Warning: I’m sure there’s a literature on this, which I haven’t read. This post is instead based on a conversation with some folks who have read more of it. So I’m “shooting from the hip” here, as they say.
Like planes, boats, submarines, and other vehicles, self-driving cars can be used in several modes. The automation can be turned off. It can be turned on and advisory only. It can be driving, but with the human watching carefully and ready to take over at any time. Or it can be driving with the human not watching very carefully, so that the human would take a substantial delay before being able to take over. Or the human might not be capable of taking over at all; perhaps a remote driver would stand ready to take over via teleoperation.
While we might mostly trust vehicle owners or passengers to decide when to use which modes, existing practice suggest we won’t entirely trust them. Today, after a traffic accident, we let some parties sue others for damages. This can improves driver incentives to drive well. But we don’t trust this to fully correct incentives. So in addition, we regulate traffic. We don’t just suggest that you stop at a red light, keep in one lane, or stay below a speed limit. We require these things, and penalize detected violations. Similarly, we’ll probably want to regulate the choice of self-driving mode.
Consider a standard three-color traffic light. When the light is red, you are not allowed to go. When it is green you are allowed, but not required, to go; sometimes it is not safe to go even when a light is green. When the light is yellow, you are supposed to pay extra attention to a red light coming soon. We could similarly use a three color system as the basis of a three-mode system of regulating self-driving cars.
Imagine that inside each car is a very visible light, which regulators can set to be green, yellow or red. When your light is red you must drive your car yourself, even if you get advice from automation. When the light is yellow you can let the automation take over if you want, but you must watch carefully, ready to take over. When the light is green, you can usually ignore driving, such as by reading or sleeping, though you may watch or drive if you want.
(We might want a standard way to alert drivers when their color changed away from green. Of course we could imagine adding more colors, to distinguish more levels of attention and control. But a three level system seems a reasonable place to start.)
Under this system, the key regulatory choice is the choice of color. This choice could in principle be set different for each car at each moment. But early on the color would probably be set the same for all cars and drivers of a type, in a particular geographic area at a particular time. The color might come from in part a broadcasted signal, with the light perhaps defaulting to red if it can’t get a signal.
One can imagine a very bureaucratic system to set the color, with regulators sitting in a big room filled with monitors, like NASA mission control. That would probably be too conservative and fail to take local circumstances enough into account. Or one might imagine empowering fancy statistical or machine learning algorithms to make the choice. But most any algorithm would make a lot of mistakes, and the choice of algorithm might be politicized, leading to a poor choice.
Let me suggest using prediction markets for this choice. Regulators would have to choose a large set of situation buckets, such that the color must be the same for all situations in the same bucket. Then for each bucket we’d have three markets, estimating the accident rate conditional on a particular color. Assuming that drivers gain some direct benefit from paying less attention to driving, we’d set the color to green unless the expected difference between the green and yellow accident rate became high enough. Similarly for the choice between red and yellow.
Work on combinatorial prediction markets suggests that it is feasible to have billions or more such buckets at a time. We might use audit lotteries and only actually estimate accident rates for some small fraction of these buckets, using bets conditional on such auditing. But even with a much smaller number of buckets, our experience with prediction markets suggests that such a system would work better than either a bureaucratic or statistical system with a similar number of buckets.
Added 1p: My assumptions were influenced by the book Our Robots, Ourselves on the history of automation.