Content uploaded by Yiyun Shou
Author content
All content in this area was uploaded by Yiyun Shou on Aug 02, 2015
Content may be subject to copyright.
Moral Reasoning as Probability Reasoning
Yiyun Shou (yiyun.shou@anu.edu.au)
Research School of Psychology, The Australian National University
Canberra, ACT, Australia
Fei Song (sfei618@gmail.com)
Department of Philosophy, The University of Hong Kong, Hong Kong
Abstract
Previous studies found that the likelihood of subjects to
choose a deontological judgment (e.g., allowing harm)
or a consequentialist judgment (e.g., doing harm) varied
across different moral dilemmas. The present paper
explored if the variation can be explained by the
differentiation of the perceived outcome probabilities.
We generated moral dilemmas that were similar to the
classical trolley and footbridge dilemmas, and
investigated the extent to which subjects were sensitive
to the outcome probabilities. Results indicated that the
majority of subjects, including both those who initially
chose a deontological decision and those who initially
chose a consequentialist decision could be sensitive to
outcome probabilities. The likelihood of being sensitive
to the probabilities was invariant across different
dilemmas. The variation of the choice behaviors across
different dilemmas might be associated with the
variation of the estimated outcome probabilities.
Keywords: probability judgment; moral reasoning;
moral dilemma
Introduction
Moral reasoning has been under long-term intellectual
scrutiny. Recent psychological investigations of moral
reasoning frequently employ moral dilemmas that render
conflicts between moral requirements (Crockett, 2013).
Moral dilemmas commonly engender conflict between two
major type of moral reasoning: deontological and
consequentialist moral reasoning. A deontological moral
judgment primarily concerns the actions per se, that is,
whether it is consistent with moral principles, rules or duties.
On the other hand, a consequentialist judgment primarily
concerns the outcome of each possible action and aims to
choose the one with the best outcome.
The trolley dilemma requires people to decide between
killing an innocent individual and allowing five innocent to
be killed. The former is often considered as a consequentialist
decision, under which the loss is minimized. In contrast,
allowing five to die is taken as deontological decision under
which the action of killing is regarded as a deontological
violation. Crockett (2013) associated consequentialist
reasoning with a model based system, in which the reasoner
starts from the current action, searches through the decision
tree and evaluates the best outcome of the action. In contrast,
the deontological reasoning is associated with the model-free
evaluation, where the current action and forward searching
are not activated.
Recent research suggests that the types of moral reasoning
may be shaped by the interaction and competition between
two distinct psychological systems: an automatic emotion
process and a controlled conscious reasoning process
(Greene, Sommerville, Nystrom, Darley, & Cohen, 2001;
Paxton, Ungar, & Greene, 2012). Greene et al.(2001) argued
that a deontological decision might be driven by emotional
arousal, while consequentialist reasoning is the result of the
controlled reasoning process.
Numerous studies have shown that the majority of people
perceive the consequentialist choice as morally preferred
option in the trolley dilemma (Crockett, 2013). However,
controversial findings rose in different variants of moral
dilemmas. For instance, the footbridge dilemma—in which
one needs to decide between pushing a fat man over bridge
and allowing five people to die—yielded distinctive decision
patterns (Lerner, Li, Valdesolo, & Kassam, 2014; Valdesolo
& DeSteno, 2006). The proportion of subjects who preferred
the consequentialist choice over the deontological choice can
vary case by case (Cummins & Cummins, 2012). These
studies suggest that some case–relevant features might
influence the decision making process.
Greene made a distinction between a personal dilemma like
the trolley dilemma and an impersonal dilemma like the
footbridge dilemma. The personal dilemma triggers the
negative response to a harmful act which treats an agent as
the only means to an end, whereas the impersonal dilemma
fails to trigger negative response to a harmful act which is
only a side effect (Greene, Nystrom, Engell, Darley, &
Cohen, 2004). It has been systematically found that personal
dilemmas commonly produce more deontological judgments
among subjects, while impersonal dilemmas commonly
produce consequentialist judgments (Cummins & Cummins,
2012; Moll & de Oliveira-Souza, 2007).The personal
dilemmas, featuring the involvement of the physical contact,
may trigger higher emotional arousal, which in turn result in
the higher likelihood among subjects to choose
deontological-like judgments.
