Sunday, February 15, 2015

To push or not to push? Should that be the question?

There is a runaway trolley bus that is headed for five innocent people who cannot escape. A nearby pedestrian notices what is going on. The pedestrian has five options (and no others):
Option 1: Do nothing and let five die.

Option 2: Flick a switch that redirects the trolley onto a side-track with four innocent people who cannot escape on it, killing them and saving the five

Option 3: Flick a switch that redirects the trolley onto a raised side-track with three innocent large men who cannot escape on it, killing them and saving the five.

Option 4: Push an innocent very large man, who is holding another innocent large man and standing on a footbridge, into the path of the trolley, killing them and saving the five.
Option 5: Push an innocent very large man on a foot bridge into the path of the trolley, killing him and saving the five.
The pedestrian takes Option 5, and as a result the innocent very large man dies and the five innocent people are saved.

Select which answer best fits your opinion (chose one and only one response):

A: Given the situation, the pedestrian’s action was morally permissible 
B: Given the situation, the pedestrian’s action was not morally permissible

According to many philosophers, including Philippa Foot (1967) and Judith Jarvis Thomson (1976), the vast majority of people realize that pushing the very large man in this kind of situation is not morally permissible (i.e. morally speaking, it’s not OK to push the guy). Empirical research on trolley scenarios like this are largely in agreement; the proportion of respondents reporting that pushing the very large man (killing one to save five) is morally permissible ranges from about 11% (Hauser et al. 2007) to 18% (Kelman and Kreps 2014). BUT, these empirical studies, and the views of Foot and Thomson, are based on versions of the scenario that at least imply that only two options are available to the agent: to push or not to push. For an example of a two-option version, consider the above five-option scenario with options 2, 3, and 4 omitted.

Among the uses of experimental philosophy (doing experiments to help resolve philosophical problems), are:
1) Testing the claims philosophers make about the proportion of people who have this or that intuition, or believe this or that proposition, etc. And,

2) Comparing intuitions about similar scenarios to test whether certain manipulations tend to produce different results (e.g., to test whether a thought experiment inadvertently elicits a bias by the inclusion of some purportedly irrelevant fact)
I conducted a paper-based survey of 468 undergraduate students so as to run these kinds of tests in relation to the ‘pushing the very large man’ (AKA ‘foot bridge’) trolley scenario. More specifically, I wanted to find out whether including several, instead of two, options had an effect on how people reported the moral status of pushing the very large man in a ‘foot bridge’ trolley scenario. The basic method for conducting such an investigation is to randomly assign each individual in the sample either the two-option version of the scenario or a several-option version of the same scenario. In my study, I randomly assigned one of the following three versions: the two-option version, the three-option version, or the five-option version. 

The five-option scenario above is the same as I used in my study. The three-option version followed the same pattern, but omitted options 2 and 4 from the five-option version. Similarly, the two-option version omitted options 2, 3, and 4.

As you can see in the chart below, the more options in the scenario, the greater the proportion of respondents indicating that pushing the very large man is morally permissible. The differences between the two-option version and the other versions are both statistically significant (p<0.05). The difference between the three- and five-option versions was not statistically significant (p=0.1213). These results mean that we can be 95% or more confident that the differences between two-option version and the other versions are not the product of chance. Notice also, that the majority verdict changes position as we move from the two- to the five-option version. In the two-option version, the majority views pushing the very large man as morally impermissible, but in the five-option case an even greater majority views pushing the very large man as morally permissible. This is interesting result #1.

Why might including several-options in the scenario change the reported judgments in this way? Peter Unger (1996) has argued for ‘Liberationism’, i.e., that including intervening options in established two-option thought experiments can liberate readers’ judgments by inducing them to think more deeply about which option reduces the most harm to innocent people. According to Liberationism, then, reasonable respondents should report more utilitarian judgments about the morality of particular actions in several-option setups compared to two-option setups.

What do you think could have caused this result? Also, assuming the cause of the difference is not pernicious, think about the implications of this finding for the trolley problem and, potentially, all those other famous two-option scenarios!

Interesting result #2 is that 41.8% of respondents to the two-option scenario reported pushing the very large man as morally permissible. This is more than double the proportion found in other published studies of ‘foot bridge’ trolley scenarios, and far more than anecdotal evidence from the ‘coal face’ of undergraduate (raise-your-hand-if-you…) lecturing would suggest. Was my result anomalous? My collaborators conducted some online surveys using very similar scenarios with very similar results. When I observed Genevieve Wallace polling students on this scenario using clickers, the results were also similar to my results here. So, it seems this is a genuine mystery. Why are all the results I have first-hand experience of different to all of the results that I don’t have that experience of?

Please help me solve these mysteries!

