A computational account of how individuals resolve the dilemma of dirty money

Abstract

Money can be tainted when it is associated with direct or indirect harm to others. Deciding whether to accept “dirty money” poses a dilemma because money can be used to help others, but accepting dirty money has moral costs. How people resolve the dilemma of dirty money remains unknown. One theory casts the dilemma as a valuation conflict that can be resolved by integrating the costs and benefits of accepting dirty money. Here, we use behavioral experiments and computational modeling to test the valuation conflict account and unveil the cognitive computations employed when deciding whether to accept or reject morally tainted cash. In Study 1, British participants decided whether to accept “dirty” money obtained by inflicting electric shocks on another person (versus “clean” money obtained by shocking oneself). Computational models showed that the source of the money (dirty versus clean) impacted decisions by shifting the relative valuation of the money’s positive and negative attributes, rather than imposing a uniform bias on decision-making. Studies 2 and 3 replicate this finding and show that participants were more willing to accept dirty money when the money was directed towards a good cause, and observers judged such decisions to be more praiseworthy than accepting dirty money for one’s own profit. Our findings suggest that dirty money can be psychologically “laundered” through charitable activities and have implications for understanding and preventing the social norms that can justify corrupt behavior.

Full text

1
Vol.:(0123456789)
Scientic Reports | (2022) 12:18638 | https://doi.org/10.1038/s41598-022-22226-9
www.nature.com/scientificreports
A computational account
of how individuals resolve
the dilemma of dirty money
Jenifer Z. Siegel 1,2,6, Elisa van der Plas3,6, Felix Heise, John A. Clithero 4,7 &
M. J. Crockett 2,5,7*
Money can be tainted when it is associated with direct or indirect harm to others. Deciding whether
to accept “dirty money” poses a dilemma because money can be used to help others, but accepting
dirty money has moral costs. How people resolve the dilemma of dirty money remains unknown. One
theory casts the dilemma as a valuation conict that can be resolved by integrating the costs and
benets of accepting dirty money. Here, we use behavioral experiments and computational modeling
to test the valuation conict account and unveil the cognitive computations employed when deciding
whether to accept or reject morally tainted cash. In Study 1, British participants decided whether
to accept “dirty” money obtained by inicting electric shocks on another person (versus “clean”
money obtained by shocking oneself). Computational models showed that the source of the money
(dirty versus clean) impacted decisions by shifting the relative valuation of the money’s positive and
negative attributes, rather than imposing a uniform bias on decision-making. Studies 2 and 3 replicate
this nding and show that participants were more willing to accept dirty money when the money
was directed towards a good cause, and observers judged such decisions to be more praiseworthy
than accepting dirty money for one’s own prot. Our ndings suggest that dirty money can be
psychologically “laundered” through charitable activities and have implications for understanding and
preventing the social norms that can justify corrupt behavior.
In early 2019, dozens of organizations began refusing charitable donations from the Sackler family, whose mem-
bers stood accused of fueling the deadly opioid crisis. Later that year, a scandal erupted upon revelation that
researchers at MIT and Harvard had accepted nearly a million dollars from convicted sex oender Jerey Epstein.
ese are just new examples of an old dilemma. e concept of “dirty money” dates back thousands of years,
and can be dened as money associated with direct or indirect harm to others. Deciding whether to accept dirty
money poses a dilemma because money can be used to help others, but accepting dirty money has moral costs1.
How do people resolve this dilemma?
Research on “moral contagion” demonstrates that the value of objects can be tainted by association with moral
misdeeds2–4. For example, participants will pay less money for items previously owned by immoral individuals5,6.
ere is also evidence that money itself is less subjectively valuable when it is obtained immorally. People imagine
they would rather not receive money that is morally tainted7, and when faced with actual decisions to accept
money from an experimenter in exchange for the experimenter inicting painful electric shocks on oneself
or another person, most people would rather receive money from an experimenter harming themselves over
someone else8. Neuroimaging studies demonstrate that morally tainted money and objects are associated with
lower activity in the brain’s valuation network5,9,10. However, it remains unknown what cognitive computations
are employed during decisions about whether to accept or reject morally tainted cash.
A recently proposed theory casts the dilemma of dirty money as a valuation conict1. According to this
theory, when people are deciding whether to accept dirty money, they must resolve a conict between its posi-
tive attributes (e.g., its ability to purchase desirable goods or help others) and negative attributes (e.g., harmful
outcomes associated with the money). Strengthening positive attributes, or weakening negative attributes, can
OPEN
1Department of Experimental Psychology, University of Oxford, Oxford, UK. 2Department of Psychology, Yale
University, 2 Hillhouse Ave, New Haven, CT 06511, USA. 3Institude of Neurology, University College London,
London, UK. 4Lundquist College of Business, University of Oregon, Eugene, USA. 5Department of Psychology
and University Center for Human Values, Princeton University, Princeton, USA. 6
These authors contributed
equally: Jenifer Z. Siegel and Elisa van der Plas. 7These authors jointly supervised this work: John A. Clithero and
M. J. Crockett. *email: mj.crockett@princeton.edu
Content courtesy of Springer Nature, terms of use apply. Rights reserved

2
Vol:.(1234567890)
Scientic Reports | (2022) 12:18638 | https://doi.org/10.1038/s41598-022-22226-9
www.nature.com/scientificreports/
shi the balance towards accepting dirty money. e idea that people might shi the subjective value of a deci-
sion’s positive and negative attributes to resolve moral conicts is not new. Past research has shown that people
can maintain a positive self-image of themselves despite harmful actions by manipulating the value of the moral
costs and personal benets of the choice, allowing the individual to benet from their immoral actions while
maintaining a positive self-view11. Moral judgments are hypothesized to play a central role in assigning positive
or negative value to attributes and subsequent decisions. For example, judging the source of the money to be
morally bad is predicted to decrease the likelihood of accepting the money, while judging the destination of the
money to be morally good is predicted to increase the likelihood of accepting the money12–14.
Here, we tested the predictions of the valuation conict hypothesis across a series of studies (summarized
in Fig.1). Across all studies we operationalized “dirty money” as money obtained by an experimenter inicting
painful electric shocks on another person, and “clean money” as money obtained by an experimenter inicting
shocks on oneself (Fig.2a). In Study 1, participants decided whether to accept or forego dierent amounts of
money which varied in their source (dirty or clean). In Study 2, we tested whether moral judgments of decisions
to accept dirty money were sensitive to the destination of the money (personal prot or charitable donation).
Finally, in Study 3, participants decided whether to accept or forego dierent amounts of money which varied
in both their source and destination. e valuation conict account predicts that people would have a lower
probability of accepting dirty than clean money, and additionally, that directing dirty money towards a charitable
cause will mitigate source eects, eectively “laundering” its value through the charitable act.
To decompose the cognitive processes that unfold during decisions about dirty money, in Studies 1 and 3
we modeled participants’ decisions and response time distributions using a multi-attribute extension of the
dri–diusion model (DDM)15,16. In this model, decisions to accept or reject money are compared by comput-
ing a subjective value over multiple attributes. Over time, this value integration represents accumulated value
in favor of one option over another. A choice is made when the decision variable passes a threshold for one of
the choice options.
By parameterizing several distinct aspects of the decision process, the DDM can also test the valuation conict
account of the dirty money dilemma (Fig.2b). e dri rate parameter v captures the rate of value accumulation
toward one of two options. is parameter relates to the eciency at which the choice attributes (i.e., shocks
and money) are processed. e bias parameter z captures the extent to which people lean toward accepting the
larger amount of money before they know how much money is at stake. erefore, the bias parameter z captures
cognitive processing that occurs before the specic attributes of the choice are known. e cognitive dierence
between these two parameters is the extent to which the choice attributes aect the evaluation of the options (as
is the case for the dri rate, v) or whether this evaluation is aected prior to observation of the attributes of the
choice (as is the case for starting point bias, z). e parameter a represents the distance between the two decision
thresholds, known as the decision boundary, a. If a decision maker is biased towards a particular response then
z will not be equidistant between 0 and a. e non-decision time parameter, T, accounts for basic perceptual
processing, such as recognizing that a choice has been presented, and captures any time required to initiate a
response, such as pressing a button. e valuation conict hypothesis1 focuses on the valuation process, and
therefore predicts that source and destination eects should impact the dri rate parameter v. Here, we tested if
the source of the money impacts the rate at which the valuation of money is accumulated, the rate at which the
valuation of harm is accumulated, or both.
Figure1. Overview of experimental design. Manipulation of source and destination as a function of study
number. In Study 1 participants made moral decisions that varied as a function of the source (dirty vs. clean).
In Study 2 participants made blame judgments about moral decisions that varied as a function of destination
(prot vs. charity). In Study 3 participants made moral decisions that varied as a function of the source and
destination.
Content courtesy of Springer Nature, terms of use apply. Rights reserved