Greene et al (2001) presented neuroimaging evidence and
showed that the footbridge dilemma was associated with
greater activity in emotion associated brain areas such as the
posterior cingulate gyrus (Brodmann Area 23/31) and
bilateral angular gyrus (Brodmann Area 39). Greene (2009)
implied that the physical contact may induce emotional
arousal, which promotes subjects to be more likely to engage
2176
in deontological reasoning. In fact these two brain areas are
not restricted to emotion relevant processes. For example, the
posterior cingulate gyrus was also found to be associated with
the cognitive process involved in the evaluation of the values
of choices (Rushworth & Behrens, 2008), while the bilateral
angular gyrus was found to be activated during decision
making under uncertainty (d’Acremont, Fornari, &
Bossaerts, 2013).
An alternative explanation for the higher likelihood of
subjects’ preferences for allowing harm in personal dilemmas
like the footbridge case can be that the decision of allowing
harm can be a result of either endorsing deontological
reasoning or endorsing consequentialist reasoning. The
consequentialist reasoning recruits the model-based
evaluation while the deontological reasoning recruits the
model-free evaluation. To avoid confusion, we distinguish
the two reasoning types from the two choices observed in a
moral dilemma. We name the choice of “doing harm” as
consequentialism-like choice (CLC), and the choice of
“allowing harm” as the deontology-like choice (DLC). The
DLC that allows five people to die can be perceived as
justifiable via either types of reasoning. A reasoner who
adopts consequentialist reasoning can make a DLC when the
perceived utility of doing harm is lower than the utility of
allowing harm. The perceived utility can be altered by the
reasoner’s probability estimates of the outcomes given the
two choices.
Most previous studies did not explicitly indicate how likely
the outcomes would occur given that each action had been
taken when presenting the moral dilemmas. Subjects may
estimate the outcome probabilities based on their prior
knowledge or experience with the scenario in a dilemma.
Subjects may be more likely to make a DLC when they
perceive the positive outcome given doing harm as less likely
compared to the one given allowing harm. For instance, in the
footbridge vignette, subjects may perceive the probability
that the fat man being pushed over the bridge can stop the
trolley and thereby five people will be saved being much
lower than 100%. The concept of “doing harm” aversion (i.e.,
prefer a choice of avoiding doing harm to a choice of doing
harm ) in the footbridge vignette is analogy to the concept of
risk aversion (prefer a choice with certainty to a choice with
risk) (Rogers, Viding, & Chamorro-Premuzic, 2013).
In a preliminary investigation (Song & Shou, 2014), we
used the classical trolley dilemma and footbridge dilemma,
and asked subjects for their preference between the CLC and
DLC in each dilemma. Depending on their preference,
subjects were then asked if they would change their decisions
if the outcome probability of their previous decision was not
100%. About 40% subjects, including those who initially
chose CLC and those who chose DLC, altered their
preferences. In addition, subjects who initially chose CLC
were more likely to alter their preference than those who
initially chose DLC. It was also found that the proportion of
subjects who chose to switch the choices was similar between
the trolley dilemma and footbridge dilemma.
Being sensitive to the outcome probability is a substantial
feature of a consequentialist reasoning, as it is in accordance
with the basic principle of consequentialism---maximizing
the expected utility of outcomes (Harsanyi, 1980; Hooker,
2000; Peterson, 2009). The results in Song and Shou (2014)
implied that consequentialist reasoning may be applied to
generate both CLC and DLC. The equal proportions of
subjects who were sensitive to outcome probability across
two dilemmas suggest that the proportion of subjects who
initiated consequentialist reasoning and subjects who
initiated deontological reasoning can be independent from the
type of dilemma. Instead, it is a matter of outcome
probability.
In the current study, we focused on the impact of outcome
probabilities on subjects’ moral decisions. We argued that the
moral decisions can be influenced by the outcome
probabilities, which may offer a new perspective for
rethinking the differences in moral decision across different
moral dilemma. We used three moral dilemmas that differed
in term of the extent to which they are like the personal or the
impersonal dilemma. We first assessed subjects’ moral
judgments without providing any probabilistic information
about the outcomes. We hypothesized that like the previous
studies, subjects are more likely to choose DLC in a personal
dilemma than when presented with an impersonal dilemma.
