Thomas Kaminsky: Blind Trolley: Effective Altruism in a World of Incomplete Information
“Oh no, it’s happening again!” You mutter to yourself, rolling your eyes. You can hardly believe it: for what seems like the hundredth time, you find yourself in a trolley car, racing down the tracks toward five unsuspecting victims. Once again, there’s a lever right in front of you that can switch the path of the train, allowing you to run over a single person instead of the five you’re currently barreling towards.
For a moment you hesitate, feeling the same trepidation that you felt the first time you were presented with this scenario, but then you take a deep breath and do some basic math:
Alive is Good, and Dead is Bad. Hence, A > D.
Since A > D, then
4A > 4D
4A + 1D > 5D
5A + 1D > 1A + 5D. ∎
So you reach a conclusion: you should switch tracks and run over the one person. Smiling at your efficient proof, you pull the lever.
It isn’t until the next morning when, while eating breakfast, you read the newspaper headline:
Nobel Prize-Winning Dog-Lover and Mother of Three On Her Way to Donate $500,000,000 Diamond to Charity Killed by Trolley After Being Robbed by Five Serial Killers Who Listen to Nickelback, Allowing Them to Get Away!
You stop eating your toast. This is not what you wanted. This is not what you wanted at all.
*****
The above thought experiment, silly as it is, demonstrates an important limitation of utilitarianism: whenever you find the answer to a moral question through some calculation, your answer is only as strong as the information you used to get it. In this situation, if it was that case that all six lives had some identical, constant value, then the right decision would have been to run over the one and spare the five. What we failed to realize was that the problem instead actually looked something like this:
where each integral is just the net utility that each set of people would contribute from the time t0 (the moment the trolley would hit them) to the time d when the last of them would naturally die. This expression is functionally unsolvable––each value function could be (and most likely is) radically different from the others, and, even if we knew enough about each person to meaningfully compare them, there’s no way to definitively calculate a solution because each term is only concerned with future utility. We’re at an impasse, then––steaming down that track, there isn’t enough information in the world to make an informed decision.
That’s a super bad situation, for sure, but it still might not be obvious why this matters for everyday altruism. People don’t really deal with these moral dilemmas often, and, if they did, most people would still probably feel like the simplification that we made in the thought experiment is the best way to approach the problem. Still, falling into this pattern could lead to very negative outcomes.
For example, suppose there was some systemic factor that impacted who was likely to be at each track––say, that the serial killers’ track was right next to a black market, and the Nobel prize winner’s was right next to a youth center. Then, if trolleys kept on reaching this intersection and their conductors always chose where to go based on the logic in the thought experiment, they’d disproportionately kill children and their parents over black market clients. In that case, accepting the simplistic metric could cause more than a single bad outcome.
Of course, there’s no way to eliminate risk entirely. For any opportunity, there’s always some probability of failure. Plus, there are compelling models (expected value, etc.) which imply that investing in riskier projects with greater potential payoffs might still be worthwhile. The problem with the above example is that you have no idea how much risk there is. There are so many factors at play that you can’t even begin to make an expected value argument for the solution.
The trolley problem is a perfect example of a situation to avoid. You know absolutely nothing about the people who you’re interacting with except for how many there are, and you have no way to gain any more information before you hit them. As such, the best-case scenario is that you have a 50% chance of making the right decision. There’s nothing ‘effective’ about this altruism––if you have no meaningful way to judge the risk of your endeavor, you might as well take the left fork because you’re left-handed. The question, then, becomes this: what steps can be taken to reduce the uncertainty of your uncertainty? How do you ensure that your risks are meaningful?
To begin to address this, I propose three alterations to the typical values:
Prioritize problems that are close to you.
Now, this flies in the face of a lot of Effective-Altruist principles. It’s vital to EA that a human life on the other side of the world is just as valuable as one in your backyard, and, if your backyard happens to be in one of the most prosperous countries in the world, you might improve more lives by investing in something far away from you than focusing on your comparatively well-off community. These are definitely worthwhile considerations, but it’s still important to recognize the advantage of focusing on nearby problems. For one, you probably have far more information about your neighbors than people on the other side of the world––both enough common culture to understand what people want from you and enough shared experience to understand what resources are needed to solve their problems. If it starts to rain, you’ll be the first to know they’ll need raincoats. By keeping your focus on a realm that you understand, you can better characterize the risks that you’re taking, allowing you to more confidently wield the utilitarian metrics that underlie EA.
Prioritize concrete solutions over institutional ones.
There’s a lot of debate over whether it’s more worthwhile to invest time and energy into direct aid (e.g. providing mosquito nets to children in regions where malaria is common) or institutional growth (e.g. investing in programs that improve access to education in areas with few academic opportunities). Don’t get me wrong; better information certainly isn’t enough to justify doing away with institutional change altogether––in fact, there are probably situations in which you can have more certainty about the value of implementing institutional solutions over direct ones––but it does make clear one of the enormous benefits of directness. Even the most well-researched institutional cause areas are justified by research gathered over a fixed time period, and, often, it’s impossible to analyze the truly long-term, intergenerational impact of such actions. Concrete solutions are way simpler––give children mosquito nets, and they are less likely to get malaria. Full stop. You’ve accomplished your goal.
Don’t limit yourself to the ‘optimal’ solution.
There’s no way to completely understand the implications of potential actions; the world is just too chaotic, and you don’t have enough time to consider every contingency when trying to do good. As a result, there’s always some probability that, even once your algorithm has spit out the EA-optimized course of action, it might not actually be the best way to solve your problem. It could still be worthwhile, then, to invest a portion of your time, energy, or money into ‘suboptimal’ solutions. For one, you probably don’t know that they’re suboptimal, at least in the beginning, and, even if they are, you can potentially learn more about the system you’re trying to impact by seeing how it’s affected by a variety of forms of aid. Sure, if your original solution continues to be the highest performer you can begin to focus on it fully, but, if some other contender turns out to have merits that you hadn’t previously noticed, you could instead shift in that direction. Of course you’ll have to weigh the material sacrifice of investing in multiple solutions, but, especially in situations where the landscape of the problem isn’t well-known, it’s important to acknowledge the fundamental limitations of your calculation.
This certainly isn’t a complete list, and it may not make sense to apply these values to every problem, but I think that they’re still worthwhile factors to consider. One of the great strengths of Effective Altruism is its emphasis on reason, but the value of reason is entirely contingent upon the completeness of the data that you use to make your decisions. If a problem’s givens aren’t subject to the same level of rigor as the stuff done with them, we lose the incredible power of EA, and that would be an incredible shame.
———
Thomas Kaminsky is a freshman from Walla Walla, Washington who plans to study Computer Science and Physics at Harvard. He loves thinking about extinction-level events, education and Truth, and the inevitable AI takeover. Catch him walking his dog, a pudelpointer named Mollie, or trying to convince his friends to listen to the Barber Violin Concerto.