3
Vol.:(0123456789)
Scientic Reports | (2022) 12:18638 | https://doi.org/10.1038/s41598-022-22226-9
www.nature.com/scientificreports/
As the valuation conict framework focuses on the valuation process, that is, the processing of the choice
attributes (i.e., money and harm), it does not make predictions about the bias parameter, z. However, it is pos-
sible that people have a prosocial default17,18, whereby people are generally predisposed against increasing harm
to others. is bias is independent of the specic attributes of the choice. e prosocial default account predicts
that source and destination eects should be captured via the bias parameter z, biasing individuals towards
foregoing dirty money even before they know the amount at stake. While theoretical accounts make a broad
claim that behavior is prosocial as a “default” or “heuristic,” they have not specied how this is implemented at a
computational level in terms of how choices are made. Here, we test whether in a two-alternative forced choice
situation people are predisposed to choose the prosocial option before they receive any information about the
precise attributes of that choice. us, using DDM we can arbitrate between temporally distinct latent processes
in complex decision making; while source eects on starting point, z, suggest a predisposition to select one option
over the other before observing the attributes of the choice, source eects on the dri rate, v, reect valuation-
based processes related to the evaluation of the choice attributes themselves. Consequently, using formal model
comparison, we generate further insight into the predictions of the valuation conict and prosocial default
accounts by investigating the eects of source and destination on model parameters.
Figure2. Moral decision-making task and modeling framework. (a) To probe moral decisions, participants
(known as the “Decider”) made a series of real decisions, where each decision involved choosing between two
options: a smaller amount of money plus a smaller number of painful electric shocks, or a larger amount of
money plus a larger number of shocks. Before observing the choice options, a screen was presented indicating
the recipient of the money and the shocks on the current trial. For half of the trials the shocks were allocated
to the Decider and for the other half the shocks were allocated to an anonymous stranger in the next room. In
study 1, the money recipient was always the Decider. In study 3, for half of the trials the recipient of money was
the Decider and for half the trials the money was donated to a charity (Children with Cancer, UK). (b) Visual
schematic for how the DDM captures choices and response times in Study 1 and Study 3. Responses are coded
as larger shocks on the upper threshold and smaller shocks on lower threshold. Model parameters are discussed
in the main text. (c) Moral decisions involving clean and dirty money (Study 1). e valuation of pain and
money are sensitive to source eects: e weight on shocks was signicantly higher when considering dirty
relative to clean money (purple), while the weight on money was signicantly lower when considering dirty
relative to clean money (green).
Content courtesy of Springer Nature, terms of use apply. Rights reserved

4
Vol:.(1234567890)
Scientic Reports | (2022) 12:18638 | https://doi.org/10.1038/s41598-022-22226-9
www.nature.com/scientificreports/
Study 1
Methods. First, we exploited a previously published dataset9 in which participants decided whether to accept
“clean” money (obtained by inicting painful electric shocks on themselves) or “dirty” money (obtained by
inicting painful electric shocks on another person) for their own personal prot. In each condition (dirty,
clean), participants faced a series of decisions which varied in terms of how many shocks were imposed and how
much money was obtained (Fig.2a). We previously reported source eects on decision making in this dataset:
participants in this experiment preferred taking clean money to dirty money9. Additionally, participants were
slower to accept dirty relative to clean money. ese results are consistent with both accounts of the dirty money
dilemma: valuation conict and prosocial default. In Study 1, we tested the predictions of each model by tting
multi-attribute DDMs using participants’ choices and response times (RT), as detailed below.
Participants. For Study 1, we conducted further analysis of a previously published dataset9 that included 28
participants aged 18–35 (12 female, mean age: 21.90years), recruited from the University College London
(UCL) Psychology Department and the Institute of Cognitive Neuroscience subject pools. All methods were
carried out in accordance with relevant guidelines and regulations. Exclusion criteria included a history of sys-
temic or neurological disorders, psychiatric disorders, psychoactive medication or drug use, pregnancy, more
than two years’ study of psychology, and previous participation in studies involving social interactions and/or
electric shocks. Because this was an fMRI study, we recruited right-handed participants only. Because this was a
previously published dataset, no statistical methods were used to pre-determine sample sizes. e sample size for
the original study was based on estimated eect size for moral preferences observed in two previous behavioral
studies using the same task8.
Moral decision task: source eects. Stimuli were presented using the Cogent toolbox (www. vislab. ucl. ac. uk/
cogent. php) in MATLAB (MathWorks Inc.). We developed a moral decision-making paradigm8,9,19 where par-
ticipants made 204 decisions that involved a trade-o between real monetary rewards and mildly painful electric
shocks (Fig.2a). In Study 1, across trials we manipulated how participants obtained money for prot (source
eects): on half the trials the money was “clean,” obtained by the experimenter inicting painful electric shocks
on oneself. On the other half of trials, the money was “dirty,” obtained by the experimenter inicting painful
electric shocks on another. Before observing the choice options, a screen was presented indicating the recipient
of the shocks on the current trial. is enabled us to test for any a priori bias individuals have towards foregoing
dirty money even before they know the amount of money and shocks at stake.
Participants were instructed to indicate their preference within 6s by pressing a button box with the le or
right index nger. Any trial where the participant did not respond within 6s was recycled and presented at the
end of the task. e exact timing of decisions, the corresponding choice and the dierence between the shock-
and money-oers were used to simulate the valuation-process with the Python-based Hierarchical Dri–Diu-
sion Model toolbox (HDDM20), which will be discussed below.
Participants were instructed that at the end of the experiment one of their decisions would be selected at
random and implemented.
Generation of choice options for the moral decision task. A single trial consisted of two options, each including
a number of shocks (0 to 20) and an amount of money (£0 to £20). One option l was always associated with a
lower number of shocks sl and a lower amount of money ml and the other option h with a higher number of
shocks sh and a higher amount of money mh. Initially, a set of 102 trials were created with each trial tted to the
indierence point of a simulated agent with a specic harm aversion coecient κ. Values of κ were chosen equi-
distant from a uniform distribution covering the whole range of possible κ values (from 0 to 1). For each value
of κ 1,000 random pairs of shock movement ∆s (i.e., the dierence between the higher number of shocks sh and
the lower number of shocks sl) and money movement ∆m were created. e value of ∆s was restricted to positive
integers ranging from 1 to 20 shocks and for ∆m ranging from 0.1 to 19.9 British pound sterling. e pair that
was closest to the indierence point of the respective value of κ was then selected and transformed into a trial.
is was done by giving the lower option a random sl constrained by 0 ≤ sl + ∆s ≤ 20, and a random ml constrained
by 0 ≤ ml + ∆m ≤ 20. e higher options values were then simply given by sh = sl + ∆s and mh = ml + ∆m. To avoid
confounding the eects of the specic trial set with the eect of the dierent conditions, the same set of trials was
used in both conditions (dirty and clean). However, within each condition the trials were counterbalanced with
regards to higher and lower options appearing on the le and right side of the screen. is was intended to make
it more dicult for participants to notice repeating trials and to keep the task more interesting.
Computational framework for valuation conict. Consistent with a growing body of work on value-based
choices we assumed that choices can be captured using a multi-attribute variant of a dri–diusion model
(DDM)15,16,21, in which noisy value signals accumulate over time and a choice is made when the accumulated
signal crosses a predened threshold for choice. In this task, we assume that the relevant attributes of each
option are the relative money (Δm) and relative shocks (Δs). e comparison process stops when a boundary
(a), representing the relative evidence in favor of one of the options, is reached and the corresponding option is
selected. We coded the decision upper and lower boundaries as corresponding to those for choices associated
with the higher and lower amount of harm, respectively. According to a valuation conict framework, the average
dri rate (v) captures the overall speed of the accumulation process and varies from trial to trial with Δm and Δs,
according to dri weights m and s, respectively:
Content courtesy of Springer Nature, terms of use apply. Rights reserved

