It is entirely possible that self-driving vehicles will not be any better than human drivers, and they'll just commit different categories of errors. It may be that it really does just take general intelligence to drive a car, and that perfection is impossible in the current environment and infrastructure.

I don't think "this is a clearly a development phase" means anything when there's no guarantee self-driving cars will actually reduce deaths. It may be that the way forward is simply more advanced driver assist technology that better accounts for human psychology, and not fully autonomous cars. All these sacrifices for a future of autonomous driving may simply be a waste.

Personally, I don't think these cars should be on public roads until they get a lot better. We're just introducing a new class of accidents; human drivers generally make the same kinds of errors and other drivers can reason about what they're thinking. Computers are more likely to fail in (to humans) surprising and bizarre ways. What we have now is the worst of both worlds.

As an aside, if we really wanted to do something, I suspect the easiest way to save thousands of lives a year right now would be to implement mandatory functionality on all smartphones to shut off the screen when the phone is moving faster than 5mph. This would inconvenience people and passengers would be annoyed, but there was, after all, a time before smartphones.

> shut off the screen when the phone is moving faster than 5mph

I like this idea

Maybe need to figure out how to avoid shutting down the passengers' phones?

Sure, my answer to the question I stated would be:

Yes, but the legal entity behind the driving of the car needs to be hold legal liable. (i.e. the AI provider)

Hand over to driver can be required but the driver needs to have reasonable time to adapt to the situation (multiple seconds).

Only with this can we make sure that companies providing driving AIs give their best to make sure they don't kill.

Else they will have just pressure to improve it to a point where it happens rarely and then stop there because its not worth the development cost.

> Else they will have just pressure to improve it to a point where it happens rarely and then stop there because its not worth the development cost.

As ever, perfect is the enemy of good. Are there currently zero fatalities on the roads because cars are 100% safe? Or have they been developed to an acceptable standard of safety?

Oh its never going to be perfect (imo). And agree with you that good solution now is better than perfect never.

But we need to set initiatives in a way that manufacturers will be improving over time. They won't do it out of goodness of their hart, and quite possible marketing and good PR will be more impactfull on sale than 0.5% safer AI

I have 0 trust that market itself will solve this issue.

We could give a grace period of some years until we apply strict laws.

We just would need to make it clear from the get to go that after that grace period the law will be strict even for cars sold before ;=)

That could lead to situation where, cars that you bought, wont be useful couple of years after you bought them in case ai company doesn't deliver.

I have even less faith for politicians to not extend grace periods in face of organized backslash(probably paid and bought by companies themself), making them infinite.

> Hand over to driver can be required but the driver needs to have reasonable time to adapt to the situation (multiple seconds).

Of course. Otherwise you could just always hand over a few milliseconds before any impact and blame the driver.

This is a solved problem in actuarial & insurance circles...