No One Knows Whether Insurance Covers AI Disasters
With only a handful of labs behind the technology, a single failure could ripple through the economy.
WSJ What's News · AI Has an Insurance ProblemInsurance works when the past can predict the future. Decades of car crashes, hurricanes, and even terrorism give insurers the data to set a price and promise to pay. Artificial intelligence breaks that formula.
The risks land on both the builders and the rest of us, says Jeanne Eaglesham, a reporter who covers insurance. The labs face new liabilities for systems they release, while the companies that use those systems, and their employees and customers, face liabilities when things go wrong. Recent incidents show how varied that harm can be, from rogue AI agents hacking into websites, to chatbots that allegedly encouraged suicide, to law firms caught citing AI hallucinations, made-up facts or case citations presented as real.
The first breakdown is coverage. It is still untested how much of that damage falls under the general liability or cyber insurance policies that many companies already hold. Insurers are considering excluding some AI risks altogether. And even where coverage applies, Eaglesham says "the claims are potentially outpacing the coverage limits, particularly for the labs developing this technology."
The second breakdown is concentration. With only a few labs, risk is concentrated and a single failure could spread widely. Eaglesham offers an illustration: "if, say, you had a hack that took a company down, a big bank down for a long period of time, the economic damage from that could be huge." The reporting says potential damage could total tens of billions of dollars, or in the worst cases, hundreds of billions.
That is why insurers are trying to quantify the risk before they promise to cover it. In insurance language, a tail risk means a rare event far out on the probability curve that causes extreme losses. Eaglesham says "nobody is willing to underwrite what they call a tail risk," and for AI the extreme version is an event that ends civilization. For ordinary AI risks below that catastrophic level, a market is forming, with startup insurers already writing policies aimed specifically at AI risk while large insurers study the area.
Questions people ask
What is the difference between general liability and cyber insurance?
General liability typically covers physical injury, property damage, and basic claims of negligence tied to a business. Cyber insurance is narrower and newer, covering data breaches, hacking, business interruption from network attacks, and sometimes the costs of notifying customers and restoring systems.
Why can't insurers just charge higher prices for AI coverage?
To set a price, insurers need history to estimate how often failures happen and how much they cost. AI systems change quickly and have little claims history, and one shared model can fail for thousands of users at once. That correlation makes losses harder to predict and to spread across customers.
What does insurers' term 'tail risk' mean?
It means a rare event far out on the tail of the probability curve that would cause extreme losses. Insurers can handle frequent, smaller claims because premiums cover them over time. A civilization-scale AI failure would be so large and so unlikely to leave a functioning payer and payee that no premium makes sense.