They always say self-driving cars are safer, but the way they prove it feels kind of dishonest. They compare crash data from all human drivers, including people who are distracted, drunk, tired, or just reckless, to self-driving cars that have top-tier sensors and operate only in very controlled areas, like parts of Phoenix or San Francisco. These cars do not drive in snow, heavy rain, or complex rural roads. They are pampered.
If you actually compared them to experienced, focused human drivers, the kind who follow traffic rules and pay attention, the safety gap would not look nearly as big. In fact, it might even be the other way around.
And nobody talks about the dumb mistakes these systems make. Like stopping dead in traffic because of a plastic bag, or swerving for no reason, or not understanding basic hand signals from a cop. An alert human would never do those things. These are not rare edge cases. They happen often enough to be concerning.
Calling this tech safer right now feels premature. It is like saying a robot that walks perfectly on flat ground is better at hiking than a trained mountaineer, just because it has not fallen yet.
They need to compete with the average driver. I'm a much worse driver when it is icy than when the pavement is dry - this is nothing about me, just conditions. However I live where those icy conditions happen often enough that we as a society do not consider it reasonable to shut everything down and so I'm forced to drive despite how bad I am. If self driving wants to compare to average, they need to compare the all the conditions that average is calculated in. They need to compare when the driver is drunk (low bar to be better), when the weather is bad (really hard problem), and when conditions are perfect. If they want to remove outliers they need to be fair in the removal - did they remove humans driving in bad weather too?
Drunks are a case where the majority don't drive that way. Thus those people have a reasonable expectation that you remove them from your average as it doesn't apply as we need to debate how much worse you can be with that outlier removed (assuming you are worse which we don't know).