[thedoc (OP)]

Driverless cars could change the world- but what legal and ethical challenges do they bring with them?
Read a transcript of the interview by clicking here
or [chapter podcast=1001529 track=16.11.22/Naked_Scientists_Show_16.11.22_1005986.mp3] Listen to it now[/chapter] or [download as MP3]

[RD]

What will the driver-AI decide when faced with the dilemma of saving the car-passengers, but at the expense of the people queuing at the bus-stop ? ... http://gizmodo.com/your-self-driving-car-will-be-programmed-to-kill-you-de-1782499265

[alancalverd]

In the event of any collision, who is legally liable?

[jeffreyH]

In the event of any collision, who is legally liable?

Skynet?

[nilak]

In the event of any collision, who is legally liable?

In aviation there is a collision avoidance system called ACAS. If two aircraft are in a conficting paths  and predicted to go below some pre-established values it will issue mandatory commands to pilots at a certain time befor the point of impact. Basically ACAS is "driving" the plane through traffic  from that moment. There have been reported deficiencies in the past but the system but newer versions fixed those problems. Now, if the system generated wrong maneuvers due to design glitches and an accident would happen who is going to be responsible? I don't think, the designer will be liable. The authority that issued approval for the inmplementation and use of the equipment could be liable but only if they failed  to follow their own protocols (deliberately or in negligence). Otherwise it could be subject to a judge decision.   
         However this system is a last line of defense (usually ATC provides separation).  In the case of driverless cars it is a matter if choice.
         My idea is, if the system is proved safer (even marginally) than an average human driver (acording statistical data) it can get approval for implementation and use, even if it is completely autonomous. Nobody will be legally liable in the event of a collision. However, This is really hard to accomplish and possibly harder ( might be impossible or perhaps by implementation in small steps) to demonstrate. An average driver can cause a serious  accident once in like 10 years or perhaps even more ( I didn't check recent statistics). For example Tesla already anwers accusations using statistical data).
This is how I see the problem.

[alancalverd]

AFAIK ACAS does not assume command. It, or a human controller, can issue deconfliction advice, and a human controller can issue deconfliction commands. In the event of a collision the pilot may be held liable for ignoring such advice, or the controller may be liable for issuing an incorrect instruction, but no machine is held liable. Faulty equipment may be considered mitigating.

The technical difference is that midair conflicts are very rare and in principle predictable because the rate of closure from initially nonintersecting paths is actually quite low (in English - planes have plenty of room to maneuver and can't turn or accelerate/decelerate rapidly) whereas dogs, children and drunks run out into roads and cyclists fall off bikes without warning. Therefore the assumption that something that works in the air (and only if all the other aircraft are carrying TCAS-compatible transponders and have them switched on) can be implemented on a city street, is preposterous. 

If nobody is legally liable for a collision, who pays for the damage?

There are about 3000 road fatalities (mostly urban pedestrians and cyclists) a year in the UK, with about 45,000,000 drivers licensed. This suggests that the average driver will kill one person every 15,000 years. Multiply by 10 for serious injuries, and it's still a fairly low probability, so maybe there isn't such a great need for driverless cars after all. Driverless trains make sense because, for the most part, they operate in what aviators call a Known Traffic Environment (like Class A airspace) but city roads and country lanes don't fit into that category. 

[nilak]

Although it is called advisory ACAS, advice must be promptly and precisely folowed but you are right, the pilot might have reasons not to follow the advice if the aircraft has serious technical problems for example. In that case it comes down to the judge  to decide if the pilot action was appropriate or not. This is only an analogy for how we could be dealing with the legal issue.
    I agree, technical  challenges are far more difficult. An algorithm that coud do what a human driver does, would be far more complex than the most complex algorithm ever created by man. Even if you do it, the reliability is another big challenge.

[nilak]

The new self learning AI software have really very promising results, so it might happen sooner than expected.