(TNS) — Many artificial intelligence systems are “trained” by interacting with human behavior and human-created information. Now, Washington University researchers have found that humans change their own behavior when they know their actions are being used to train AI. And not only do people change, but those changes can last into the future — creating new habits in the human trainers.
And tendencies or biases in a person’s behavior, including those the person isn’t even aware of, can change, too. So, who is training whom, here? “This study clearly shows us that we need to understand these behaviors of people interacting with AI, specifically when they are helping train these tools, so that we can measure that bias and mitigate it,” said Dr. Philip R.
O. Payne, director of the WashU Institute for Informatics and a professor of medicine. But there could be a downside to people improving their behavior if their actions are being used to train AI.
Lead researcher Lauren Treiman, a WashU graduate student, said that if people were helping train an AI for self-driving cars, for example, they might drive especially carefully. This might make the AI a perfect driver. However, in a place like St.
Louis where people often run short yellow lights, it might be more dangerous for a self-driving car to try to be the perfect driver. “AI might need to learn to run yellows, if that’s what people tend to do,” she said. Treiman said she first started thinking about the effe.