[klmadfejno]

Self driving is, in my opinion, ironically, a trolley problem. If enabled in mainstream usage, we'd probably see total deaths drop, but they'd be shifted to a more random set of people dying for stupider reasons.

Would you rather two drunk drivers and an overworked sleepy trucker die; or one guy who could drive safely but was instead driven into an overturned truck.

[sandworm101]

We can cross that bridge when, if, we come to it. At the moment, AI products are nowhere near as safe as human drivers in no-compromises realworld driving situations (ie a foggy/snowy/icy night, driver asleep, mountain highway etc).

If we really cared about using AI to promote safe driving, we would turn it away from the road and towards the human driver. Take the above scenarios. While AI driving is a difficult problem, an AI that can detect a drunk/tired/inept driver is not. Any car could easily be equipped with internal cameras or other systems to tell if a driver is unfit. Any car cold be equipped with a dead man's switch to safely deal with a driver that has fallen asleep. How about a car that calls the cops whenever it thinks its driver may be drunk? Heck, you don't need AI to install a speed governor that would curtail any and all speeding[1]. The fact that the market repeatedly rejects such simple AI implementations tells me that all-up AI driving is a long long way away.

[1] My dream is an automatic switch that turns on a police car's lights/sirens every time they break the speed limit. Why else would a marked cop car ever speed unless it was chasing someone?

[klmadfejno]

> We can cross that bridge when, if, we come to it. At the moment, AI products are nowhere near as safe as human drivers in no-compromises realworld driving situations (ie a foggy/snowy/icy night, driver asleep, mountain highway etc).

I'm not entirely convinced this is true, even though it's commonly stated. People are terrible at driving in fog and snow and ice (and... asleep?). It's intuitive why a self driving car company would not want to release the cars to do this in extra dangerous situations while they're still improving the easier stuff, but we don't exactly have stats to say the cars would necessarily do worse.

[sandworm101]

>> People are terrible at driving in fog and snow and ice

They are considerably better than the AI systems who currently just stop and say "I see no road" or facetiously stick to markings/signs that mean little during winter conditions. The last time I rented an SUV (2020 Jeep grand Cherokee) it wouldn't let me reverse into a parking spot because the rear camera/sensor was caked in snow/ice. In order to be dangerous at winter driving you must first be able to actually move.

[klmadfejno]

There's a line of can't/won't/shouldn't here though. The AI systems probably could drive in these conditions. Most of them utilize GPS to navigate. But they don't because that's a high risk activity. I'm not totally convinced that even now, the average accident rate of a self driving car wouldn't be lower in adverse conditions compared to a random driver- if the car were just told to do its best and let loose.

Not that I'm advocating this is a good idea.

[MattRix]

To be clear, human driving that is augmented by “AI” (lane following, collision detection, blind spot warnings, etc) is much safer than human driving alone.

I

[sandworm101]

To a point. There is an inflection when the AI becomes a crutch that allows the human to stop paying attention. Tesla may be at/near this point. For instance: lane assist is great until it causes people to take their hands off the wheel, to stop looking at where they are going. I'd rather see such systems implanted as enforcement mechanisms. Let the AI prevent a car from drifting out of its lane. Let it monitor the lane position and scream at the driver when they start to drift. Don't let the AI become a comfortable crutch that allows the driver to take their mind off the task.

[MattRix]

I agree, driver monitoring should absolutely be a core part of self driving systems. This is something that George Hotz (Comma.ai) is getting right, and Tesla is getting wrong.

The problem is that Elon/Tesla has the hubris to think that their self driving is soon going to be so good they don’t need to worry about driver monitoring, but that’s obviously wrong at this point.

[ghaff]

I'm not sure it's a trolley problem so much as an automated system that is statistically safer than manual operation but will "randomly" kill people on a regular basis. Which is a legislative/liability problem because outside of possibly rare drug side effects, we don't normally accept consumer-facing products, even if used and maintained properly, randomly killing people--and just shrug our shoulders because stuff happens.

[bumby]

I think the OP was referencing the trolley problem because autonomous driving AI is often regarded as a utilitarian problem. In other words, "intent" does not matter, only consequences. In that regard, any ML cost function is purely focused on consequences. In the context of minimizing deaths it wouldn't matter if the software deliberately chose to kill one person, if the cost (presumably total deaths) was minimized. Where I think it gets sticky is that our society does factor in intent (see the various degrees of homicide) and if we're thinking in terms of pure utilitarianism, I'm not sure how this plays into the liability of autonomous software.