5
Vol.:(0123456789)
Scientic Reports | (2022) 12:18638 | https://doi.org/10.1038/s41598-022-22226-9
www.nature.com/scientificreports/
e accumulation of evidence starts from a starting point (z), which represents an a priori preference towards
one of the choice options before any trial information is presented (e.g., a bias towards choosing the option with
the smaller shock, regardless of shock magnitude). Models testing for a prosocial default suggests that source
eects should impact the bias parameter z, biasing individuals towards foregoing dirty money even before they
know the amount at stake. e DDM assumes RT is the sum of the duration of the decision time and a non-
decision time component (T). Each of these parameters is depicted in Fig.2b. Additional details on various DDM
specications are provided in the Supplementary Methods.
All DDM estimation was performed using a soware package freely available in Python20. Models were esti-
mated using a Bayesian hierarchical framework, with Markov chain Monte Carlo (MCMC) sampling methods
employed to estimate a joint posterior distribution of the model parameters. Hierarchical Bayesian estimation
allows individual participant estimates to be constrained by a group distribution, but also to vary to the extent
that their data demonstrate separation from the data of others. Estimation used non-informative priors: all priors
were uniform distributions over large intervals of possible parameter values. As outlier RT can cause estimation
issues for the DDM22 we removed all trials with RT faster than 200ms prior to model tting.
We used two measures for model comparison. First, the Deviance Information Criterion (DIC) is a exible
measure for goodness-of-t in hierarchical Bayesian models23. e DIC combines a measure of deviance (i.e.,
lack of t) with a penalty for model complexity, and a lower DIC represents a better model t. Second, we com-
pared observed data to simulated data using the mean of sampled posteriors. For each model, 10,000 simulations
were run with choice and RT data. e squared error between the observed data and the simulated data for each
condition (e.g., shocks for self, shocks for other) was computed for both choices and RT, with the mean squared
error (MSE) computed across conditions. is allowed us to identify how well a tted model can recreate the
observed data24. ese statistics are summarized in Supplementary Results.
Model convergence was assessed using the Gelman-Rubin (G-R) R statistic25. e R statistic compares within-
chain and between-chain variance of dierent runs of the same model. Perfect convergence across runs would
result in an R of 1. e R statistic was computed for all of the model parameters based on ve runs of the model15.
ese statistics are summarized in Supplementary Results.
Bayesian estimation also makes it straightforward to compare dierences in parameter estimates. Signicance
inference is possible using the posterior distributions generated from the MCMC sampling process. To compare
coecients to the null hypothesis (β = 0) we assessed the percentage of the N = 10,000 sampled coecients that
were greater than zero. Similarly, to determine signicant dierences between two regression coecients, we
computed what percentage of samples indicated a dierence that was greater than zero. In our analysis, we are
particularly interested in the ratios of various parameters (e.g., Δm and Δs in various conditions). For these
comparisons, we compared the ratios using the full posterior distribution of each ratio of interest.
Results. To test how source aected decisions to accept money, we computed the proportion of decisions
in each condition where participants chose to accept more money (at the expense of more shocks) and entered
these proportions into a repeated-measures analysis of variance (ANOVA) with source (dirty vs. clean) as
within-subject factors. We found that participants preferred accepting money from a clean rather than a dirty
source when the money was for personal prot (t27 = 2.190, p = 0.037).
To test the predictions of the valuation conict and prosocial default accounts of the dirty money dilemma, we
t multi-attribute DDMs using participants’ choices and response times (RT). e valuation conict hypothesis
predicts that the source of the money will impact the relative weighting of its positive and negative attributes,
such that positive attributes (i.e., money) would outweigh negative attributes (i.e., harm) of decisions to accept
clean money, while the reverse will be observed for dirty money. Meanwhile, the prosocial default hypothesis
predicts that the source of the money will impact the starting point of the valuation process, with a starting point
biased more towards ‘reject’ for dirty relative to clean money.
We formalized these hypotheses with separate models capturing dierent eects of source on model param-
eters. e prosocial default model t separate starting point parameters for dirty and clean money. e family
of valuation conict models t separate dri weights for positive and negative choice attributes for dirty versus
clean money in various combinations (see Methods for details). As a baseline, we also t a naïve model for which
parameters did not vary at all by condition or trial.
We performed formal model comparisons along three dimensions, as outlined above. First, we used the DIC,
a common measure for goodness-of-t in hierarchical Bayesian models23. We also ran simulations using the t-
ted parameters to determine the extent the models could recreate the observed data24, looking at MSE between
true and simulated data for both choice and RT. ese three measures allowed us to triangulate the best tting
model. Additional model comparison details are provided in the Supplementary Materials.
Across all three measures of model t, the prosocial default model performed worse than any of the valuation
conict models (DIC = 20,635.44, Choice MSE = 0.2163, RT MSE = 0.7695). e best-tting model of those we
tested was in the family of valuation conict models, where dri rate is modeled as a multi-attribute combination
of money and shocks, with distinctive weights on these attributes for dirty vs. clean money (DIC = 16,653.22,
Choice MSE = 0.0992, RT MSE = 0.6788). As an additional test of the prosocial default hypothesis, we also modi-
ed the best-performing valuation conict model to include separate starting points for dirty vs. clean money.
is hybrid valuation-default model performed worse than the best-tting valuation conict model on all three
dimensions (DIC = 16,662.66, Choice MSE = 0.0993, RT MSE = 0.6790), and the two starting point parameters
were not signicantly dierent from one another (p = 0.228). Overall, the model comparisons provide decisive
support for the valuation conict hypothesis and no evidence for the prosocial default hypothesis in this dataset.
v
=
βm�m
+
βs�s
Content courtesy of Springer Nature, terms of use apply. Rights reserved