Dan Weijers
Department of Philosophy
Sacramento State


Many thanks to Peter Unger and Justin Sytsma for advice and comments on the study set up, analysis, and discussion. Thanks also to the Provost’s Research Incentive Fund for effectively bankrolling the study.


Foot, P. (I967). The Problem of Abortion and the Doctrine of the Double Effect, The Oxford Review, 5: 5-15.

Hauser, M., Cushman, F., Young, L., Kang‐Xing Jin, R., & Mikhail, J. (2007). A Dissociation Between Moral Judgments and Justifications. Mind & language, 22(1): 1-21.

Kelman, M., & Kreps, T. A. (2014). Playing with Trolleys: Intuitions About the Permissibility of Aggregation. Journal of Empirical Legal Studies, 11(2): 197-226.

Thomson, J. J. (1976). Killing, Letting Die, and the Trolley Problem. The Monist, 59: 204-217.

Thomson, J. J. (1985). The Trolley Problem. Yale Law Journal, 94: 1395-1415.

Unger, P. (1996). Living High and Letting Die: Our Illusion of Innocence. Oxford University Press: New York.


  1. Dan, do you think that something like a decoy effect could be occurring in your studies?

    (For those who are not familiar with it: The decoy effect works by introducing a dominated option. So, for example, rather than being given a simple choice between a small regular hamburger for 5 dollars and a fancy large hamburger for 15 dollars, you are given a third option, which is a fancy small hamburger for 15 dollars. We say this latter option is dominated by the large fancy, because the large fancy burger is clearly a better value for anyone who values fancy. What's interesting is that by introducing a decoy of this kind a greater percentage of people will choose the large fancy than in the previous two option example. In the decoy effect, it seems that people are simply attracted to one of the options as a result of ruling out the option it dominates.)

    In your 5 option study it seems to me that option 5 clearly dominates option 4. So I'm wondering if it triggers decoy thought process (an alternative to the deep thinking Unger suggests.)

  2. Very interesting suggestion Randy. There is one empirical paper that tests for this kind of effect (and finds it) in trolley problems. They tested 3-option scenarios and got respondents to rate all three (different to our method). They positioned their extra option so that it was somewhere in between the two extreme options (in terms of net lives saved and (roughly) amount of physical contact), but much closer to one option than the other. I say that these authors find that such positioning of the options did have an effect, but there were alternative explanations that fit their data.

    Anyway, I tried to devise my options to make them balanced to avoid this effect (just in cases). I.e. I tried to make all the options fairly evenly spread along the (supposedly) key dimensions of net lives saved and amount of purposeful physical contact. But, thanks to your comment, I can see that perhaps option 4 is too close to option 5 - it's dominated.

    So, if I have created an unbalanced set of options, specifically too much grouping at the "push" end, then the "do nothing" option might be more attractive than a fair test would make it. If this is right, then we would expect the % of respondents morally endorsing "push" to be less in my actual 5-option scenario than it would have been in a fair 5-option scenario test. I don't think the setup allows us to meaningfully test this hypothesis anyway (the 3-option version seems balanced, but there are obviously other differences between the 3- and 5- option versions that make direct comparison problematic; e.g. different number of options!).

    So, unless I'm getting this mixed up, if your hypothesis is correct, the results would be even more surprising in both of the ways I mention.

  3. Dan, thanks for the reply. It seems like you are right that interesting result 2 remains surprising, and perhaps even more surprising (unless it is just fully explained by something else like poorly designed experiments or an evolving moral consciousness.)

    I'm not sure I completely understand why Interesting result 1 is more surprising. When we introduce a dominated option we effectively create a preference for the dominating one that wouldn't have been previously expressed. So there is a possible explanation of the increase that we wouldn't have had if there were no such relation. So it seems like the question would be how strong the decoy effect is relative to the opposite effect that one predicts based on the smaller number of choices. I guess that would be pretty straightforward to test. Just run your 5 option as a 4 option with the dominated option removed.

    I am probably not getting something.

  4. What I mean by result #1 being more surprising is that: assuming your hypothesis is correct, the actual gap between the '% reporting it permissible to push' scores of the 2- and 5-option versions would be even bigger (and therefore, more surprising).

    Maybe we disagree about how the decoy affect would work in this situation. The way that I think adding options in between the two original options works is that it creates a preference for the original option that is most different from the new option. This is the way it was described as working in the empirical paper I mentioned. They call it the similarity effect:

  5. But, if the decoy effect makes the "push" option more acceptable, then I can see how my comments would have been mysterious.

    If the effect of option 4 is to make option 5 more attractive simply because it is clear that option 5 is better than at least one other option (and that is not clear for the other options), then I can see that option 4 should be changed to "poke with a stick" rather than "push" to get a less bias result.

