Insurers setting rules for their own use of AI first have to address certain prerequisites including data management, transparency, detecting bias and establishing governance principles, according to AI experts from the life, P&C and brokerage fields of insurance, who spoke at a recent webinar on the ethical use of AI in the insurance industry.
To guard against bias in AI, insurers need to choose the right data or information set, according to Richard Wiedenbeck, chief AI officer at Ameritas, a life insurance and wealth management company. For instance, a bank applying AI to credit card data from the past 70 years would end up with a bias against women. That is because, in the U.S., women could not have credit cards in their own name until the 1974 passage of the Equal Credit Opportunity Act.
In addition, Wiedenbeck cautioned, information sets or data may not be structured – such as collections of PDFs or images – which means the AI may not pick up on all the available information.
Life insurers applying AI to medical records data could also end up with a bias against women because of another societal bias, noted Suzanne Grover, vice president of underwriting at Coastal Wealth, a wealth management solutions provider.
"Pharmacological studies were really only on men for a long time," she said. "What is the impact on women, and how often they are prescribed particular things versus not, and how physicians write that into records – all of those things have to be taken into consideration. Then we're asking an AI model to make sense of all of it. It's really complex, and we've got some work to do on some of those fronts."
State Farm pointed out, in a response to email questions, that it has a "Responsible AI Principles" policy to ensure its use of AI is "compliant, explainable, reliable, secure, and protected."
Newfront, an insurance brokerage, makes "explainability" a core principle of its use of AI, according to Patrick Miller, head of data and AI at the company. "There's a lot of error and uncertainty with AI," he said. "Explainability is the clearest path towards actually building trust with your users."
Another tack insurers can take is pre-emptively establishing AI governance principles, as Frank Neugebauer, vice president, Generative AI, Cincinnati Insurance Companies, explained.
"We don't want to implement AI that is going to be either illegal at some point, prohibited, or otherwise heavily changed," he said. "In property and casualty, we don't even bother talking about AI for pricing selection. Humans do all of that. They have generalized linear models, which are totally different than a neural network for pricing. But everything else that involves pricing and selection, humans are controlling everything."
Cincinnati Insurance Companies chose not to apply AI to hiring efforts – in the form of a resume reader, because of its potential for bias, Neugebauer said. "Language models don't understand all the text," he said. "We'd rather have that inefficiency than the risk of running afoul of the law."
Neugebauer offered a guideline for insurers to use in setting AI policies. "If you're not 100% sure what you're doing, you're not relatively sure of what's going to happen on the other end, the answer should be no," he said. "We should be working towards yes, but working towards yes in a way that balances the outcome that we want. Start with no, work towards yes."