We then measured subjects’ sensitivity to outcome
probabilities. Subjects were presented with several paired
choices, each of which had the outcome probabilities
specified in two choices. If a subject adopts consequentialist
reasoning and evaluates the utility of the outcomes, his or her
decision should eventually shift to the alternative when the
expected outcome utility of their previsions decision (the
utility of an outcome discounted by its likelihood) becomes
lower than the expected outcome utility of its alternative.
In contrast, if a subject evaluates the action itself rather
than its outcome, then he or she should be insensitive to the
change of probabilities in the outcomes. We hypothesized
that subjects who applied consequentialist reasoning could
also make the DLC when there was no probability
information. That is, there would be subjects who chose DLC
initially without probability information shifting their
decisions when probability information was provided.
Subjects who initially selected CLC were more likely to
change the decision than those who initially selected DLC.
In addition, we argue that whether consequentialist or
deontological reasoning is applied does not depend on the
types of the dilemma. It is a matter of the judgmental model
a subject usually adopts. The proportion of subjects who are
applying the consequentialist reasoning and are sensitive
should be similar across different moral dilemmas. We
hypothesized that the proportions of consequentialist
reasoning based subjects who eventually shifted their
decisions should be independent from the types of moral
dilemma.
2177
Method
Participants and Procedure
A total of 161 subjects (109 females) were recruited via
online crowd-sourcing service CrowdFlower. Subjects aged
between 20 and 74, with a mean age of 39.47 years (SD =
11.73). Subjects were randomly assigned to one of the three
moral dilemmas vignettes described below. They read the
consent information, completed the demographical
questionnaire, the moral judgment task and the subsequent
pairwise choice comparison task in order.
Materials
There were three moral dilemma vignettes. The detailed
vignettes were available in online appendix. The first
vignette—flood vignette was an impersonal dilemma that was
similar to the trolley vignette. The CLC was to initiate the
explosion to sacrifice one person to prevent five people from
being flooded, while the DLC was to do nothing and allow
the five people to be flooded. The decision maker did not
have physical contact with the victims and the decisions
influence the victims remotely.
The second vignette – truck vignette was also an
impersonal dilemma. The CLC was about turning the truck
into one bystander in order to protect five people in a car,
while the DLC was about allowing five people to be hit
instead of sacrificing one person. In comparison to the flood
case, the truck case had greater distance between the decision
maker and the victims in the dilemma.
The final vignette—hostage vignette was a personal
dilemma and was similar to the footbridge case, where the
decision maker needed to physically contact the victim. The
CLC was to push a person over the cliff to prevent five
hostages from being killed by the gangster, while the DLC
was to do nothing and allow five hostages to be killed
1
.
The description of the vignette did not contain any
probabilistic information. After reading the vignette, subjects
were asked to judge “Which action do you think is morally
better?” The question asks subjects to compare two choices
in aspects of morality. The question, unlike other common
moral judgment questions that ask about “permitted” or
“wrong”, can draw subjects’ attention to morality per se from
law or convention (Baron, 2014). To engage subjects with the
dilemmas, black and white illustrations for each dilemma
were presented on screen throughout the whole task.
Next, we specified the probabilities for the outcomes of the
two actions and asked subjects to judge which action is more
morally correct. For example, the first comparison for the
flood dilemma was that “Now suppose we know that the
outcomes of your choices may not be 100% sure. Suppose if
you choose do nothing, it is 80% sure that the five miners will
die. On the other hand, if you choose to explode the
floodwall, it is 100% sure that the individual miner will die.
1
The details of materials as well as example illustrations are
available in online supplemental materials at http://goo.gl/hknhMJ
Given this new information, if you are asked to re-do the
judgment, which action do you think is morally better?”
We choose to specify the probability of this type of loss
(i.e., how likely the victims would die) for the following
reasons. First, we avoided the expression that “the victim
would be killed’ because “kill” implied an action that is not
morally neural, and may bias subjects’ decisions. Second, we
avoided using negative wording (i.e., how likely the victims
would not die) as subjects may have difficulty in judging the
probabilities for the negatively worded statement (Peterson,
2009). Finally, we avoid using vague outcomes such as “how
likely the victims would be alive” as the degree of harm
relating to being alive is more ambiguous than death.
