Will insurers apply AI oversights in time?

A sometimes-overlooked cyber exposure stems from how companies — including insurers — are using AI.

Businesses are becoming the targets of class action lawsuits prompted by inappropriate AI-generated outcomes, observes Kirsten Thompson, partner and the national lead of the privacy and cybersecurity group at Dentons LLP in Toronto.

“Insurers were very slow to adopt electronic and digital means, and then they all sort of moved as a pack at the same time to adopt artificial intelligence,” Thompson said during a Denton’s insurance media briefing last November. “Their first use cases were mostly around fraud, fraud prevention, detection, etc., because they have great data sets….

“We’re now starting to see it used in claims and claims adjudication, which is a higher-risk area. We started to see class actions come out of the U.S. in the area of health, where AI was used in the process of adjudication. It denied claims because the AI was looking at its data sets and basically saying, ‘No, old people are a risk. We’re going to deny the claims.’ And that was, to the AI’s mind, a perfectly reasonable thing [to do].”

Other AI applications

Insurance companies are also testing the use of AI in underwriting and pricing. Sources say there is a potential exposure if the pricing outcomes of AI are not supervised, particularly if the pricing is based on small data sets.

Theoretically, AI could spit out a premium of less than a dollar for a data set that included only a handful of drivers who have no claims history, as one broker tells CU.

And, at the National Insurance Conference of Canada last September, Intact Insurance encountered an opposite outcome while testing AI for pricing auto policies.

“I’ll tell you about a funny example — the regulator might not think it’s funny,” Brad Neilson, vice president of personal lines pricing at Intact Financial Corporation, said at an NICC panel addressing underwriting trends. “Very early in our process of exploring machine learning, we had a case where an auto comprehensive premium was generated of $4 million. I think even in the high-theft market, that’s probably too high.

Related: Do cyber policies cover AI-generated crimes?

“We did a deep dive into what was going on here. Getting into the nitty-gritty of the data, an assumption was made that the person was 95 years old. Not many [drivers of that age are] on the road, so you have limited data. And you had a model that was overfit to this limited data.

“I think when you scrub all that out, there were points in the modelling process when the business and the data science teams needed to come together to make sure these assumptions didn’t lead to that output. And, of course, testing and whatnot can catch those things as well.

“But I don’t think I can overemphasize the modeling expertise you need to build up in your company before you go full speed on deploying some of these models.”

This article is excerpted from one that appeared in the February-March print edition of Canadian Underwriter. Feature image by iStock/ValeryBrozhinsky