The Assumption Business

hat would happen if a century of assumptions were available on demand? Actuaries are finding out one prompt at a time. Our cover story reveals a profession leaning into a future where AI agents function as genuine teammates. But this makes the future more likely to resemble the past than we appreciate. Because LLMs formulate from patterns in training data, agents are born of the assumptions of yesterday.

The review of The Zorg presents a stark example of the consequences of passive compliance with assumptions. The murder of 130 or more enslaved people thrown overboard by the crew of a British slave ship in 1781 became a legal matter that treated the murdered as insured property. Underwriters and jurists debated whether the loss was covered, not whether it should have been allowed in the first place. The case caught the attention of abolitionists, but slavery was not abolished in the U.K. until decades later. While professionals in the courtroom executed their duties within the boundaries they knew, none took ownership of the underlying premise.

The Zorg is a historical failure of unparalleled extremity. This month’s AR explores narrower areas where convention risks complacency in the modern day. We examine how a calm hurricane century and wildfire seasons more severe than modeled inform catastrophe models built on too short a record. We consider whether changing case mix over the past decade is a cause or effect of rising litigation. We compare silica to asbestos and explore whether the industry learned from its last solvency event. Climate change, social inflation, and systemic solvency are broadly agreed upon because evidence dictates that they should be. But even if answers stay the same, we should be able to explain why they are still true.

The challenge with AI is that it has no implicit mechanism to do that. It optimizes to the objective it is given, not the one it should have been given, and is indifferent to what follows. The more it is used, the more it risks training on its own output, whose ingredients were yesterday’s assumptions. Actuaries are in the business of assumptions, and the standards require ownership of them. The ownership does not transfer to a tool such as AI. The greatest risk and opportunity live in the tail, where training data is thinnest. Someone has to go looking.