(Inside Science) -- It sounds like a philosophical riddle for actuaries: If there's a car accident, and no one is driving, who pays for the damages?
But when driverless cars hit the road, as is expected sooner or later, such a question won't be hypothetical anymore. For auto insurance companies, it's an issue that threatens to upend their industry.
Human drivers cause 94 percent of car crashes, according to a recent survey by the National Highway Traffic Safety Administration. So eliminating the driver is expected to make autonomous cars much safer. Computers can't fall asleep, get drunk or be distracted.
But the anticipated drop in accidents will likely lead to lower insurance premiums. And driverless cars may spur a growing ride-sharing industry, as people forgo buying their own vehicles, further shrinking the auto insurance market.
These changes could cause insurance companies to lose 80 percent in business by 2040, according to a recent report by Morgan Stanley.
But companies may be able to recoup at least some of their losses if they adapt to an anticipated shift in liability. As drivers take a back seat to machines, experts say that in an accident, the car manufacturer -- not the driver -- will become liable. So instead of buying your own auto insurance, companies like Google, Tesla and others will buy product liability insurance for the cars they sell.
The altered liability landscape will demand a new way for insurers to assess risk. Rather than setting premium prices based on your driving history or behavior, insurance companies will have to evaluate the risk of the car itself.
But no one has established how to do that yet, though doing so will be crucial for determining who pays for an accident, said Cian Ryan, a graduate student at the University of Limerick in Ireland. Without such a system in place, consumers would have to pay for their own damages. "Manufacturers and insurers need to take responsibility now to price these risks, to make sure customers aren't left empty-pocketed in the end," he said.
Ryan is developing a method to assess the risk of an accident by examining a driverless car's behavior, which he said should look similar to a human's reckless driving: swerving, braking hard or accelerating too fast.
The research, which he presented this week at the meeting of the Society of Risk Assessment in Lisbon, Portugal, is preliminary. But it's the beginning of a fundamental change in how insurers will have to think about liability.
Driverless cars may be safer, but they won't be perfect. Last year, a man died when his Tesla's autopilot feature failed to brake before colliding with a trailer. The system isn't fully autonomous, and an investigation cleared Tesla of fault. But even improved and truly driverless cars will crash on occasion, whether it's because of software bugs, malfunctioning sensors or simply bad weather. And some cars will be better drivers than others.
To assess the likelihood a car will crash -- and thus how much an insurance company should charge the manufacturer -- Ryan, along with Finbarr Murphy and Martin Mullins at the University of Limerick, designed an algorithm to monitor a car's behavior, compare it with others of the same kind, and identify any anomalies.
"If a vehicle is speeding or accelerating or braking hard, those are aspects that are going to lead to a crash," Ryan said.
The advantage of tracking behavior is that it accounts for hazards that can originate from every aspect of the car, from software to hardware, he said. Just flagging errors in a sensor, for example, won't suffice, as not every malfunction will threaten a passenger's safety. If an error manifests itself in a few hard turns, however, then that might be something to worry about.
But risky behavior may sometimes be safer, such as if a car swerves to dodge a collision. Not distinguishing a good swerve from a bad one may indeed be a limitation to their approach, Ryan said, although autonomous vehicles should be able to avoid most dangerous scenarios and not need evasive maneuvers in the first place.
In the future, he added, they hope to incorporate anomaly detection systems that monitor sensors and hacking attempts, creating a more comprehensive analysis of the risk.
Tracking behavior isn't new, however. Some insurance companies have already been monitoring drivers with smartphone apps or devices you plug into your car. But the system only allows a driver's risk level to increase or stay the same. Presumably, an aggressive driver's habits aren't going to change on the way to the grocery store.
But Ryan's method does allow a car's risk level to return to normal. This way, he said, it can track how risk fluctuates during its journey. If risk levels remain high for the majority of the time, then it probably means something is wrong with the software or hardware. But if it's high only for a few minutes, then maybe the car is just driving through a temporary storm.
The researchers now plan to apply their algorithm to data collected from 50 normal, human-driven cars. The analysis would provide a baseline to compare with semi-autonomous and fully autonomous cars -- although it's data that car companies aren't so readily willing to give up, Ryan said.
Still, insurance companies should be interested. "They would find it very valuable," said James Lynch, the chief actuary of Insurance Information Institute, a trade group. "It would be an important part of rating a fleet, I would say, if the system worked as promised."
Shifting liability is just one of a host of legal, ethical, societal and cultural changes that will accompany the emergence of driverless cars. But today's cars will last a long time, and technology tends to take decades before becoming mainstream. "Most likely," Lynch said, "the transition will go very slowly over a period of decades."
But that doesn't mean no one's paying attention to the changes. "Insurers," he said, "are watching very carefully."