2
$\begingroup$

There's a real possibility that self driving cars become more than just a high tech novelty and they start changing the market.

As seen in Logan, self driven commercial freight might be among the first. This makes sense to me. The automotive industry might be harder to overtake, but truckers aren't highly qualified. Anybody can get a CDL.

Also in the movie, we see that the trucks aren't perfect. One causes a large accident, and sets spooked horses running frantically across the highway.

Sam Harris has questions about the fundamental flaw of a technology that is expected to take lives into its own hands autonomously existing when it has to make a choice between one life or another.

If a car is hurdling towards a mass of people on the side of the road through some mistake caused by road conditions, should it veer towards a nearby ditch where it may endanger its passengers?

That's just one example of this. I've spoken with my friend about this a bit, and we've talked about Asimov's laws of robotics, which I think are a good starting point. As far as I know, there's four laws. Feel free to add if I've missed any:

  1. A robot cannot harm humanity, or through inaction allow humanity to come to harm.
  2. A robot cannot harm a human, or through inaction allow a human to come to harm, except where this would conflict with the zeroth law.
  3. A robot must obey any orders given to it by a human, except where this would conflict with the zeroth or first laws.
  4. A robot must do everything it can to protect itself, except where this would conflict with the zeroth, first, or second laws.

Under this simple system, we already have our automated cars choosing to save a group of lives over one life. Discussion about the worth of elderly lives versus children's lives I think is unimportant.

No company should be building machines that prioritize police over civilians or civilians over criminals. Nobody over 60 is getting in the Nissan they know cares more about everyone that's not in the car than it does it's passenger. The job of the machine is to make unbiased choices between saving one life or two. And it should always pick two.

But one-for-one choices are going to likely be determined by factors like what action is the easiest to perform at X miles per hour. Of course it makes the most sense that the vehicle itself is at the bottom of its own list of priorities. It's a tool that can be easily replaced.

But there's one thing I've not mentioned yet, and that's the processing speed of the on board computer. If self driving cars prove incapable of making split second decisions because they take their sweet time analyzing, we won't want to ride them.

Rather interested in philosophy and ethics, the programming part is where my knowledge is lacking. I can assume some basic facts. You'd have to make space in the vehicle for a computer, some kind of image recognition software would need to tell what's a human. Perhaps pets and whatnot could also be included.

And maybe it wouldn't be unreasonable to think since the passengers don't need to look out the windshield, that it could be replaced with a widescreen TV, or some other ancillary features.

Could we bridge the gap in knowledge here, and somebody give me an idea of whether this kind of thing is even possible?

$\endgroup$
  • $\begingroup$ You have a lot of interesting perspectives here but that makes it hard to understand what you are asking. Can you be more specific or direct in your question? $\endgroup$ – Jaden Travnik Sep 29 '17 at 18:24
  • $\begingroup$ Sure thing. New to the app, I might have to start shortening my opening statements. I'm wondering how one might design and train a computer that drives a car and can make split second decisions of ethics. I'm skeptical it's possible at all due to the speeds a vehicle might be traveling on the freeway. $\endgroup$ – RealisticWuxia Sep 29 '17 at 22:38
1
$\begingroup$

The 'Three Laws' and similar ideas are very abstract and no existing or near-term AI would be good at applying them to real-world situations. In other words, no existing or near-term AI will be able to calculate the possible harms resulting from an action in a way meaningful enough to guide its course of action in terms of those rules. To a self-driving car at the current level of sophistication, a school bus and a cement mixer are both just obstacles to be avoided-- it doesn't know that one is full of children and the other isn't. The car software will search for a sequence of braking and steering actions to avoid colliding with either obstacle, but if an accident is unavoidable it won't make any choice more sophisticated than slamming on the brakes.

To make the kind of evaluations you are suggesting, an AI would need to understand the world in a way that would make it (at least close to) an Artificial General Intelligence (AGI), with the ability to reason about the world in ways that approach the sophistication of the human mind. We don't know what kind of hardware and software is required to make an AGI, but at this point it seems to be a fairly distant goal (measured in decades of required advances). Most AI research is focused on more readily achievable goals, and handling of very abstract concepts like robot ethics is just not very relevant.

$\endgroup$
1
$\begingroup$

Asimov's Three Laws really aren't useful when talking about or trying to come up with guidelines for AI to follow. There are many reasons for this, which I'm sure have been covered endlessly on here and elsewhere, so I will only touch on a few briefly.

For example, it is nontrivial to define what 'harm' is and encode it in a way that an AI can understand. Even harder is defining what 'harm a human' means. We as human beings have a hard enough time defining what a human is ethically to begin with. Just look at the pro-life, pro-choice debate in the US.

For more on why the 3 laws don't work, check out this video

To better answer your question, I am not sure we necessarily want our driverless cars or smart devices making ethical decisions. Let's take an edited version of the infamous trolley problem as our example, your smart car is driving down the road and is about to slam into 5 people, it can do some maneuver and only kill one person.

This is nice in theory but it would involve the car trying to perform complex calculations with probabilities on whether these actions would reduce possible life lost all in a matter of seconds and having good confidence in those calculations would be extremely difficult. Not to mention, how do we value the driver's life in these calculations? Are they more valuable than the individuals outside of it? Would automakers alter their algorithms to value the driver over everything else as a selling point?

I think this stack post does a great job of pointing out and addressing the questions involved with this sort of topic.

Hope I answered your question!

