morning, banks were unable to process transactions, hospitals were rerouting patients, airports were grounded, and broadcasters had gone dark. A faulty content configuration file was included in the upgrade. The devices it came into contact with crashed into blue screens that required human intervention to recover. Later estimates put the losses to Fortune 500 businesses alone at $5.4 billion. Insurance claims came in swiftly. The insurers subsequently began closely examining the wording of their policies.
The faulty file was pushed by an automated system rather than a massive language model producing incorrect information, hence AI was not directly responsible for the CrowdStrike outage. However, it illustrated a theoretical concern of the insurance industry: what happens when a single software component, deployed concurrently across thousands of businesses, fails in a way that simultaneously causes associated losses across different sectors? That risk differs structurally from a single building fire or a localized flood. The fundamental actuarial premise that individual losses should be essentially independent of one another is challenged by this type of systemic accumulation event.
In several aspects, modeling the AI versions of this problem is more challenging. The harm that an AI system causes when it hallucinates—that is, produces confident, particular, incorrect outputs—depends totally on the environment. A legal AI exposes itself to malpractice when it references made-up cases in a brief submitted to a federal court. When a medical AI misidentifies a diagnosis and influences a treatment choice, it can lead to more significant problems. When a customer-facing chatbot gives inaccurate product instructions, it raises product liability issues that the current policy language was not designed to address. In response to these eventualities, carriers have mostly taken a practical approach: instead of trying to price risks they are unable to adequately estimate, they are expressly incorporating that goal into policy wording and excluding or rigorously sublimating AI-related losses up front.
Since 2024, the Cyber Market Association of Lloyd’s of London has been providing its syndicates with guidelines for AI exclusion language. Three of the biggest reinsurers in the world, Munich Re, Swiss Re, and Allianz, have all updated their underwriting questionnaires to include information about the AI tools used by an insured organization, whether any AI systems are making decisions on their own without human supervision, and the governance practices surrounding the behavior of AI models. These are not scholarly tasks. They are an effort to provide a risk picture for exposure categories for which there is currently no historical data in the actuarial tables. Loss histories, or years’ worth of claims data that inform actuaries of the frequency and cost of particular incidents, are the basis for standard insurance pricing. Large-scale AI failures are too recent for that data to be available in significant amounts.
The category of deepfake fraud serves as an example of the definitional issues that occur when completely new types of loss are combined with preexisting regulation frameworks. After taking part in a video conversation with AI-generated deepfakes of the company’s CFO and other senior colleagues, a finance employee at a Hong Kong corporation deposited $25 million in January 2024. The wording of the policy, which was developed before real-time synthetic video in corporate fraud was considered, determines whether such loss triggers a cyber policy, a criminal policy, a social engineering endorsement, or none of the above. The cases are making their way through the claims procedures and, occasionally, the courts. Before the precedents are fully established, insurers observing that process are deciding on coverage for every comparable policy they currently have on file.
A limited but expanding range of specialized AI liability solutions is emerging from the specialized end of the industry. Businesses like Coalition, Cowbell, and a number of Lloyd’s syndicates are testing enhancements to current cyber policies that expressly handle AI operational faults as well as stand-alone endorsements for AI failures. These goods are priced with a large degree of uncertainty built into the premium, are early-stage, and have little capitalization compared to the possible exposures. However, their presence indicates that a portion of the industry is attempting to anticipate the issue rather than just ignore it. Another layer is added by the EU AI Act, which designates some AI applications as high-risk and creates new liability frameworks. Before full enforcement begins, carriers are attempting to determine what compliance risk looks like for their insureds, but they do not yet have the regulatory clarity to price it reliably.
