Problem:
Let’s say you see a trolley headed towards a track with 5 people tied on the tracks. You have the ability to switch the tracks to instead go towards a track with only 1 person tied to it.
Do you swap the tracks?
Now for the twist. Lets say you know that the 5 people all faced the same question above and all 5 chose to swap the track, killing only one person. The one guy also faced the same question and chose to NOT swap the tracks allowing 5 to die.
Variants of the trolley problem are found with writing software for autonomous driving. How do we program the trolley to make the best decision with the limited information it has? Let’s say our trolley knows only the information above. In this exploration we are not going to worry about how it knows that 1 person killed 5 people or the other 5 killed 1 person. Nor is this a discussion about whether or not a trolley should be programmed to make such decisions. This is only a thought experiment of different ways that we can program a computer to come to the best decision with the limited knowledge it has.
Finding a Solution:
How can our decision process be measured? A decision will affect the universe in both direct & indirect ways. There can be many possible ways to come to a decision. When a human comes to a decision they may prioritize life over everything else. Another might pick the option that results in the least amount of deaths. Identifying all the possible desired outcomes or strategies to coming to a decision would be a good start. Depending on our situation our computer may prefer one strategy over the other or even could pick a strategy at random when there are more than one valid strategies to pick from. Having explicitly defined our desired outcome we can evaluate all the choices and determine which choice gets us closest to our desired outcome.
This is a method in which we can evaluate AND measure for an optimal solution. If we were given more information as to their backgrounds and experiences any moral or religious reason(s) for making a decision would be difficult to measure. Being difficult to measure, it would become difficult to use in finding a solution. So we will try to avoid anything in our thought process that will be difficult to measure.
Lets recognize all possible desired outcomes, establish any assumptions, and consider everything we know before finding the optimal decision. After all the possible inputs have been defined our solution is the choice where the desired outcome is maximized.
Given Information (INPUTS)
- Each person tied to the track faced a SIMILAR question but it wasn’t the exact same experience. Because we aren’t told the people on the tracks were given the “twist” part. So we can assume they made their decision without any background information that we have. We know some history about the people tied to the tracks while they did not.
- We cannot change the past. Our choice will not change what has already happened. So to remain completely neutral, the “twist” part should not affect our decision. However we as humans are incapable of not using past information to influence our decisions. Recognizing some one’s history should help us make a more educated decision but that does not mean it will always work out to be a better judgement. Individual experiences and past will influence that judgement bringing existing bias into play. A computer not having individual experiences, can still use past information to come to a “better” decision.
- Only 2 choices.
- Do NOTHING (conscious decision or not, it is the same choice)
- Switch the Tracks
Possible Desired Outcomes (Strategies)
- Minimize the change in the universe. In this scenario we want the choice that results in the smallest change or delta (Δ) of the universe. Creating life or destroying life would both impact our Δ of the universe in different amounts. The weight of change is dependent upon additional information we have. Removing 2 lives is double removing 1 life unless we know that the one life is a serial killer. Removing that one serial killer would have a stronger weight in our Δ assuming he would continue on killing. Also direct vs indirect Δ would also have different weight. A direct life saved would be measured stronger than a life saved indirectly. In the direct life saved scenario we are certain of the outcome as it is immediate. An indirect life saved is not immediate and may or may not happen due to unknown possibilities. Given we have no other information that we can use as a measurement we will use the following
Important Notes About Weight:
– Direct Life Saved =Direct Life Removed
– Indirect Life Saved = Indirect Life Removed
– 2 Indirect Lives Saved = 1 Direct Life Saved
– 2 Indirect Lives Removed = 1 Direct Life RemovedThose last two notes are subjective. One could argue and with good reason that the potential to save million’s of lives is greater than saving 1 life. However, the number of Indirect Lives are 5 in both choices. So the weight we pick (in this case .5) is irrelevant to our problem as long as:
Indirect Life Saved < Direct Life Saved
Indirect Life Removed < Direct Life RemovedAnother assumption to point out is we are looking for the smallest delta where -1 = +1 in terms of change. In this scenario we are NOT concerned with whether adding life is right or wrong. As we pointed out earlier that life could very well live on to kill many other living things. We are only concerned with minimizing the change in the universe. In other words a Δ of -1 would be preferred than a Δ of +2 even though that choice would result in a death.
