A self-driving car doesn't have to anticipate stupid drivers, it can keep appropriate distance and planning needed regardless of the inferred skill of the drivers. For example, why is anyone not exiting the freeway even in the lane that you know will get 30% bad mergers 500 meters ahead?
And frankly, yes, it's not hard to spot the stupid drivers and you could probably have algorithms for that. But the obvious stupid drivers aren't the danger, it's the good ones missing something as you won't be expecting them to do something idiotic. It's the biker you see slowing down and classifying as 'smart' who then rams straight ahead when you're looking in the other direction. It's the merger who perfectly accelerates up to the right speed, starts blinking, and then just doesn't see you. It's the guy who's stopped at the same red light a thousand times, but who's brain played a trick on him the 1001's time and had him think the green light for going ahead was the one that applied, not the red one for the left turn he just blew straight through.
A smart car will assume that everyone is stupid and that it, itself, has to be capable to counter any physically possible action. We should, as well, but the fact is that we're constantly hampered by our intelligence, assuming things that fit the general rules, anticipating based on experience, projecting states and emotions onto everything else. We shorten distances as we know that will make the driver ahead of us remember that he's forgotten to get out of the passing lane (or get uncomfortable enough to move out of the way). We maintain speeds that are usually ok, even in rain, when its dark, and our visibility is half of our stopping distance. If there's a stupid pedestrian about to pass we'll notice on the bump-bump. When we're the fourth car about to merge in to the freeway, do we stop and wait for the three idiots doing the bumper-to-bumper merge to get onto the freeway? No, most of us will at best give them a slightly wider berth, but most likely we'll see if we increase distance a bit and plan a higher acceleration and cut two lanes and get past them as fast as possible.
A well programmed smart car should know better than to get into a situation where an accident is inevitable, and it would at least theoretically be better at that. And the fact is, when the choice comes and it is inevitable I'd rather trust the car, because frankly, I've never, _ever_ been in that situation and I have absolutely no idea what I'd actually do or any time to think about it. The car, at least, might have knowledge of the statistics, its own capabilities and the physics of the situation.