6
Vol:.(1234567890)
Scientic Reports | (2022) 12:18638 | https://doi.org/10.1038/s41598-022-22226-9
www.nature.com/scientificreports/
We next tested a series of theoretical predictions by examining the parameter estimates from our best-tting
valuation conict model. e valuation conict hypothesis posits that decisions about dirty money involve the con-
current activation of conicting impulses toward materiality and morality1. In the context of our model, this predicts
that the source of money (dirty vs. clean) will impact the relative weights on positive and negative choice attributes
(money and shocks, respectively). To test this, we compared the posterior distributions for the weights on money
and shocks across source conditions (dirty relative to clean). Ratios signicantly dierent from 1 indicate the pres-
ence of source eects on each attribute. Here, ratios signicantly lower than 1 indicate that the dri weight for that
attribute was lower when considering dirty relative to clean money, while ratios signicantly greater than 1 indicate
that the dri weight for that attribute was signicantly higher when considering dirty relative to clean money. We
found that the weights on money and shocks were both sensitive to source eects (Fig.2c). e weight on shocks was
signicantly higher when considering dirty relative to clean money (mean ratio = 1.068, p = 0.0181), while the weight
on money was signicantly lower when considering dirty relative to clean money (mean ratio = 0.865, p < 0.001). To
provide another test for the key ratio comparisons, we computed the 95% Highest Density Interval26, to determine
if it overlapped with 1 or not.e HDI of the dirty/clean ratio for money did not overlap with 1 [0.8149,0.9185],
nor did it for shocks [1.0014,1.1317]. is is consistent with the posterior distributions displayed in Fig.2c.
Together, the ndings from Study 1 reveal that source eects on decisions to accept money can be explained
through a shi in the relative valuation of positive and negative choice attributes, consistent with the valuation
conict hypothesis. is suggests that resolving the dilemma of dirty money involves integrating over money’s
positive and negative attributes, rather than adopting a heuristic strategy whereby deciders have an a priori bias
against increasing harm to others and thus are inclined to reject any morally tainted money regardless of its
attributes.
Study 2
Methods. Having found initial support for the valuation conict hypothesis, we next turned to the possibil-
ity that altering the destination of dirty money could “launder” its value27. For example, donating dirty money
to charity (rather than keeping it for oneself) could make initial decisions to accept dirty money more palatable,
perhaps because such decisions might be seen as more morally justied. A prerequisite for testing this hypothesis
is establishing that disinterested observers do in fact judge it to be less blameworthy to accept dirty money on
behalf of a charity than to keep it for oneself. We tested our pre-registered prediction (https:// aspre dicted. org/
blind. php?x= zb65vc) that participants would judge decisions to take dirty money as more blameworthy when
the destination of the money was personal prot relative to charitable donation using a Moral Judgment Task.
Participants. For Study 2, 152 UK university students between the ages of 18 and 35 were recruited from the online
crowdsourcing platform, Prolic (www. proli c. ac), and randomized to either a prot condition or a charity condi-
tion. All participants provided written informed consent prior to participation and were compensated for their
time. e Yale UniversityHuman Investigation Committee approved the procedures, ethics number 2000022385.
All methods were carried out in accordance with relevant guidelines and regulations. Eleven participants rand-
omized to the prot condition and eleven participants randomized to the charity condition were excluded from the
analysis for failing pre-registered attention checks (see below). Final analysis was carried out on the remaining 64
participants in the prot condition (40 female, mean age: 22.48years) and 66 participants in the charity condition
(49 female, mean age: 22.74years). An a priori power analysis indicated that the study required 64 participants in
each condition to have 80 percent power to detect a moderate eect (0.5) in a parametric between-groups analysis.
us, our study was suciently powered to observe an eect in our between-groups design.
Moral judgment task. e moral judgment task implemented in Study 2 asked participants to judge how blame-
worthy they thought it would be for an agent (known as a ‘Decider’) to make each of 42 moral decisions. For each
decision, the Decider chose between a smaller amount of money plus a smaller number of electric shocks for a
stranger (helpful option), or a larger amount of money plus a larger number of electric shocks for the stranger
(harmful option), as shown in Fig.2a. We parametrically modulated the amount of money and shocks in each
trial to produce a series of choices that varied in their relative harmfulness (see Generation of choice options for
the Moral Judgment Task below). Upon seeing each pair of options, participants judged how blameworthy they
thought it would be if a Decider chose the more harmful option (i.e., the larger amount of money and shocks) on
a continuous scale ranging from 0 (extremely praiseworthy) to 100 (extremely blameworthy).
Generation of choice options for the moral judgment task. A single trial consisted of two options, each including
a number of shocks (1 to 20) and an amount of money (£0.10 to £20). One option l was always associated with
a lower number of shocks sl and a lower amount of money ml and the other option h with a higher number of
shocks sh and a higher amount of money mh. A set of 42 trials were created with each trial tted to the indier-
ence point of a simulated agent with a specic price value for shocks. is value was parameterized as a ‘harm
aversion’ coecient κ. Values of κ were chosen equidistant from a uniform distribution covering the whole range
of possible κ values (0 to 1). For each value of κ 1000 random pairs of shock movement ∆s (i.e. the dierence
between the higher number of shocks sh and the lower number of shocks sl) and money movement ∆m were
created. e value of ∆s was restricted to positive integers ranging from 1 to 20 shocks and for ∆m ranging from
0.1 to 19.9 British pound sterling. e pair that was closest to the indierence point of the respective value of κ
was then selected and transformed into a trial. is was done by giving the lower option a random sl constrained
by 1 ≤ sl + ∆s ≤ 20, and a random ml constrained by 0.1 ≤ ml + ∆m ≤ 20. e higher options values were then sim-
ply given by sh = sl + ∆s and mh = ml + ∆m. We ensured that ∆m and ∆s did not covary across trials (Pearson’s
R = 0.004, p = 0.978).
Content courtesy of Springer Nature, terms of use apply. Rights reserved

7
Vol.:(0123456789)
Scientic Reports | (2022) 12:18638 | https://doi.org/10.1038/s41598-022-22226-9
www.nature.com/scientificreports/
Attention check. We included two attention check questions mixed within the 42 moral judgments. For each
attention check, we presented participants with a pair of options, and instead of asking participants to judge
how blameworthy they think it would be if a Decider chose the more harmful option, we asked participants to
respond on the rightmost side of the scale (100 = extremely blameworthy). Participants who did not respond on
the rightmost side of the scale for either of these questions were excluded from the analysis.
Manipulation. To investigate whether moral judgments were sensitive to the destination of the dirty money
participants were randomly assigned to one of two conditions: a prot condition, where for each decision the
Decider receives the money, or a charity condition, where for each decision the money is donated to Children
with Cancer UK, a leading charity organization that helps provide better care for children with cancer and their
families, and supports research into causes, prevention and treatment of childhood cancer. Accordingly, partici-
pants in the prot condition judged harmful decisions that benet the Decider personally, while participants in
the charity condition judged harmful decisions that benet a charity.
Analysis. All data analysis was completed in Matlab (Mathworks) and all statistical tests were two-sided. We
hypothesize that participants in the prot condition will assign greater blame for harmful decisions than partici-
pants in the charity condition. To test our primary preregistered hypothesis that people assign greater blame for
ill-gotten harms than harms for the greater good, we carried out a between-subjects parametric test (independent
samples t-test) on subjects’ average blame rating across trials. As secondary preregistered analyses, we used a lin-
ear mixed eects model with a random intercept to model how money and shocks in each choice option predict
blame judgments, and whether the eects of money and shocks on blame judgments dier as a function of the
money recipient. e regressors in the model included: the dierence between the chosen and unchosen amount of
money (Δm, β1), the dierence between the chosen and unchosen amount of shocks (Δs, β2), the money recipient
(‘condition’, dummy coding for prot versus charity, β3), the interaction between condition and Δm (β4), and the
interaction between condition and Δs (β5), and a random intercept for participants (u0i). Our regression included
a xed intercept term (β0), capturing the average judgment across trials, and all other coecients expressed mean
deviations from this judgment. Parameters were estimated using maximum likelihood estimation.
We report means and standard error of the mean (sem) as mean ± sem. Eect sizes were computed for sig-
nicant results using Cohen’s d.
Results. We t a linear mixed-eects model to describe how blame judgments varied as a function of the
amount of harm imposed, the amount of money obtained, and the destination of the money (see Methods for
details and Supplementary Table4 for results of the full model). Consistent with past work (Siegel, Crockett,
& Dolan, 2017), across both conditions, blame judgments were positively correlated with the amount of harm
(β = 1.697 ± 0.081, t = 20.894, p < 0.001) and negatively correlated with the amount of money (β = -1.879 ± 0.078,
t = -24.211, p < 0.001), Fig. 3a. As we predicted, participants in the prot condition assigned signicantly
more blame overall than participants in the charity condition (mean ± sem, prot: 62.842 ± 0.215; charity:
49.543 ± 0.204; t = 5.563, p < 0.001, d = 0.976). To illustrate these eects, we t the linear mixed-eects model
separately for the prot and charity condition and plotted the estimated blame judgments as a function of money
and shocks in Fig.3b. e heatmap for Prot (le) demonstrates an increase in blame relative to the Charity
(right) condition. To summarize, study 2 demonstrates that blame for accepting dirty money is sensitive to the
destination of the money, where blame is mitigated when dirty money is directed toward a charitable cause.
Study 3
Following the results of Study 1, which supported the valuation conict account, we predicted that participants
would prefer to accept clean over dirty money for their own personal prot, but that source eects would be
reduced when the money was destined for a charity. Furthermore, we expected that both source and destination
would impact the relative weighting of the positive and negative attributes of money. Specically, we expected
to replicate the source eects observed in Study 1, such that decisions about dirty relative to clean money would
involve higher weights on shocks and lower weights on money. Following the results of Study 2, we expected that
directing dirty money toward charity would reduce the impact of negative attributes on the valuation of money.
Methods. Study 1 provided initial support for the valuation conict hypothesis, demonstrating that deci-
sions about whether to accept dirty (versus clean) money involve integrating across the money’s positive and
negative attributes. Study 2 showed that accepting dirty money is less blameworthy when the money is directed
to charity than when it is kept for personal gain. In Study 3, we build on these ndings to further test the predic-
tions of the valuation conict and prosocial default accounts of the dirty money dilemma. e valuation conict
account, but not the prosocial default account, predicts that directing dirty money to charity can eectively
“launder” its value. Consequently, we adapted the Moral Decision Task implemented in Study 1 to test the eects
of both source (dirty, clean) and destination (prot, charity) on moral decision making.
Participants. For Study 3, healthy participants aged 18–35 were recruited from the University of Oxford’s
Department of Psychology subject pool. Participants with a history of neurological or neuropsychiatric disor-
ders, pregnant women, and more than one year of study in psychology were excluded from participation. Partici-
pants who had previously participated in studies involving deception of electric shocks were also excluded due
Blame
=
(β0
+
u
0
i)
+
β
1(
�m
)+
β
2(
�s
)+
β
3(
condition
)+
β
4(
condition
∗
�m
)+
β
5(
condition
∗
�s)
Content courtesy of Springer Nature, terms of use apply. Rights reserved