- Maximize Life. In this scenario we want the choice that results in creating or maintaining existing life. The number of deaths has no place in our decision here. A Δ of +5 would be preferred than a Δ of +2.
- Minimize Death. In this scenario we want the choice that results in smallest number of deaths. The number of lives added or maintained has no place in our decision here. A Δ of -1 would be preferred than a Δ of -2.
It is important to understand the differences between “Minimize Death” & “Maximize Life”, & “Minimize Change”. Let’s look at an example. In the following example when using minimizing change as our desired outcome either choice becomes valid as they both are equal. With maximizing life we choose the first decision and minimizing death we choose the latter.
There may be other strategies that I have not thought about. Without any other information about the people behind the lives, this seems to be a good ruler to measure the outcomes of our decisions.
Bias or Not?
We have to discuss bias. Do we use the history we are given to influence our decision? After all, the guy on the track by himself chose to kill 5 people over 1. As to why, we have no idea. This is important to acknowledge that we may not always have access to the right information. He may have had a reason to kill all 5 that we would even agree with. But we rarely have the why or all of the information which is what makes it a BIAS!
For our thought experiment, we are going to treat the deaths caused by the people on the track as potential lives saved if we chose to let them die on the track. This introduces a sort of Karma into our decision process. To cause the death of someone that has caused 5 deaths vs someone that has caused 1 death. Anything said about the future can never be said for certain so bias can only be based on past events. This is really the only bias we have at our disposal to use. So we will use it.
Next let’s find optimal solutions per desired outcome with the bias toggled on or off. First choice would be to do NOTHING. Second choice to switch the tracks.
Solution by Desired Outcome (NO Bias)
- Minimize Change (Δ)
Either choice is equivalent - Maximize Life
Switch Tracks is optimal - Minimize Death
Switch Tracks is optimal
Solution by Desired Outcome (Bias)
- Minimize Change (Δ)
Either choice is equivalent. - Maximize Life
Switch Tracks is optimal - Minimize Death
Switch Tracks is optimal
Solution Explained & Lessons Learned for AI
In all scenarios switching tracks is optimal. Whatever strategy we use for our desired outcome we come to the same optimal solution to switch the tracks. This is also true when we introduce our bias.
In this scenario, the best choice for the computer to make is apparent. There may be other desired outcomes that we have not thought of but with every strategy we have defined (“Minimize Change”, “Maximize Life”, “Minimize Death”) the optimal choice is to switch the tracks.
How do we trust companies to program ethical decisions? Can we trust the entity programming the rules for a self driving car to make such decisions? No of course not. Theoretically the blockchain solves our trust problem. Blockchain can guarantee the computer is making it’s judgement based on the exact strategies defined and non of the inputs are being manipulated. Such parameters could be defined inside a smart contract and executed on the blockchain with its inputs, rules, & results published for all to see.
This would be only useful in the top right quadrant of our pie chart below. These are the problems, like the trolley problem example, where the optimal outcome is the same choice regardless of the strategy or bias toggle being used. We still would need to explore what to do where the choice varies based on strategy, bias, or both.
How do we handle these situations? How should we prioritize strategies? Should bias be toggled on? As we get more information it becomes more ethically complicated in how the different elements should be weighted. In a car, do we give harm to self more weight than harm to others? It can get ugly very quickly and as you could imagine if not implemented correctly could lead to a very real dystopian future. Just because we can do something doesn’t always mean we should. Some might argue there should be laws or regulations put into place where robots and self driving cars are restricted from being programmed to make such ethical decisions. In that scenario the trolley would only continue on its set path regardless of the outcomes effectively killing all five people.