We changed the probabilities until subjects changed their
mind to prefer the alternative choice. Figure 1 illustrates the
paired judgments process. The probability-attached choice
questions started by a comparison of a DLC with 80% chance
of loss against a CLC with 100% chance of loss (the red
circle). If a subject chooses CLC, which means that the
subject perceives 80% chance of that five people die is
morally worse than 100% chance of that one person dies, then
he or she proceeds to the second question, in which the
probability of the loss in DLC decreased to 20%. If the
subject altered the choice and preferred DLC, the third
question would increase the probability of the loss in DLC to
50%. After this question, we also narrowed down the possible
equivalence of the two choices to the subject. If the subject
chose DLC in the third question, then the subject perceives
the 100% chance of one death in DLC is morally equivalent
to 20% - 50% chance of five deaths in CLC. On the other
hand, if the subject chose CLC in the third question, then the
subject perceives 100% chance of one death in DLC is
morally equivalent to 50% - 80% chance of five deaths in
CLC.
Likewise, if a subject chooses DLC in the initial
comparison, which means that the subject perceive 100%
chance of one death is morally worse than the 80% chance of
5 deaths, then he or she proceeds to the next question, in
which the probability of the loss in CLC further decreased.
There are eight resultant categories of this decision task as
shown in Figure 1, the moral equivalence of the outcomes are
summarized in Table 1.
Subjects who stopped in category 1 are those who insisted
DLC regardless how low the chance that one individual
would die. Stopping in category 8 suggests the subject
insisted CLC as the morally better choice even when the
expected utility of the two choices are indistinguishable.
2178
Table 1: Types of subjects based on the paired comparison
where subjects shifted their decisions
Outcome probability
Category
DLC (5 deaths)
CLC (1 death)
1
100%
[0%, 20%]
2
100%
[20%, 50%]
3
100%
[50%, 80%]
4
100%
[80%, 100%]
5
[80%, 100%]
100%
6
[50%, 80%]
100%
7
[20%, 50%]
100%
8
[0%, 20%]
100%
Results
Moral Decisions
Table 2 displays the frequencies and percentages of the
moral decisions across different dilemmas. Logistic
regression was conducted to model the choices in different
dilemmas by using the CLC as the baseline choice. The type
of the moral dilemmas had substantial contribution to the
regression model, χ2 = 33.54, p < .001, indicating subjects’
initial moral decisions were significantly different across
different dilemmas. Subjects in the hostage dilemma were
slightly less likely to choose CLC than DLC, b = -0.51, p =
.064
2
. In support of our first hypothesis, subjects in the
hostage dilemma were significantly less likely to choose CLC
than subjects who were in the other two dilemma, b =-1.74, p
2
We change the dummy coding scheme to obtain the coefficient
estimate. The results of the comparisons across the different
dilemmas were obtained by conducting three versions of logistic
< .001 in compared to the truck dilemma, and b = -2.37, p <
.001 in compared to the flood dilemma.
Subjects in the truck and flood dilemma were more likely
to choose CLC than DLC, b = 1.23, p <.001, odds ratio = 3.4
for truck dilemma, and b = 1.86, p <.001, odds ratio = 6.4 for
flood dilemma. The likelihood of choosing CLC among
subjects in the flood dilemma was not significantly different
from those in the truck dilemma, b = 0.63, p = .226.
Table 2: Frequencies and percentages of the moral decisions
across different dilemmas
Vignette
DLC
CLC
Total
Truck
12 (22.6%)
41 (77.4%)
53
Flood
7 (13.5%)
45 (86.5%)
52
Hostage
35 (62.5%)
21 (37.5%)
56
Sensitivity to Probability Information
Subjects who were finally in category 1 (insisting DLC)
and category 8 (insisting CLC) were regarded as those who
were insensitive to probabilistic information. The numbers
and proportions of those subjects are displayed in Table 3.
On average, the majority of subject (79.2%) were influenced
by the change of the probabilities of the outcomes and finally
shifted their choices.
regression; each of each treats one of the dilemma as the base
comparison group.