8
Vol:.(1234567890)
Scientic Reports | (2022) 12:18638 | https://doi.org/10.1038/s41598-022-22226-9
www.nature.com/scientificreports/
to concerns that prior experience with being deceived would inuence belief in the outcomes of the current task,
which did not involve deception. All participants provided written informed consent and were compensated for
their time. Study 3 was approved by the University of Oxford ethics committee (R50262/RE001). All methods
were carried out in accordance with relevant guidelines and regulations.
Participants in Study 3 attended a single testing session at the University of Oxford’s Department of Experi-
mental Psychology. Sixty-two pairs of participants took part in the study. e two participants arrived at stag-
gered times and were led to separate testing rooms without seeing one another to maintain anonymity. Aer
providing informed consent, a pain thresholding procedure was used to familiarize participants with the electric
shock stimuli prior to the task. Next, participants were randomly assigned to either the role of the ‘Decider’ or
the ‘Receiver’. Receiver participants completed a moral learning task (data to be reported separately) and decider
participants completed a moral decision task, which we focus on here. One decider participant did not com-
plete the task due to a technical error, and therefore was excluded from the analysis. is le a total of sixty-one
participants in the role of the decider whose data were analyzed (25 female, mean age: 23.95years). A power
analysis indicated that the study required 34 participants in each condition to have 80 percent power to detect
a moderate eect (0.5) in a parametric within-groups analysis of dirty versus clean money. us, our study was
suciently powered to observe source eects when money is destined for prot and for charity.
Moral decision task: source and destination eects. Stimuli were presented using the Cogent toolbox (www.
vislab. ucl. ac. uk/ cogent. php) in MATLAB (MathWorks Inc.). In Study 3, participants made 176 decisions in the
Moral Decision Task (Fig.2a). As in Study 1, across trials we manipulated how participants obtained money
for prot (source eects): on half the trials the money was “clean,” obtained by the experimenter inicting pain-
ful electric shocks on oneself. On the other half of trials, the money was “dirty,” obtained by the experimenter
inicting painful electric shocks on another person. In this study, we additionally manipulated who obtained the
money across trial (destination eects): on half of the trials the money was for the decider (‘prot’ condition)
and on the other half of trials the money was donated to a charity organization for children with cancer (‘char-
ity’ condition). is resulted in a 2 × 2 factorial design in Study 3, with source (dirty vs. clean) and destination
(prot vs. charity) as the independent variables. Closely matching the instructions given in Experiment 1, before
observing the choice options, a screen was presented indicating the recipient of the shocks (self, other) and
money (prot, charity) on the current trial. is allowed us to test for a tendency towards harming for charity
before participants knew what amount of money and shocks were at stake.
As in Study 1, participants were instructed that at the end of the task, one of their decisions would be selected
at random and implemented at the end of the experiment. is allowed us to ensure that any eort exerted in
the process of inicting shocks did not aect the valuation. Information about the charity was derived from its
website (www. c hild r enwi thcan cer. o r g. uk) and presented to the participants in the following form: “e decisions
you make during [charity] trials aects how much money we donate to Children with Cancer UK. is organiza-
tion is a leading charity in the UK that helps provide better care for children with cancer and their families, and
supports research into causes, prevention and treatment of childhood cancer in the UK.”
Trials were generated using an identical procedure to those outlined in Study 1. However, in Study 3, the
trials from each condition were randomly sorted into sections of 10 trials. One section from each of the four
conditions (see Fig.1) were combined to form a block of 40 trials. ese blocks were then randomly presented
one aer another with no block starting with the same condition that the previous block ended with.
Figure3. Moral judgment task and destination eects on blame. (a) In Study 2 participants rated how
blameworthy they thought it would be if a person chose the more harmful option (i.e., the larger amount of
money and shocks, highlighted in red) on a continuous scale ranging from 0 (extremely praiseworthy) to 100
(extremely blameworthy). (b) Estimated blame judgments from a linear mixed-eects model as a function
of money and shocks. Blame for accepting dirty money is sensitive to the destination of the money. Dashed
diagonal lines indicate mix of money and shocks that are neither praiseworthy nor blameworthy.
Content courtesy of Springer Nature, terms of use apply. Rights reserved

9
Vol.:(0123456789)
Scientic Reports | (2022) 12:18638 | https://doi.org/10.1038/s41598-022-22226-9
www.nature.com/scientificreports/
Computational framework for valuation conict. In addition to allowing the dri weights v to vary as a function
of source, Study 3 t models that allowed v to vary as a function of destination. Additional details on various
DDM specications are provided in the Supplementary Methods.
Results. To test how source and destination aected decisions to accept money, we computed the proportion
of decisions in each condition where participants chose to accept more money (at the expense of more shocks)
and entered these proportions into a repeated-measures analysis of variance (ANOVA) with source (dirty vs.
clean) and destination (prot vs. charity) as within-subject factors. We found that harm was mitigated when
the money was destined for a charitable cause versus for prot (eect of destination: F1, 60 = 11.627, p = 0.001).
Decisions to harm were not signicantly aected by the source of the money (eect of source: F1, 60 = 0.0835,
p = 0.365). However, as predicted, we observed a signicant interaction between source and destination (inter-
action: F1, 60 = 30.929, p < 0.001). Simple eects analyses indicated that participants preferred accepting money
from a clean rather than a dirty source when the money was for personal prot (t60 = 2.850, p = 0.006), but not
when the money was for charity (t60 = − 0.851, p = 0.398, Fig.4a. Furthermore, participants preferred accepting
larger amounts of money/shocks for prot relative to for charity only when the money was clean (t60 = 4.8637,
p < 0.001), but not when the money was dirty (t60 = 1.216, p < 0.229). is suggests that people may be particularly
averse to their own harm when it benets another, as opposed to themselves.
Again, we t a series of DDM specications to assess the cognitive computations employed when deciding
whether to accept dirty money. e prosocial default model, as in Study 1, allowed the starting point parameter
to vary across source conditions. e family of valuation conict models t separate dri weights for positive and
negative choice attributes across source and destination conditions. Consistent with Study 1, we found evidence
in support of the valuation conict hypothesis. On all three model comparison indices, the valuation conict
model performed better than any other model (DIC = 35,535.33, Choice MSE = 0.2499, RT MSE = 0.6355). e
best-tting model was a valuation conict model that allowed the dri rate to vary in terms of both source and
destination eects (DIC = 28,023.20, Choice MSE = 0.1170, RT MSE = 0.5900). is model had separate param-
eters for dri weights on money and shocks in each of the four experimental conditions. As an additional test of
the prosocial default account, we modied the best-tting valuation conict model to include separate starting
points for source eects. As in Study 1, this hybrid valuation-default model performed worse on all three dimen-
sions (DIC = 28,091.39, Choice MSE = 0.1171, RT MSE = 0.5910).
We next examined the parameter estimates from the best tting model to test the eects of source and destina-
tion on the relative weighting of positive and negative attributes in decision making. We rst compared param-
eter estimates for dirty versus clean money separately for each destination condition (Fig.4b,c). In the prot
condition, as in Study 1, the weight on money was lower for dirty relative to clean money (mean ratio = 0.735,
p < 0.001 compared to ratio of 1). In contrast to Study 1 however, there was no evidence for source eects on
the weighting of shocks (mean ratio = 1.056, p = 0.189). Moreover, the dierence in the size of these eects was
signicant, such that source impacted the weighting of money to a signicantly greater degree than the weighting
of shocks (p = 0.011). Conversely, in the charity condition, no source eects were observed on money weights for
dirty relative to clean (mean ratio = 1.006, p = 0.457). However, source impacted the weight on shocks, which was
lower when accepting dirty, relative to clean, money for charity (mean ratio = 0.919, p = 0.036). Collectively, these
ndings demonstrate that considering whether to accept dirty money for prot is associated with a diminished
valuation of money. However, when dirty money is directed toward a charitable cause (i.e., “laundered”), this
reduction in value is no longer observed. Quantitatively, the impact of source on the weighting of money was
approximately four times larger in the prot condition than the charity condition.
To provide another test for the key ratio comparisons, we computed the 95% HDI26, to determine if it over-
lapped with 1 or not.In the prot condition, the 95% HDI for the dirty/clean ratio of money did not overlap with
1 [0.6701, 0.7977], whereas for shocks it did [0.9358,1.1838]. is result is consistent with the above ndings
indicating a signicant eect of source on money but not shocks in the prot condition. For the charity condi-
tion, the 95% HDI for money did overlap with 1 [0.9140,1.1012], which is consistent with our previous ndings
suggesting no eects of source on the weighting of money in the charity condition. However, we nd a weaker
result for shocks, as the 95% HDI slightly overlapped with 1 [0.8360, 1.0076].
Finally, as an out-of-sample check on the validity of the latent model parameters, we tested if there was a
correlation between the blame judgments collected in Study 2 and the estimated value of accepting dirty money
in Study 3. As noted in Study 2, participants judged it to be more blameworthy to accept dirty money for prot
compared to charity. Here, we compared this gap in blame judgments between the prot and charity conditions to
the “temptation” of selecting the option with more money in Study 3. “Temptation” was computed as the weighted
sum of the dierence in money and the dierence in shocks, using the estimated DDM dri weights from the
best-tting model. To compute the gap in blame judgments between prot and charity conditions we computed
the dierence in the estimated blame for each possible combination of shocks and money, given the extracted
beta weights from our linear mixed-eects models that were t separately for the two conditions in Study 2.
where:
Gap in blame judgement
=
Blameprofit
−
Blamecharity
Blameprofit
=
β
1
profit (�m)
+
β
2
profit (�s)
Blamecharity
=
β
1
charity(�m)
+
β
2
charity(�s)
Content courtesy of Springer Nature, terms of use apply. Rights reserved