$\endgroup$
  • $\begingroup$ Antlersoft and hisairnessag3, you've both shown what I had suspected from the beginning, that questions of this kind are essentially irrelevant at least at the moment. I also had considered that Asimov's stuff was a bit outdated, but I think they still work as a basic example for people not well versed in computer engineering. It makes a lot of sense that a Roomba sees obstacles and tries to avoid them in general. If the cat hops out of the way, it updates, sees no obstacles, and continues on its way. Thanks for putting up with my poorly phrased question, and understanding it and answering it. $\endgroup$ – RealisticWuxia Oct 3 '17 at 17:16
  • $\begingroup$ We would both probably appreciate it if you choose an answer that satisfies your question $\endgroup$ – hisairnessag3 Oct 4 '17 at 23:30
  • $\begingroup$ If the question is "how would AI prioritize ethically?", "They simply wouldn't. An AI of this type is not yet possible, perhaps impossible, and perhaps unwanted," is likely the shortest and most comprehensive answer, rephrased from your answers." $\endgroup$ – RealisticWuxia Oct 6 '17 at 1:14
  • $\begingroup$ Not what I meant $\endgroup$ – hisairnessag3 Oct 6 '17 at 3:42
0
$\begingroup$

The question dips into a number of issues that are only vaguely related. The primary topic is how ethics would be applied to situations in vehicle control.

Priorities in Vehicle Control

There has not been much ambiguity or historical variance in the priorities that govern situational ethics in transportation, listed here in order of descending priority.

  • Minimum risk to human life
  • Minimum risk to high value property
  • Minimum risk to pets
  • Minimum time to destination
  • Minimum consumption of fuel
  • Minimum wear on vehicle components

Some pet lovers or conservationists may slightly modify this list, but the weight of legislation and case law affirms the above order. It is the effective cultural norm expressed in rules, penalties for violations, and case law.

The degree of automation does not impact this norm. Instead, the norm is the standard of acceptance imposed on the automation.

Automation Challenges More Technical Than Ethical

The accuracy and reliability of risk detection is the primary design concern. The automaton driving or piloting the vehicle must follow rules of law regarding the operation of the vehicle, recognize common and uncommon forms of danger, and select from among available options to avoid collisions. This requires identification of stationary and moving objects, projection of trajectories, comparison of collision avoidance strategies in terms of the above priorities, and execution of the best choice. All of this must be done in real time and repeatedly or continuously.

Regarding the statement that the action taken in an unavoidable collision with multiple possible outcomes would be the, "Action [that] is the easiest to perform at X miles per hour," is incorrect. The automation would not care what is easy. The actions that are reliably achievable are the equal except for their outcome in terms of the above priorities.

Asimov's General Robotics Laws

Isaac Asimov's robotics laws do not apply to vehicle automation because the user input to the vehicle would not be, "Turn left now, and disregard those school children walking." It would be, "Take us to the grocery store on Lincoln Avenue." The above established priorities would replace Asimov's general robotics laws in the vehicle automation context.

A vehicle trajectory on an intercept course with a pedestrians will, if driven correctly, try to save lives of both passengers and pedestrians. If automated vehicles properly identify the pedestrians as pedestrians, the automation is more likely to model the physics of the vehicle paths quickly and accurately and preserve human life than a human would, given the same view and time window.

The question indicates a concern that automated systems might be slow. That is not likely. Any real time learning would be designed to run in parallel with operational control, not in series with it.

The automation would also have less reason to assign greater values to one human over another. For instance, a normally unselfish pilot or driver might, in the rush of a moment, instinctively prioritize them self over a child. The type of automation used to control an automated vehicle would assign only the value to itself that it is worth on the market to use in comparing collision loss mitigation options. It would not preserve itself over human life.

Fuzzy Implementation Practicalities

The one interesting ethical option that is not codified in law except through the school zone laws of some jurisdictions. Should decisions of minimizing loss of life be based on the number of lives or the number of projected years. For instance, if the system were to judge the life expectancy of two grown adults to be forty years total but the life expectancy of one small child to be fifty, should the value of life ethic cause the sacrificing of the two adults to save the child if no other steering or breaking options could save all three?

It is likely that this would be a matter of either legislation or case law.

Another issue is whether a 0.1% probability of loss of life is equivalent to a 100% probability of loss of some property, such as a museum or a historic landmark, or should the value of property be completely ignored when even the remotest risk to life exists. This later option might be impractical. Determining probabilistic rules regarding remote risks to humans in comparison to other priorities is an interesting area where ambiguity does exist.

That there are school zone laws implies that humans left unaccountable may place too high a priority on getting to their destinations to produce the level of child safety that the community wishes. It is likely that school zones will remain in place for automated vehicles to cover this community concern.

Ancillary Issues

Whether a vehicle spooks horses will not be a requirement imposed on vehicle automation, just as what horses might do as a human driver does not appear to have sanity or full control over the vehicle is not a cause for liability. It will be up to horse owners to keep horses away from heavy thoroughfares as is the present case. Automated vehicle will pass at safe distances unless a human's trajectory cuts off steering and breaking options.

The question's characterization of truck driving skill is dismissive. Trucker drivers are qualified drivers and the methods of qualification are in place. Not everyone can obtain commercial certification, and the certification can be lost. Statistically, they have a lower incidence of accidents and moving violations than typical drivers. Automation makes sense for trucking largely because many of the road hours are on long stretches of sparsely occupied highway.

The socio-economic challenge includes a transition away from perspectives that marketing operations have formed and solidified in the cultures of industrialized countries for a dozen generations. The primary one is that nobility, strength, sophistication, and uniqueness is expressed through a persons car, truck, boat, and plane assets. Automated vehicles will not likely change this or be hindered by it.

Regarding space for the hardware that runs the vehicle control automation, it has yet to be determined what the space and energy costs are of such computing resources after the appropriate period of research and development minimizes these requirements.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.