Reply to: Vehicles with no human driver

[DoctorStrangelove]

I never said perfect, I just want it clear where liability lies when two competing algorithms battle it out on the high street

There'll be an exclusion from liability clause in the 15,327 page pdf governing the use of the car.

[ladymuck]

I never said perfect, I just want it clear where liability lies when two competing algorithms battle it out on the high street

[Peoplesoft bloke]

That's a good point.

Also, at present, it's pretty easy (in most cases) to determine who caused an accident and their insurance and/or the law deals with them. If an autonomous car causes an accident because it has a brain fart and mistakes a child for a paper bag, who pays? Tech companies seem to be very good at wriggling out of liability.

Everyday folk are great at

Not being insured
Driving away without stopping
Having false plates/lying about who they are
Arguing straight out lies about what happened
crash for cash

So, once again, I'm not sure why technology has to be uniquely perfect all of a sudden. Unfortunately, tulip happens.

[ladymuck]

I do have a concern about decisions that might be made on 'cost' (not just financial cost)

Say an accident is inevitable and that the vehicles can take actions to 'optimise' injury outcome.
Perhaps morally, the party causing the accident should accept greater injury than an innocent party.

But algorithms could apply lots of factors
- minimise overall financial cost e.g. let a pensioner receive more injuries since no loss of income and lower remaining years of life
- minimise injury to a party who has paid a premium for a defensive algorithm
- injure single occupant to protect a vehicle with multiple occupants
etc.

That's a good point.

Also, at present, it's pretty easy (in most cases) to determine who caused an accident and their insurance and/or the law deals with them. If an autonomous car causes an accident because it has a brain fart and mistakes a child for a paper bag, who pays? Tech companies seem to be very good at wriggling out of liability.

[Croesus]

Great on a grid system like LA I'm sure but when you come face to face with a farmer in a combine harvester on a windy, single track country lane with no passing places - I'd be interested to see the results!

[Protagoras]

I do have a concern about decisions that might be made on 'cost' (not just financial cost)

Say an accident is inevitable and that the vehicles can take actions to 'optimise' injury outcome.
Perhaps morally, the party causing the accident should accept greater injury than an innocent party.

But algorithms could apply lots of factors
- minimise overall financial cost e.g. let a pensioner receive more injuries since no loss of income and lower remaining years of life
- minimise injury to a party who has paid a premium for a defensive algorithm
- injure single occupant to protect a vehicle with multiple occupants
etc.

[malvolio]

It seems to me that the speed of assessing the situation was key - along with the river as a possible emergency landing place. Nothing in any of that is incompatible with AI as far as I can see.

I'm not arguing Capt Sully isn't a hero for what he did - just that it's becoming less and less necessary to have human intervention as machines become cleverer.

It seems a lot of people won't accept a car driven entirely by a machine unless/until it carries zero risk - which I would imagine is all but impossible. What if it carried the same statistical risk of death/injury/property damage, but enabled people who otherwise couldn't, to go places?

Given almost all traffic crashes are caused by human error (a very small percentage are down to vehicle failure, hence the official switch to RTColision from RTAccident) anything that takes human error out of the equation has to be a good move. The real fear at present is that the error-free autonomous driver will lose out to a random and typically inept human one. That risk reduces as more autonomous vehicles are on the road

[ladymuck]

It seems to me that the speed of assessing the situation was key - along with the river as a possible emergency landing place. Nothing in any of that is incompatible with AI as far as I can see.

I'm not arguing Capt Sully isn't a hero for what he did - just that it's becoming less and less necessary to have human intervention as machines become cleverer.

It seems a lot of people won't accept a car driven entirely by a machine unless/until it carries zero risk - which I would imagine is all but impossible. What if it carried the same statistical risk of death/injury/property damage, but enabled people who otherwise couldn't, to go places?

I think it's more about which decision an automated vehicle will make and when.

There's lots of examples of such cars ignoring police tape and driving into crime scenes, or failing to stop / unnecessarily stopping because it can't properly work out what's going on.