10
Vol:.(1234567890)
Scientic Reports | (2022) 12:18638 | https://doi.org/10.1038/s41598-022-22226-9
www.nature.com/scientificreports/
Each possible combination of money and shocks is plotted in Fig.4d. We observed a signicant positive
correlation between “temptation” and Prot-Charity blame (Pearson r = 0.520, p < 0.001). Furthermore, this
positive correlation was tightly coupled to the underlying dierence in shocks (Pearson r = 0.544, p < 0.001), as
shown in the color coding of the dots in Fig.4d. No such relationship existed with the amount of money (Pearson
r = 0.100, p = 0.528, see corresponding plot in Supplementary Results). is demonstrates a proof of principle that
the attribute weights assigned to available options predict both moral decisions and RT in the decision-making
task, as well as moral judgments in the blame task28.
Discussion
Folk intuitions and empirical evidence show that people are generally disinclined to accept dirty money, but
sometimes do so anyway7,8. However, it remains unclear how people resolve the dilemma of dirty money. e
valuation conict hypothesis suggests people trade o the moral and material costs and benets of dirty money1.
By analyzing patterns of decisions and response times with multi-attribute dri–diusion models, we nd evi-
dence to support this hypothesis. Participants in our experiments were more likely to accept dirty money when it
was directed toward a charitable cause, suggesting that moral destinations can “launder” dirty money. Response
time data showed that the source and destination of money impacted the accumulation of value during the deci-
sion process (as predicted by the valuation conict account), but did not inuence the starting point of value
Figure4. Source and destination eects on moral decision making. (a) Proportion of harmful choices made
as a function of the source (dirty/clean) and destination (prot/charity) of the money. (b) In line with Study 1
(Fig.2c) the weight on money was signicantly lower when considering dirty relative to clean money for prot.
(c) e weight on shocks was signicantly lower when considering dirty relative to clean money for charity. (d)
Scatterplot showing out-of-sample correlation between dierences in blame judgments (Prot-Charity) in Study
2 (horizontal axis) and the estimated dierences in value associated with choosing the more harmful option
(Prot-Charity) in Study 3 (vertical axis). Each dot represents a possible combination ofmoney andshocks
oers. Dots are color coded according to the highest/lowestshock oers. Error bars represent standard error of
the mean.
Content courtesy of Springer Nature, terms of use apply. Rights reserved

11
Vol.:(0123456789)
Scientic Reports | (2022) 12:18638 | https://doi.org/10.1038/s41598-022-22226-9
www.nature.com/scientificreports/
accumulation or the decision boundary. Together these ndings suggest that individuals resolve the dilemma of
dirty money by integrating over money’s positive and negative attributes.
In addition to providing empirical support for the valuation conict framework, our ndings further suggest
that resolution of the conict is accomplished primarily via changing the valuation of positive rather than nega-
tive attributes. Across Studies 1 & 3 we found consistent evidence that the source of money impacts the weight of
positive attributes in the valuation process, but inconsistent evidence that source impacts the weight of negative
attributes. ese ndings are consistent with past neuroimaging results9, which suggest that the decision to forego
ill-gotten gains is driven by decreased activation in prot-sensitive brain regions (dorsal striatum) rather than
increased activation in pain-sensitive brain regions (insula and anterior cingulate cortex). Together, the results
enrich previous work in non-social decision making9,29,30 by further demonstrating that moral decision making
shares the principles of information integration found in many other domains16.
Our results also reveal how dirty money can be “laundered” when it is directed towards a noble cause. Accept-
ing dirty money that is meant for a charity is judged to be less blameworthy than using such money for personal
gain (Study 2; Fig.3b), and dirty money is devalued less when the money is for charity relative to when it is for
prot (Study 3; Fig.4b,c). ese ndings are consistent with previous studies showing that individuals will spend
dirty money on more virtuous items as opposed to more hedonic ones3. Indeed, moral preferences are malleable
to contextual changes that justify antisocial behavior: for example, people oen donate less to charity if stinginess
does not harm their reputation31,32. Our modeling work shows a plausible decision process that leads to such
behaviors: changing the moral consequences of decisions alters the integration of distinct attributes10. Looking
across studies, our ndings provide evidence for the impact of third-party moral judgments on the formation
of independently made moral decisions.
By studying moral judgment and moral decision making within the same experimental framework, we were
able to explicitly link these cognitive processes28 to examine how people resolve the dilemma of dirty money. e
valuation conict account suggests that strengthening the positive attributes of dirty money can make people
more likely to accept it. Here we show that making those positive attributes moral – for example, by directing
the money towards a charity rather than one’s own prot – changes the way that money is valued. is process
may operate via changing people’s beliefs about how third parties evaluate decisions to accept dirty money.
Supporting this idea, we found that dri rates derived from our computational model of the decision-making
process made accurate out-of-sample predictions regarding third parties’ blame judgments about decisions to
accept dirty money (Fig.4d). ese ndings suggest that people might simulate how observers would judge their
decisions about whether to accept dirty money, and adjust their own decisions in tune with the estimated moral
consequences of their decisions33. More broadly, this perspective suggests that integrative appraisals may be
involved in understanding one’s own preferences (‘thinking about thinking’) and aect the subjective decision-
making process accordingly34.
Our ndings also align with a recent study by Qu and colleagues14, who asked participants to decide whether
to accept money for themselves or a charity, at the expense of a moral cost (proting a morally bad cause). ey
found that participants were more willing to accept ill-gotten gains for themselves than for a charity, which
was driven by a diminished valuation of the negative attributes (i.e., moral costs) when deciding for their own
interests. Similarly, Study 3 demonstrates that participants are also more willing to accept negative outcomes
when deciding for their own interests, versus the interests of a charity. However, this was primarily associated
with a diminished valuation of the decision’s positive attributes (i.e., money) when they came at a moral cost
(i.e., shocks for another), as opposed to the decision’s negative attributes. Several dierences across studies might
explain this discrepancy, such as the stimuli used (shocks vs. proting a morally bad cause) and the number of
experimental manipulations (source and destination vs. destination alone). Future research is needed to elucidate
the discrepancies between studies and determine under what conditions a willingness to accept ill-gotten gains
for one’s own interest, over another’s, is driven by increased valuation of the decision’s positive attributes, versus
diminished valuation of the decisions negative attributes.
One important limitation of our design is that our experimental context, trading money for electric shocks, is
unlikely to be encountered in real life. Notably, while electric shocks are aversive, the stimulation was relatively
mild and monetary rewards were small. Whether individuals resolve the dilemma of dirty money with similar
cognitive computations in more familiar, real-world situations, e.g., stolen money or drug money, or when more
extreme harms and money are at stake, is unknown. Past work has shown similar devaluation eects of dirty
money in vignettes using other forms of dirty money, such as stolen money7, which suggests there might be a
common mechanism in more familiar contexts. Additionally, our design requires participants to impose some
amount of harm regardless of their decision. e cognitive computations underlying dirty money decisions may
be dierent when deciding between inicting some amount of harm for money versus no harm at all, i.e., zero
shocks35. Future work should investigate whether an option where harm can be avoided entirely changes the
decision dynamics as measured with the DDM.
Although our computational models t latent parameters at the individual level, our hypothesis tests focus
on the group-level eects of source and destination. Using these group-level parameters, we found that resolv-
ing the conict of dirty money operates primarily via changing the valuation of positive rather than negative
attributes. However, these group level ndings may obscure systematic individual dierences. For instance, some
individuals may resolve the conict via a response bias against increasing harm to others, consistent with the
prosocial default account. Such individual dierences may relate to other aspects of moral cognition, such as
deontological versus utilitarian preferences. ose who rely on more rule-based strategies, such as those applied
in deontology, may be more likely to resolve the conict by adopting a prosocial default, while those who adhere
to utilitarian preferences may be more likely to weigh up the positive and negative attributes of the conict. Future
work could use a combination of electroencephalogram (EEG) recordings and the DDM36, to obtain the temporal
resolution required to infer if certain aspects of the moral choice are considered earlier in the valuation process.
Content courtesy of Springer Nature, terms of use apply. Rights reserved