[MattRix]

This is a trolley problem as much as seatbelts and airbags are.

The number of people who get into car accidents is staggering, and it’s certainly not just drunk and tired drivers.

Drive assist features like lane keeping already make driving much more safe. There’s no trolley problem here, just huge numbers of reduced deaths across the board.

[bumby]

>This is a trolley problem as much as seatbelts and airbags are.

I disagree. The trolley problem is concerned with ethical decisions. Seatbelts are passive safety devices and don't make decisions. Airbags make "decisions" based on a sensor input, but my hunch is there is very little ethical dilemma in whether or not to deploy an airbag when a sensor threshold is met.

Current driving assist is probably not much of a trolley problem but future self-driving software almost certainly will be tasked with making choices between "bad and worse" which opens up the ethical can of worms about how to define "bad".

[MattRix]

Yes, those (very rare, mostly hypothetical) situations are trolley problems, but OP was making an entirely different argument, referring to self driving as a whole being a form of trolley problem.

I think the concerns of self driving “moral dilemma” problems are largely overblown. First, they are such rare situations: I’ve been driving for 20 years and never experienced one, I imagine few people have. Second, the solution is simple anyway: favor the passengers of the current car.

[ghaff]

I almost have trouble imagining realistic scenarios that involve deliberately endangering the occupants of the vehicle in order to protect others. (And any manufacturer who designed such a system is probably not going to find many takers.) And, for that matter, there are very few scenarios where swerving and running into something at speed is a better decision than braking hard.

[Digit-Al]

I'm the same, I can't really come up with any realistic vehicle accident scenario that would be a classic trolley problem. I may just be lacking in imagination though. I'd be interested if anyone could present a realistic example of such a scenario, and one that isn't wildly unlikely ever to happen.

Something that does occur to me is that since humans tend to just automatically react in extreme situations whereas an AI would presumably have precious milliseconds to consider different scenarios, maybe an AI could create a trolley problem where it wouldn't exist for a human.

For example, if someone steps out from behind a bus a human will most likely slam the brakes on instantly and will likely slide into the pedestrian at a slower but possibly still fatal pace; whereas an AI having a bit of time to think calmly about it might slam the car hard into the bus whilst also braking, which might slow the car down enough to prevent too much injury to the pedestrian whilst causing damage to the bus, writing off the car, and possibly injuring the occupant (though the car safety features should help).

[bumby]

>possibly injuring the occupant (though the car safety features should help)

This seems to take the opposite viewpoint of the OP which said the "obvious" answer is to do whatever is best for the passengers.

>one that isn't wildly unlikely ever to happen.

A lot of the comments here are along these lines and, to a certain extent, they miss the point. Risk is a combination of probability and severity even before considering ethics. In the cases where probability can be reliably calculated, there needs to be threshold about what risks are accepted in order to implement informed decisions. I'll try to illustrate a couple examples.

Say there is software that decides to take a specific action based on sensor input. Maybe the action is to accelerate and swerve to the left if an obstacle is detected to be oncoming from on the right at an intersection. Let's make it somewhat more complicated by having a jaywalker stepping off the curb on the left. There are (at least) two prospective events:

1) The car does not perform evasive maneuvers and increases the risk of being hit by the other vehicle 2) The car does perform evasive maneuvers and increases the risk of hitting the pedestrian

Each has a probability and a severity. To make things simple, say the probability of each is 5%. However, they have different severities. In scenario 1) the severity to the pedestrian may be risk of injury or death even if the severity to the driver is fairly low (probably repair work covered by insurance). But in scenario 2) the severity to the pedestrian is negligible while the severity to the driver(s) is moderate. With modern cars, it's unlikely either driver will be killed is fairly small but there is an increased risk of sub-fatal injuries and likely much more damage to the vehicles.

Remember, the risk = probability x severity. So if we take the simplistic approach of only taking in the perspective of the driver ("do whatever favors the passengers of the car") we have effectively minimized the passengers risk of moderate injury by increasing the pedestrian risk of grievance injury. Adding ethics gets even more complicated; does it change if the pedestrian is a child? A mother pushing a stroller? A homeless person?