I 100% agree that such modes of transport will be a life saver for many people for whom current public transport offerings are simply not suitable. These are exactly the sort of people who deserve a good solution, not a good-enough solution.

[Peoplesoft bloke]

It seems to me that the speed of assessing the situation was key - along with the river as a possible emergency landing place. Nothing in any of that is incompatible with AI as far as I can see.

I'm not arguing Capt Sully isn't a hero for what he did - just that it's becoming less and less necessary to have human intervention as machines become cleverer.

It seems a lot of people won't accept a car driven entirely by a machine unless/until it carries zero risk - which I would imagine is all but impossible. What if it carried the same statistical risk of death/injury/property damage, but enabled people who otherwise couldn't, to go places?

[malvolio]

Why are you both insisting AI couldn't have identified the Hudson as an option?

Because AI isn't actually intelligent and would have the river marked as "Not land" on the maps it would be using to determine its course of action? All those nice long straight roads would vastly outweigh a water landing. And, of course, it wouldn't care (or even realise) that the roads would be occupied.

AI doesn't reason (at least, not yet...), it weighs up alternatives via strict rules.

[Protagoras]

As an aside, there's a prevalence for pilots who execute recoveries to be experienced, and often current glider pilots i.e. they have excellent hand flying skills.

If we take the scenario being discussed - double engine failure while ascending over built-up area.

AI / automation, would need to

Run the checklists - probably this could be done faster and more accurately than a human - which may or may not be better. (In the scenario discussed, I seem to recall that APU start was performed out of checklist sequence)

Select destination - reaching a destination after multiple failures is often about optimising energy conservation, such as trading speed for altitude. It seems to me that this is computationally possible and that seeded with alternative landing locations a system could choose a preferred option - and probably very quickly. But the question remains of whether a river would be an acceptable option when setting up such a system.

Reaching the destination - again, I think that this is possible, being an extension of autopilot and autoland capability with energy conservation management added on.

But caveats to the above are
- all sensor data is available and correct, or the system can determine what data is valid
- AI needs to be able to understand the weather picture and use that in making decisions
- there needs to be some consideration of ethical risk in decision making. Is it better to try the river and risk nn people or try the tarmac overflying mm people?

Of course if the AI could detect birds in the first place ...

[Peoplesoft bloke]

Why are you both insisting AI couldn't have identified the Hudson as an option?

[ladymuck]

You miss the point, as Sully pointed out to the enquiry. You could fly back to the airport and land provided you did all the right things immediately. It took him and the co-pilot about 20 seconds to work out what went wrong by which time it was too late. The pilots who made the theoretical landing weren't panicking, knew what the problem was beforehand and reacted immediately.

An autopilot, even an AI enhanced one, would not recognise the Hudson as a landing option. If it worked out it couldn't get home, it would most likely have picked the I25 as the best option...

Yes, I saw that testimony. The simulation pilots were told what the problem was before they started. That's information Sully and his copilot didn't have in real life.

[malvolio]

No argument from me about keep flying skills - BUT in simulator exercises, other pilots got back to the airport (or landed at Teterboro) about half the time - and that was relying on human decisions too. That landing is (rightly) hailed because no-one died, but ditching in a river is a big risk - lucky for everyone it paid off that time, but it isn't proof that an automated system couldn't have done as good, or a better job.

You miss the point, as Sully pointed out to the enquiry. You could fly back to the airport and land provided you did all the right things immediately. It took him and the co-pilot about 20 seconds to work out what went wrong by which time it was too late. The pilots who made the theoretical landing weren't panicking, knew what the problem was beforehand and reacted immediately.

An autopilot, even an AI enhanced one, would not recognise the Hudson as a landing option. If it worked out it couldn't get home, it would most likely have picked the I25 as the best option...

[Peoplesoft bloke]

Imagine trying to program a computer to predict the behaviour of BMW drivers.

Pretty easy - will ignore all speed limits, tailgate the flip out of anything/everything and ignore any rules/courtesy that doesn't suit them. A bit like people trying to barge you out of the way on the tube - expect everyone else to recognise their innate superiority and importance and defer to them at every opportunity.