  6. My impression is that this last comment does summarize the basic idea behind the decoy effect. Marketers use it to specifically create a preference for an option that did not exist before.

    To me "poke with a stick" is still too touchy. I would say activate a trap door on the bridge with your cell phone.

    1. OK, you're right about decoy effect. I was thinking similarity effect. Confusingly, they are very similar but have opposite effects. And, both are used in marketing. For example, similarity effect explains this: Originally there are two options for sides at a restaurant, fries and baked potato. When a third option is added that is similar to baked potato, such as salad (both are healthy), a greater proportion of people chose fries than when there were just two options.

      They key difference seems to be whether the new option is dominated (in all respects) by one of the existing "poles". In my 5-option version, you are right to point out that option 4 is dominated by option 5. So, the difference between the 3- and 5-option versions might be explained by triggering the decoy effect, not by the presence of more options.

      Thanks for pointing this out!

  7. I have one other very half-baked thought. My guess is that when most people report what they believe to be morally permissible or impermissible they are making a prediction about how guilty they would feel if they were to do it. I wonder if the more examples we are required to consider the less in touch we are with our feelings of guilt, and the more likely we will make a judgment based on impersonal utility calculations. This could make it a typical System 1 vs. System 2 situation. It also seems vaguely consistent with research showing that people are more inclined to come to the aid of others the less time they have to think about it.

    1. I do think guilt affects some respondents (at least). When students justify their view in class, they often mention feeling bad when meeting their victim's family (on push and switch scenarios). But then someone points out that the 5 potential victims on the track have families too. This is where I think subconscious processing of the scenario favors not pushing. It is generally accepted in society that not doing something is less likely to cause offense or elicit condemnation. So, I think we have a default leaning toward inaction in cases like this. We subconsciously sense the approbation that would follow killing one, and this weighs more heavily in the subconscious processing than the positive feedback elicited by saving the 5.

  8. And I have a crazy explanation for interesting result #2. The way that people deal with information has changed dramatically since the dawn of the internet. People now process fantastically greater degrees of information and text than ever before. Possibly this has contributed to the new generations making a quantum leap in WEIRDness.

    1. This is not obviously crazy... but maybe :)

      Right now, I'm thinking about this in relation to the experience machine, and wondering whether weirdness is causing more young people (especially) to want to connect to an experience machine.

  9. Very interesting study, Dan! On Result #2: I'm not sure what to make of it. Maybe there is something different about the way you presented the cases (compared to previous research)? Maybe you could post the exact wording of the vignette and question? Or is what you have here the exact wording?

    On Result #2: I like Randy's ideas too, but here are some additional explanations:

    (1) It could be that increasing the number of options puts participants under increased cognitive load. If Josh Greene is right, this will make them switch more to manual mode thinking, which he says tends to be more consequentialist.

    (2) Or it could be that participants in the latter two conditions experienced increased cognitive load and just latched onto the most salient differences between your options, which are the numbers of people that would die as a result of the choice.

    On either story, the idea is that those who say "Push!" lose sight of the consideration that it involves harming as a means or using personal force or whatever.

    Finally, would you mind if I made a post on the x-phi blog that links to this discussion?

    1. Hi Josh, the wording I used is the wording at the start of the blog.

      I see your other suggestions as potentially compatible with Unger's view that the consequentialist action is morally preferable. Unger argues that when we think deeply about situations like this, most of us realize that saving lives is the most important thing. But, as you say, perhaps this is merely the most accessible feature of the scenario given that net lives lost is a feature of every option, but pushing is only in a subset of the options.

      And, yes, feel free to link to the x-phi blog.

    2. Okay, so regarding failure to replicate previous results (of low frequency permissibility judgments in the standard Footbridge case):

      One thing your text leaves out is referring to the one man as a "heavy object" (Mikhail 2007) or "heavy weight" (Hauser et al 2007). Some have suggested this may account for lower permissibility ratings in those experiments (Greene et al 2009, following Waldmann & Dieterich 2007). After all, then the one man is then being described as being used like a mere object, which might seem even less morally acceptable (I mention this issue in my Phil Compass paper).

      Still, though, I think other studies might have avoided this language and still seen a much lower percentage of permissibility judgments than you saw. But I'd have to dig into some studies to tell.

    3. For some reason I am now curious how many people would say it is not permissible to push the big man off the bridge even if he is just weight, i.e., already dead. Josh, hanks for mentioning your Phil Compass article. I wasn't aware of it.. We do have a subscription here and I've been using them a lot for teaching in the last few years..

    4. Thanks Josh, Your paper is great. I should have known about that. I think I'd need to see some harder evidence to believe that referring to the man as a "heavy object" (or some such) makes much difference. Pushing someone to death to save some others is using the person as an object in many ways. Additional priming for objectifying him may have some effect, but it seems much less relevant than the battery/assault/killing that is taking place.