12
Vol:.(1234567890)
Scientic Reports | (2022) 12:18638 | https://doi.org/10.1038/s41598-022-22226-9
www.nature.com/scientificreports/
Our paradigm taps into how people make decisions about money associated with harm to others. is reects
a shared aspect of many real-world cases of dirty money, for example, institutions that had to decide whether
to accept money from the Sackler family associated with the misfortune of others during the opioid crisis. at
said, there are many dierences between our paradigm and real-world examples of “dirty money.” For instance,
our paradigm may additionally tap into preferences for money that feels “earned” relative to money that feels
“unearned.” Future research could tease apart these possibilities by comparing behavior in the present paradigm
with one where participants make decisions about money they earn themselves (e.g., by exerting eort37), versus
money earned through others’ eort.
Many transactions in our economy involve morally tainted goods. Our ndings suggest multiple routes to
target in the decision-making process with the ultimate goal of inducing behavioral change. For example, past
work on value-based decision-making shows that attention and the valuation of choice attributes are tightly
linked21,35,36,38,39. In the context of the current study, this suggests that drawing individuals’ attention to the harms
of dirty money might reduce the likelihood of its acceptance. Similarly, we speculate that appraisal processes
could play a role in guiding decisions involving dirty money. Narratives that highlight the potential positive
impact dirty money can have on others (e.g., ill-gotten charitable contributions) may be especially enticing40.
erefore, developing interventions to shi individuals’ appraisals of dirty money from “good” to “bad” may
increase rejections of dirty money.
In sum, we provide a computational characterization of cognitive processes that drive moral decisions involv-
ing dirty money. Our account of how these decisions unfold helps unpack an issue oen discussed aer a choice
has been made. Understanding how dirty money can be psychologically “laundered” through charitable activities
is an initial step towards the development of interventions that can help prevent the cycle of corrupt behavior.
Data availability
Studies 1 and 3 were not formally preregistered. All methods and analyses were preregistered for Study 2 and can
be accessed at https:// aspre dicted. org/ blind. php?x= zb65vc.All data, materials, and analysis code are available
via OSF https:// osf. io/ qg8m7/.
Received: 26 July 2022; Accepted: 11 October 2022
References
1. Tasimi, A. & Gross, J. J. e dilemma of dirty money. Curr Dir Psychol Sci 29, 41–46 (2020).
2. Huang, J. Y., Ackerman, J. M. & Newman, G. E. Catching (up with) magical contagion: A review of contagion eects in consumer
contexts. J. Assoc. Consum. Res. 2, 430–443 (2017).
3. Levav, J. & Mcgraw, A. P. Emotional accounting: How feelings about money inuence consumer choice. J. Mark. Res. https:// doi.
org/ 10. 1509/ jmkr. 46.1. 66 (2018).
4. Stellar, J. E. & Willer, R. e corruption of value: Negative moral associations diminish the value of money. Soc. Psychol. Personal.
Sci. 5, 60–66 (2014).
5. Liu, J., Liao, C., Lu, J., Luo, Y. & Cui, F. Moral contagion: Devaluation eect of immorality on hypothetical judgments of economic
value. Hum. Brain Mapp. 40, 2076–2088 (2019).
6. Newman, G. E. & Bloom, P. Physical contact inuences how much people pay at celebrity auctions. PNAS 111, 3705–3708 (2014).
7. Tasimi, A. & Gelman, S. A. Dirty money: e role of moral history in economic judgments. Cognit. Sci. 41, 523–544 (2017).
8. Crockett, M. J., Kurth-Nelson, Z., Siegel, J. Z., Dayan, P. & Dolan, R. J. Harm to others outweighs harm to self in moral decision
making. PNAS 111, 17320–17325 (2014).
9. Crockett, M. J., Siegel, J. Z., Kurth-Nelson, Z., Dayan, P. & Dolan, R. J. Moral transgressions corrupt neural representations of
value. Nat. Neurosci. 20, 879–885 (2017).
10. Qu, C., Météreau, E., Butera, L., Villeval, M. C. & Dreher, J.-C. Neurocomputational mechanisms at play when weighing concerns
for extrinsic rewards, moral values, and social image. PLoS Biol. 17, e3000283 (2019).
11. Mazar, N., Amir, O. & Ariely, D. e dishonesty of honest people: A theory of self-concept maintenance. J. Mark. Res. 45, 633–644
(2008).
12. Erat, S. & Gneezy, U. White lies. Manage. Sci. 58, 723–733 (2012).
13. Yin, L., Hu, Y., Dynowski, D., Li, J. & Weber, B. e good lies: Altruistic goals modulate processing of deception in the anterior
insula. Hum. Brain Mapp. 38, 3675–3690 (2017).
14. Qu, C., Hu, Y., Tang, Z., Derrington, E. & Dreher, J.-C. Neurocomputational mechanisms underlying immoral decisions beneting
self or others. Soc. Cognit. Aect. Neurosci. 15, 135–149 (2020).
15. Harris, A., Clithero, J. A. & Hutcherson, C. A. Accounting for taste: A multi-attribute neurocomputational model explains the
neural dynamics of choices for self and others. J. Neurosci. 38, 7952–7968 (2018).
16. Ratcli, R., Smith, P. L., Brown, S. D. & McKoon, G. Diusion decision model: Current issues and history. Trends Cogn. Sci. 20,
260–281 (2016).
17. Rand, D. G. Cooperation, fast and slow: Meta-analytic evidence for a theory of social heuristics and self-interested deliberation.
Psychol. Sci. 27, 1192–1206 (2016).
18. Zaki, J. & Mitchell, J. P. Intuitive prosociality. Curr. Dir. Psychol. Sci. 22, 466–470 (2013).
19. Crockett, M. J. et al. Dissociable eects of serotonin and dopamine on the valuation of harm in moral decision making. Curr. Biol.
25, 1852–1859 (2015).
20. Wiecki, T. V., Sofer, I. & Frank, M. J. HDDM: Hierarchical Bayesian estimation of the Dri-Diusion Model in Python. Front.
Neuroinform. 7 (2013).
21. Tusche, A. & Hutcherson, C. A. Cognitive regulation alters social and dietary choice by changing attribute representations in
domain-general and domain-specic brain circuits. Elife 7, e31185 (2018).
22. Ratcli, R. & Tuerlinckx, F. Estimating parameters of the diusion model: Approaches to dealing with contaminant reaction times
and parameter variability. Psychon. Bull. Rev. 9, 438–481 (2002).
23. Spiegelhalter, D. J., Best, N. G., Carlin, B. P. & Linde, A. V. D. Bayesian measures of model complexity and t. J. R. Stat. Soc. Ser. B
(Stat. Methodol.) 64, 583–639 (2002).
24. Palminteri, S., Wyart, V. & Koechlin, E. e importance of falsication in computational cognitive modeling. Trends Cognit. Sci.
21, 425–433 (2017).
25. Gelman, A. & Rubin, D. B. Inference from iterative simulation using multiple sequences. Stat. Sci. 7, 457–472 (1992).
Content courtesy of Springer Nature, terms of use apply. Rights reserved