Another example would be what if the car is driving by a school playground and notices a ball roll between parked cars and into the road? A human can intuit playground + ball = a higher probability a child may run into the street after the ball. This might be enough to cause a human driver to apply the brakes hard even if a child isn't immediately visible. Would software intuit the same? Even if it did, would the car do the same if it increased the risk of a collision from behind?

The safety-critical software world spend a lot of time on small-probability events. The more recent 737-Max issue has similar through-lines, though there are many confounding issues as is typical with safety mishaps. The MCAS was classified as "hazardous" severity and by company policy required two AOA sensors to reduce the probability of failure. This would effectively reduce the risk since risk = probability x severity. However, since the severity was "hazardous" and not "catastrophic" (i.e., they didn't think it could cause a plane to crash), the extra sensor was an optional software feature. Had the severity been appropriately attributed, I think there is a greater likelihood that the second AOA sensor would be a mandatory feature because they (I assume) have a specific risk threshold they are aiming for. All this to say, it's not enough just to say "These are small probability events so they don't have to be addressed" but rather probability is just a part of an overall risk strategy.

The decision space with safety critical software is extremely large precisely because you must mitigate low probability events if the severity is large enough. I sometimes wonder if the fact that self-driving technology targets are missed is attributable to naivety of oversimplifying the problem.

[bumby]

Managing those edge cases effectively is extremely important in safety critical software. Addressing these is often what separates quality critical software from low-quality. Would you want an aircraft software engineer or nuclear power plant engineer to disregard low probability events? It worries me when I get the impression the SV mindset of glorifying “moving fast and breaking things” infiltrating safety critical applications, particularly those that impact the general public.

>Second, the solution is simple anyway: favor the passengers of the current car.

I don’t think this is a given. Would the ethical software, for instance, drive a car through a crowd in order to save a lone passenger? Would you be okay with a human driver being absolved of responsibility for the same choice? My intuition is most would not, because we recognize there is an obligation to others as part of the social contract. It seems a naive oversimplification to not extend the same obligations to software that makes the choices in our place.

[MattRix]

How is the car saving the passenger by driving through the crowd? That seems like an entirely imaginary scenario. In almost any real world scenario the solution is going to be to apply the brakes and come to a stop.

[bumby]

It is an imaginary scenario. But so is the scenario where they can always apply the brakes and come to a stop.

What if there is a cement truck tailgaiting and they can't hard brake? What if there are jersey barriers constraining the maneuverability? Either of those would risk the passenger more than running into pedestrians.

Performing failure-mode-effects-analysis on software is largely an exercise in creating imaginary scenarios and then effectively mitigating the risks to within an acceptable level. I agree, in most real-world scenarios the tough problems can be mitigated largely by "slow down and re-assess". But that's not what safety-critical software risk mitigation is always about. You have to mitigate low-probability events as well, especially in a domain with governmental regulation. According to the NTSB report, the Uber accident that killed a pedestrian applied the "slow down and re-assess" scenario that attributed to the death. The software was misclassifying enough that an artificial delay was put in to avoid (I assume) nuisance braking. So in the real-world, they didn't want to always apply the "just brake if in trouble" strategy because of trade-offs to the passengers and it didn't work out well.

[omgJustTest]

Well given that the genetic lottery works differently for different gene pools… I’d say that two drunk drivers or a sleepy trucker are already pretty random.

Most success is arbitrary, even genetic success, and we should have compassion replace self-righteousness. A lot of people who are failures or successes don’t “deserve” their fates.

Self driving is obviously a vain per suit of personal space for those who can afford it. Automated public transport, or even slightly improved public transport is a much better goal for the us.

[alkonaut]

Exactly this. I'd go so far as to also include honest mistakes. I'd rather have 10x the risk of being killed by a human making a mistake, than by a computer with a bug. The human making the mistake has a skin in the game (they are physically involved in the accident). Accidents happen, but I know humans try to avoid them.

Being killed by a piece of obsolete AI-code done by the lowest bidder for a company that went bankrupt 10 years ago seems much worse.

So not only will AI drivers need to be as good as human drivers, they need to be orders of magnitude better before we'll accept them.

[AnimalMuppet]

I think I'm the opposite. I'd take the 10x reduction in the probability of dying. I'd take it for myself (I think I'm a fairly good driver, but I know I'm not as safe as I should be). And I'd take it for the 90% of deaths it would save, even if they aren't me. Those other lives matter, too.