Figure 1. Illustration of the logic flow in probability-attached decision making questions
2179
Table 3: Frequencies and proportion of subjects who did not
switch the choice due to the change of probability
information
Dilemma
DLC
CLC
Total
Truck
5/12
(41.67%)
3/41
(0.07%)
8/53
(15.1%)
Flood
2/7
(28.57%
10/45
(22.22%)
12/52
(23.1%)
Hostage
13/35
(37.14%)
0/21
(0%)
13/56
(23.2%)
Total
20/52
(38.46%)
13/107
(12.15%)
38/161
(20.5%)
A logistic regression model was conducted on the
likelihood of subjects being sensitive to the different
probabilities, predicted by the type of dilemmas and their
initial choices. The likelihood of shifting choices was
significantly different between subjects who initially chose
DLC and those who initially chose CLC, χ2= 12.98, p <.001.
As expected by the second hypothesis, subjects who chose
CLC were significantly more likely to be influenced by the
probabilistic information and changed their choice than those
who chose DLC, b = 1.84, p <.001, odds ratio = 6.2.
Furthermore, in support to the third hypothesis, the type of
dilemmas did not have significant contribution to the model
fit, χ2= 1.47, p =.479, suggesting the proportion of subjects
who was sensitive to the probabilities was similar across the
three dilemmas.
Discussion
In the current study, we used three moral dilemmas with
similar attributes as the personal/impersonal cases (i.e.,
trolley case and the footbridge case.) The results were similar
to previous studies, where subjects were substantially more
likely to choose the consequentialism-like choice (CLC) in
the impersonal dilemma (flood and car) than the personal
dilemma (hostage). When being provided the probabilistic
information about outcomes, about eighty percent of subjects
eventually changed their decision on which was a morally
better choice. Being influenced by the outcome probability
indicated that those subjects might be employing
consequentialist reasoning. Subjects who initially chose CLC
were more likely to change their choice with the change in
probabilistic information, indicating that subjects who chose
CLC were more likely to engage in consequentialist
reasoning than those who initially chose DLC. This result
implied that most people may apply model-based evolution
associated with probabilities in moral reasoning. The
resultant decisions from the reasoning algorithm are not
restricted to a deontology-like decision.
It was also found that the proportions of subjects who were
sensitive to the probabilities were similar across the three
dilemmas. This suggests that the likelihood of subjects to be
engaged in consequentialist reasoning in moral judgments
might be independent from the type of moral dilemmas. The
large proportion of choice shift in the hostage case among
subjects who chose DLC initially, further suggests that the
higher likelihood of choosing DLC in the personal dilemmas
can be associated with the perceived probabilities of outcome
being different from those in the impersonal dilemmas. As
indicated by Crockett (2013) evaluation of the consequences
in model-based moral reasoning can be influenced by the
prior experience of subjects in associated with the event in
the dilemma. The different levels of experience across
different events contribute to the different evaluations for the
doing harm and allowing harm across different dilemmas.
Interestingly, several subjects who initially chose CLC did
not change their choices even when the expected utility of the
CLC (the one person has 100% chance to die) was lower than
the DLC (the five people have 20% chance to die). One
possible explanation is that a subject who chose CLC may
also adopt a decision heuristics in system 1, which involve
fast and intuitive processes (Evans, 2003). Those subjects
made their decisions by comparing the number of loss in each
case without evaluating the relative weight of outcomes. We
may call them outcome probability insensitive
consequentialist reasoner. The other explanation is that, to
those subjects, the aggregated utility of five peoples’ life can
be greater than the single person’s life, result in the perceived
utility of 20% chance of five people dying as greater than the
utility of 100% chance of one person dying.
Another interesting finding was that many subjects
changed their decision as soon as the outcome probability of
the previous one option had an outcome probability lower
than 100%. The majority of subjects stopped at the category
where they perceived the consequence of 50% to 80% of the
loss (five people would die) in CLC as being equivalent to
the 100% of the loss in DLC (one person would die One
possible explanation is that, subjects were risk seeking (i.e.,
preferred a choice whose outcome is between 0% and 100%
over a choice whose outcome is 0% or 100%) in the loss
domain even in moral reasoning. Risk seeking behaviors as a
result of loss aversion was well documented in decision
making literature (d’Acremont et al., 2013). Subjects in the
present study might prefer a choice whose probability of loss
(five deaths) was lower than 100%, over the alternative
choice whose outcome (i.e., one death) probability was
100%.
