Watch out for these AI misreads of claims images

Takeaways:

• False positives and negatives in analysis of claims images
• AI follows object shapes without thinking through what they actually are
• Well-trained image analysts needed to review AI determinations

AI still has far to go to reliably and accurately assess damage claims with aerial and satellite imagery, according to David Heathcote, head of intelligence at McKenzie Intelligence Services, a London-based insurance intelligence service.

As insurers set guidelines for their use of AI, and imagery and data advances change insurers' methods for assessing risk, they struggle with image elements that corrupt AI output about claims.

"You will have a lot of false positives, you'll have a lot of false negatives that you aren't going to be able to identify very easily," said Heathcote, who spoke in an InsTech webcast on May 13. "Even some of the better AI tools say they're getting an 80% accuracy rate. If I had an analyst who I was paying and employing, who 20% of the time was wrong, I would fire that person."

Heathcote gave a few examples of mistakes AI frequently makes when reading aerial imagery:

• Automobiles. Going by shapes, AI recognizes cars as property, but can designate them as buildings instead.

• Baseball diamonds. Because these have a common shape, AI reads them as a structure that has been destroyed, since they are flat with no elevation.

• Swimming pools. AI may designate these as undamaged just based on their shape, ignoring what a person can see reading the image, such as discolored dark water because a flood has passed over the pool, or ash from wildfires that has polluted the water.

The more common AI mistakes Heathcote described can be easily caught by anyone, but more expert human analysis is required when images include obstructions, he added. Electro-optical images, from sensors detecting visible light, can be obscured by cloud cover or wildfire smoke. However, synthetic aperture radar (SAR) – applicable to both satellite and aerial images – can cut through clouds to show landscapes and structures underneath, solving those obstruction issues.

Still, even with the innovation of SAR, expert human review of AI assessments of the images is necessary, according to Heathcote. "There are certain times where you need someone who has a serious amount of skill and expertise in imagery interpretation," he said.

When McKenzie conducted property assessments in the days right after the L.A. wildfires of January 2025, even casual views of some images were enough to reach conclusions about claims, Heathcote said. However, McKenzie staff members with U.K. military experience, including imagery analysis training, interpreted more complicated images.

Insurance companies want quick answers for claims, so McKenzie aims to use images to understand loss events as quickly as possible, Heathcote said. The right combination of AI and human review can accomplish this.

"You need a balance. There are positives with both. There are negatives with both," Heathcote said. "A combination of the two gives you the most coverage and allows you to access the most data."