The study was led by Helene Cymek of the Technical University of Berlin, who has previously studied social loafing involving human partners. “When two pilots are in a cockpit monitoring the dashboard, they each reduce their effort. [Airlines require] two people because they want to increase safety. They call this the four-eye principle. But we found social loafing,” Cymek said.

For the new study, she and her colleagues recruited 44 volunteers to inspect electronic components for manufacturing errors such as bad welds or seams—a task that, in factories, is often performed by humans paired with robots. The volunteers were divided into two groups. One worked alone, while the other was told to double-check components that had already been inspected by a robot named Panda. People in that group were shown Panda—an articulated arm with a visual sensor on the end—on their way into the lab and heard it humming along as they worked.

But in fact, there was no robot at work. Both groups were deliberately given a set of components that included the same number of mistakes, so if the two groups were giving the same degree of attention to the task, their results should have been roughly the same. Instead, the researchers found that the humans working alone picked up an average of 4.2 out of 5 errors, while those who thought they were being assisted by a robot detected an average of 3.2—20% worse.

It’s not a huge difference, but if you think about quality-control teams where humans and AIs work together on medical imaging or aircraft navigation, it’s clear that the phenomenon of social loafing could potentially carry a high cost. For instance, Cymek notes, it’s known that in mammography screening, it makes a difference if a radiologist is checking an image for the first time or double-checking someone else’s work: “If the person knows it’s been checked first, they slack off.”

This finding doesn’t exactly boost my confidence in self-driving cars, even if there’s a human at the wheel to take over in an emergency. “People take advantage of the support that is offered—they’re over-reliant on the system,” Cymek said. “They look but do not see.”