It seems pretty clear that if the vehicle becomes capable of doing real active driving then there should be the equivalent of a human driving test. Only the fea

It seems pretty clear that if the vehicle becomes capable of doing real active driving then there should be the equivalent of a human driving test. Only the features that pass the test can go into the vehicle, so as per your example - an algorithm that cuts people up when merging would simply fail to pass the driving test and wouldn't be acceptable for deployment. I think the likeliest case for full self driving is almost the opposite of what you're describing. Because it's all code and it's reproducible and testable the car company is liable if they put out something that speeds or drives dangerously. The result will be FSD cars all driving at the speed limit, leaving appropriate breaking distances etc. - driving much more conservatively than the equivalent human driver. It's also about incentives, you as the consumer might want a car that drives a hundred miles per hour and weaves in and out of traffic, but if the car company is the one facing the liability for it then it's not going to happen.