Ironically, this is the flaw of any automation process that replaces a complex set of actions. A human driver is more likely to make mistakes, but the human mind is also going to be more adept at reacting to irregular and unconventional circumstances than a machine. This is why driverless technology is really only feasible right now for applications where the "fail-safe" protocol involves the car stopping itself completely. This works in an automated parallel-parking process, but it will never work at freeway speeds.
The biggest difference that the research found, however, is fault. So far, driverless cars were to blame for zero of the crashes.
...
The cars are too perfect. They don’t give the same cues as a human driver. But this problem will be solvable.