13
Vol.:(0123456789)
Scientic Reports | (2022) 12:18638 | https://doi.org/10.1038/s41598-022-22226-9
www.nature.com/scientificreports/
26. Kruschke, J. K. Bayesian estimation supersedes the t test. J. Exp. Psychol. Gen. 142, 573–603 (2013).
27. Lobel, T. E. et al. Being clean and acting dirty: e paradoxical eect of self-cleansing. Ethics Behav. 25, 307–313 (2015).
28. Yu, H., Siegel, J. Z. & Crockett, M. J. Modeling morality in 3-D: decision-making, judgment, and inference. Top. Cogn. Sci. 11,
409–432 (2019).
29. Hutcherson, C. A., Montaser-Kouhsari, L., Woodward, J. & Rangel, A. Emotional and utilitarian appraisals of moral dilemmas are
encoded in separate areas and integrated in ventromedial prefrontal cortex. J. Neurosci. 35, 12593–12605 (2015).
30. Shenhav, A. & Greene, J. D. Moral judgments recruit domain-general valuation mechanisms to integrate representations of prob-
ability and magnitude. Neuron 67, 667–677 (2010).
31. Andreoni, J. Impure altruism and donations to public goods: A theory of warm-glow giving. Econ. J. 100, 464–477 (1990).
32. Dana, J., Weber, R. A. & Kuang, J. X. Exploiting moral wiggle room: experiments demonstrating an illusory preference for fairness.
Econ. eor. 33, 67–80 (2007).
33. Volz, L. J., Welborn, B. L., Gobel, M. S., Gazzaniga, M. S. & Graon, S. T. Harm to self outweighs benet to others in moral decision
making. PNAS 114, 7963–7968 (2017).
34. van der Plas, E., David, A. S. & Fleming, S. M. Advice-taking as a bridge between decision neuroscience and mental capacity. Int.
J. Law Psychiatry 67, 101504 (2019).
35. Berman, J. Z. & Kupor, D. Moral choice when harming is unavoidable. Psychol Sci 31, 1294–1301 (2020).
36. van Vugt, M., Simen, P., Nystrom, L., Holmes, P. & Cohen, J. EEG Oscillations Reveal Neural Correlates of Evidence Accumulation.
Frontiers Neurosci. 6 (2012).
37. Lockwood, P. L. et al. Prosocial apathy for helping others when eort is required. Nat Hum Behav 1, 1–10 (2017).
38. Hare, T. A., Malmaud, J. & Rangel, A. Focusing attention on the health aspects of foods changes value signals in vmPFC and
improves dietary choice. J. Neurosci. 31, 11077–11087 (2011).
39. Smith, S. M. & Krajbich, I. Attention and choice across domains. J. Exp. Psychol. Gen. 147, 1810 (2018).
40. Bénabou, R., Falk, A. & Tirole, J. Narratives, Imperatives, and Moral Reasoning. https:// www. nber. org/ papers/ w24798 (2018).
https:// doi. org/ 10. 3386/ w24798.
Acknowledgements
is work was supported by the John Templeton Foundation Beacons Project and grant #61495, a Wellcome
Trust ISSF award (204826/Z/16/Z), the John Fell Fund and the Academy of Medical Sciences (SBF001\1008).
J.Z.S. was supported by a Clarendon and Wellcome Trust Society and Ethics award (104980/Z/14/Z). e funders
had no role in study design, data collection and analysis, decision to publish or preparation of the manuscript.
Author contributions
M.J.C., J.Z.S. and F.H. designed the research, J.Z.S. and F.H. performed the research, J.Z.S., E.V.D.P., F.H., and
J.A.C. analyzed the data, J.Z.S., E.V.D.P., F.H., J.A.C., and M.J.C. wrote the paper.
Additional information
Supplementary Information e online version contains supplementary material available at https:// doi. org/
10. 1038/ s41598- 022- 22226-9.
Correspondence and requests for materials should be addressed to M.J.C.
Reprints and permissions information is available at www.nature.com/reprints.
Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and
institutional aliations.
Open Access is article is licensed under a Creative Commons Attribution 4.0 International
License, which permits use, sharing, adaptation, distribution and reproduction in any medium or
format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the
Creative Commons licence, and indicate if changes were made. e images or other third party material in this
article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the
material. If material is not included in the article’s Creative Commons licence and your intended use is not
permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from
the copyright holder. To view a copy of this licence, visit http:// creat iveco mmons. org/ licen ses/ by/4. 0/.
© e Author(s) 2022
Content courtesy of Springer Nature, terms of use apply. Rights reserved

1.
2.
3.
4.
5.
6.
Terms and Conditions
Springer Nature journal content, brought to you courtesy of Springer Nature Customer Service Center GmbH (“Springer Nature”).
Springer Nature supports a reasonable amount of sharing of research papers by authors, subscribers and authorised users (“Users”), for small-
scale personal, non-commercial use provided that all copyright, trade and service marks and other proprietary notices are maintained. By
accessing, sharing, receiving or otherwise using the Springer Nature journal content you agree to these terms of use (“Terms”). For these
purposes, Springer Nature considers academic use (by researchers and students) to be non-commercial.
These Terms are supplementary and will apply in addition to any applicable website terms and conditions, a relevant site licence or a personal
subscription. These Terms will prevail over any conflict or ambiguity with regards to the relevant terms, a site licence or a personal subscription
(to the extent of the conflict or ambiguity only). For Creative Commons-licensed articles, the terms of the Creative Commons license used will
apply.
We collect and use personal data to provide access to the Springer Nature journal content. We may also use these personal data internally within
ResearchGate and Springer Nature and as agreed share it, in an anonymised way, for purposes of tracking, analysis and reporting. We will not
otherwise disclose your personal data outside the ResearchGate or the Springer Nature group of companies unless we have your permission as
detailed in the Privacy Policy.
While Users may use the Springer Nature journal content for small scale, personal non-commercial use, it is important to note that Users may
not:
use such content for the purpose of providing other users with access on a regular or large scale basis or as a means to circumvent access
control;
use such content where to do so would be considered a criminal or statutory offence in any jurisdiction, or gives rise to civil liability, or is
otherwise unlawful;
falsely or misleadingly imply or suggest endorsement, approval , sponsorship, or association unless explicitly agreed to by Springer Nature in
writing;
use bots or other automated methods to access the content or redirect messages
override any security feature or exclusionary protocol; or
share the content in order to create substitute for Springer Nature products or services or a systematic database of Springer Nature journal
content.
In line with the restriction against commercial use, Springer Nature does not permit the creation of a product or service that creates revenue,
royalties, rent or income from our content or its inclusion as part of a paid for service or for other commercial gain. Springer Nature journal
content cannot be used for inter-library loans and librarians may not upload Springer Nature journal content on a large scale into their, or any
other, institutional repository.
These terms of use are reviewed regularly and may be amended at any time. Springer Nature is not obligated to publish any information or
content on this website and may remove it or features or functionality at our sole discretion, at any time with or without notice. Springer Nature
may revoke this licence to you at any time and remove access to any copies of the Springer Nature journal content which have been saved.
To the fullest extent permitted by law, Springer Nature makes no warranties, representations or guarantees to Users, either express or implied
with respect to the Springer nature journal content and all parties disclaim and waive any implied warranties or warranties imposed by law,
including merchantability or fitness for any particular purpose.
Please note that these rights do not automatically extend to content, data or other material published by Springer Nature that may be licensed
from third parties.
If you would like to use or distribute our Springer Nature journal content to a wider audience or on a regular basis or in any other manner not
expressly permitted by these Terms, please contact Springer Nature at
onlineservice@springernature.com