      Also, in an online study using a slightly different scenario (the man had on a large heavy backpack like in some of Greene's scenarios) we got very similar results.

      We are thinking about another round of experiments, so we may get more data to help figure this out.

      Randy, hopefully most would think that we should push the dead man... but one thing I'm learning is that it's hard to get complete agreement on any kind of moral issue when using surveys. So, let's say 10% wouldn't push the dead man. Should we think they haven't bothered to read the question? Perhaps they are joking? Perhaps they have some moral justification for it (e.g., I couldn't do it, and ought implies can). Our new study might try to address this kind of problem in relation to trolley scenarios.

    5. Dan, I bet a lot would say it is morally worse to use a dead man to stop the trolley than to use, say, a live horse. Defiling dead human bodies is taboo. And so I also wouldn't be surprised if the taboo is strong enough to override the felt moral difference between causing and letting something happen. But a lot of thing wouldn't surprise me that are no doubt false.

  10. Dan,

    I'm inclined to believe that the participants merely weighed the 5 options against each other, saw that the other options were much more impermissible than 5 (or that 5 was the least impermissible) and then drew the conclusion that the action was permissible from that inference.

    1. Hi Brandon, we actually ran another study at the same time as this one that produced data relevant to this idea. Using a very similar question as for the 5-option version above, we said that this situation occurs in 5 different states around the country and there was a bystander in each state. The bystander in state 1 took option 1 (do nothing), the bystander in state 2 took option 2 (flick a switch so the trolley kills 4 new people instead of the 5 original people), the bystander in state 3 took option 3 (flick a switch so the trolley kills 3 new large men on a raised side track, saving the 5 originals), etc.
      We then asked them to rate on a scale how moral all of the actions were. The scale had 7 points with one end labeled "very morally bad", the other labeled, "very morally good", and the middle point labeled "morally neutral".

      We could assign points and average the results or count up the number who rated each action permissible (morally neutral or better) and work out the proportions. Either way, option 5 was not the most moral.

      I'd say more, but I'll miss my bus!


    This is a link to a Ted talk that discusses how limiting choice can make a person more inclined to participate in an event. Or something like that, sorry about the lack of empirical backing.
    Personally, I'm interested in how responses might vary according to active or passive participation in such an event, i.e. physically pushing a person as opposed to controlling the direction of the cart. I believe the delusion these respondents are under while proposed a hypothetical situation involving the use of physical force has a significant impact on the decision; as hypothetics removes the necessity of physical action. Should they actually be placed in that situation I believe you would quickly see the answers they give become devalued.
    I cannot think of any biases congruent with this hypothesis, just my thoughts for now.
    Good luck with your research.

    1. Thanks for this Brenton, I agree that people would behave differently if they were
      really faced with the dilemma. But, I think the main reason for that is the idealized nature of thought experiments. I cannot imagine a real life scenario in which I would be sure that pushing a fat man off a bridge would stop a runaway trolley.

  12. It's the sample.
    You're comparing your results with some others. How were your samples taken as compared to theirs? Your college is a "commuter" college and you would expect the students less diverse, generally, than at a big university.
    How did you select and administer the poll? All at the same time. If not, perhaps?
    Ralph ...
    those who took it later were (innocently) briefed on its substance.
    Did all the subjects know what "morally permissible" means, pretty much exactly? When I was an undergraduate, I doubt if I did.
    Ultimately, you have to ask yourself whether the subjects were answering the question you asked and not something else related to the narrated situation. The "large" man business might easily have thrown them off. Did they understand the reason for that specification?
    I'm really interested in exactly how you chose the sample and how the question was administered, i.e. all together, or what

  13. Thanks for these suggestions. The surveys were actually conducted in New Zealand At a university that is very WEIRD and homogeneous compared to Sac State. Also the results match quite well with online audiences from two different sources.

    Not explaining morally permissible did trip a couple of people up. I removed from the analysis 2 respondents who admitted not knowing what it meant.

    I probably should have explained the significance of the man being large too because a few people suggested the bystander should have jumped instead. Still, explaining this would probably have made even more people report pushing as permissible.

  14. Thanks for these suggestions. The surveys were actually conducted in New Zealand At a university that is very WEIRD and homogeneous compared to Sac State. Also the results match quite well with online audiences from two different sources.

    Not explaining morally permissible did trip a couple of people up. I removed from the analysis 2 respondents who admitted not knowing what it meant.

    I probably should have explained the significance of the man being large too because a few people suggested the bystander should have jumped instead. Still, explaining this would probably have made even more people report pushing as permissible.