"For example, participants who were asked whether they agreed that the death penalty should be legal might receive an AI suggestion that disagreed."

They couldn't have come up with a less agreeable example?

This just makes it sound like the introduced bias is real but hey it might be good.