While March’s men’s and women’s basketball tournaments brought unexpected excitement, a new set of topics has captivated the attention of property/casualty insurance professionals. These issues, recently highlighted in Verisk’s emerging issues bracket, could remain central to industry discussions for years to come.

Verisk, a global leader in data analytics, hosted an engaging one-hour webinar on June 10 to spotlight the top concerns voted on by industry experts. Despite fierce competition from topics like climate change, microplastics, and infrastructure, artificial intelligence (AI) and generative AI (Gen AI) secured the top spot again. As AI continues to evolve, with states increasingly looking at AI-related legislation, concerns about edge cases and AI “hallucinations” are emerging as significant risks.

AI Growth Sparks Legislative Action

AI’s rapid technological progress has driven state legislatures across the U.S. to take action, despite the risks still being largely uncharted.

According to Laura Panesso, associate vice president of government relations at Verisk, policymakers have responded with remarkable speed in drafting AI-specific laws and regulations. While most states have proposed some form of AI regulation this year, only a few have actually passed these laws into legislation, Panesso noted during the webinar, which can be accessed on Verisk’s website.

“This trend shows the growing recognition that we need to regulate a technology evolving faster than we can fully understand it,” Panesso remarked.

The National Association of Insurance Commissioners (NAIC) developed a bulletin outlining how insurers should handle the development, acquisition, and application of AI in decision-making. It also urges the adoption of testing methods to detect bias and discrimination. So far, 24 states have embraced this bulletin, while California, Colorado, and New York have introduced their own specific AI regulations.

Without a national framework for AI oversight, Panesso explained that states are creating their own regulatory paths to manage the challenges posed by the emerging technology. To date, 40 states have either introduced or passed some form of AI legislation, ranging from studies on AI’s impacts to regulating certain aspects of the technology.

Key topics discussed in the webinar included regulations on AI deployment, ownership rights for data used in training models, and concerns over algorithmic pricing and discrimination. On the generative AI front, regulators are focusing on the use of deep fakes and the creation of intimate images, with several states already establishing boundaries for how Gen AI systems can be used, Panesso added.

For instance, Utah recently passed Senate Bill 26, which addresses the use of generative AI in consumer interactions. The law requires businesses to disclose when Gen AI is used, holds companies liable for violations of consumer protection laws, and provides a safe harbor for certain disclosures.

Edge Cases and Gen AI Hallucinations

While AI is celebrated for its innovation, it is far from flawless. Greg Scoblete, a principal with Verisk’s emerging issues team, highlighted two critical areas: edge cases and statistical errors.

Edge cases are instances where a specific AI model has insufficient data for a particular scenario, making them prime opportunities for AI systems to cause damage or errors.

AI and machine learning are integral to many automotive safety features, and as vehicles become more autonomous, these technologies are expected to expand. However, as Scoblete explained, these advancements also introduce risks.

“We’re already seeing reports of edge case errors leading to accidents,” Scoblete noted. For example, in the U.K., a luxury vehicle equipped with adaptive cruise control reportedly accelerated to over 100 mph in a 30 mph zone after its object recognition system misinterpreted a road marking as a signal to speed up.

In the U.S., a vehicle with adaptive cruise control collided with the top of an overturned truck. A federal investigation suggested that the AI image recognition system failed to recognize the top of the truck as an obstacle because it wasn’t part of the model’s training data. This represents an edge case, an unlikely but dangerous situation that AI isn’t prepared to handle.

“The key difference between AI and humans is that, while we might not have encountered the top of a truck before, we’d intuitively understand that driving into one is a bad idea,” Scoblete explained. “With AI, though, we can’t predict how it will react when faced with an edge case.”

Beyond edge cases, Scoblete discussed the issue of “hallucinations” in generative AI. These are errors that occur not from missing data but from the way a model generates an output. While generative AI often produces accurate results, it isn’t infallible.

So far, at least 121 legal professionals have filed briefs with errors attributed to generative AI tools. This supports findings from two Stanford University studies, which revealed that generative AI could produce a considerable number of mistakes, especially when dealing with legal topics.

“We need to ask ourselves: Is the legal field the only profession where precision and accuracy are critical? And is it the only sector under pressure to increase efficiency with generative AI?” Scoblete posed.

Moreover, a data set from George Washington University reveals that at least 11 product liability lawsuits related to generative AI have been filed. Scoblete pointed out that there’s a growing debate on whether the current product liability laws designed for physical products should also apply to virtual AI products.

“One major issue here is the nature of the injury,” Scoblete continued. “As AI becomes more embedded in physical products like vehicles, its potential to cause property damage or personal injury may increase.”

The full scope of these risks, especially as AI technology becomes more widespread, is yet to be fully understood.