An alternative explanation is that the outcome utility
calculation may also involve the evaluation of action cost.
The action of killing may yield action costs associated with
social conventions, moral responsibilities and law
obligations. That may be the reason why the outcome utility
of one death with 100% certainty in CLC is greater than the
five deaths with 20% certainty. Both explanations may need
further investigation in future studies.
Limitations and Conclusion
The current experiment demonstrates that most subjects
were evaluating consequences in moral reasoning when the
probabilistic information was provided. One may argue that,
the explicit probability information may induce the adoption
of consequentialist reasoning, as consequentialist reasoning,
in contrast to deontological reasoning, is a cost-benefit
2180
calculation with probability. This may not be the direct
evidence for what people did naturally and intuitively. Future
studies in moral dilemmas may assess subjects’ prior belief
on the probabilities for both the positive outcomes and the
negative outcomes to better understand how the probabilistic
factor could influence people’s moral reasoning.
In conclusion, the results indicate that choosing a DLC
does not entail that people engage in consequentialist
reasoning, whereas choosing a CLC does not entail that
people engage in consequentialist reasoning. The differences
across different moral dilemmas are very likely due to the fact
that subjects perceive the outcome probability as insufficient
for one to choose the alternative choice.
References
Baron, J. (2013). Moral judgment: Acts, omissions, and rules.
Unpublished manuscript, Department of Psychology,
University of Pennsylvania, Philadelphia, United
States.
Crockett, M. J. (2013). Models of morality. Trends in
Cognitive Sciences, 17(8), 363–6.
Cummins, D. D., & Cummins, R. C. (2012). Emotion and
deliberative reasoning in moral judgment. Frontiers in
Psychology, 3:328.
d’Acremont, M., Fornari, E., & Bossaerts, P. (2013). Activity
in inferior parietal and medial prefrontal cortex signals
the accumulation of evidence in a probability learning
task. PLoS Computational Biology, 9.
Evans, J. (2003). In two minds: dual-process accounts of
reasoning. Trends in Cognitive Sciences, 7(10), 454–
459. doi:10.1016/j.tics.2003.08.012
Greene, J. (2013). Moral Tribes: Emotion, Reason, and the
Gap Between Us and Them. Penguin Press HC.
Greene, J. D., Sommerville, R. B., Nystrom, L. E., Darley, J.
M., & Cohen, J. D. (2001). An fMRI investigation of
emotional engagement in moral judgment. Science
(New York, N.Y.), 293, 2105–2108.
Greene, J., & Haidt, J. (2002). How (and where) does moral
judgment work? Trends in Cognitive Sciences, 6(12),
517–523.
Harsanyi, J. C. (1980). Rule utilitarianism, rights, obligations
and the theory of rational behavior. Theory and
Decision, 12(2), 115–133.
Hooker, B. (2000). Ideal Code, Real World: A Rule-
consequentialist Theory of Morality. Oxford University
Press.
Lerner, J. S., Li, Y., Valdesolo, P., & Kassam, K. (2014).
Emotion and decision making. Annual Review of
Psychology, 66: 799-823.
Moll, J., & de Oliveira-Souza, R. (2007). Moral judgments,
emotions and the utilitarian brain. Trends in Cognitive
Sciences, 11, 319–321.
Paxton, J. M., Ungar, L., & Greene, J. D. (2012). Reflection
and reasoning in moral judgment. Cognitive Science,
36(1), 163–177.
Peterson, M. (2009). An Introduction to Decision Theory (1st
Ed). New York: Cambridge University Press.
Rogers, J., Viding, E., & Chamorro-Premuzic, T. (2013).
Instrumental and disinhibited financial risk taking:
Personality and behavioural correlates. Personality and
Individual Differences, 55, 645–649.
Rushworth, M. F. S., & Behrens, T. E. J. (2008). Choice,
uncertainty and value in prefrontal and cingulate cortex.
Nature Neuroscience, 11, 389–397.
Song, F. and Shou, Y, (2014). Affect and moral judgments,
paper presented at Intentional Conference on
Epistemology and Cognitive Science, Xiamen, China,
Jun 2014.
Valdesolo, P., & DeSteno, D. (2006). Manipulations of
emotional context shape moral judgment.
Psychological Science, 17(6), 476–7.
2181