Morality constrains the default representation of what
is possible
Jonathan Phillipsa,1 and Fiery Cushmana
aDepartment of Psychology, Harvard University, Cambridge, MA 02138
Edited by Barbara H. Partee, University of Massachusetts, Amherst, MA, and approved March 21, 2017 (received for review November 30, 2016)
The capacity for representing and reasoning over sets of possi- a graded contrast between default and deliberative judgments
bilities, or modal cognition, supports diverse kinds of high-level rendered by a common system. In either event, the aim of the
judgments: causal reasoning, moral judgment, language compre- present study is to test for a form of nondeliberative modal cogni-
hension, and more. Prior research on modal cognition asks how tion, identify its signature properties, and determine its influence
humans explicitly and deliberatively reason about what is pos- on higher-level cognition.
sible but has not investigated whether or how people have a Critically, two independent research programs point toward a
default, implicit representation of which events are possible. We candidate signature property of the default representation: the
present three studies that characterize the role of implicit repre- relevance of prescriptive norms, such as morality and rationality.
sentations of possibility in cognition. Collectively, these studies This stands in contrast to prior research on an adult’s deliberative
differentiate explicit reasoning about possibilities from default reasoning about the (im)possibility of an event, which shows sen-
implicit representations, demonstrate that human adults often sitivity to descriptive norms, such as the likelihood of the event
default to treating immoral and irrational events as impossible, occurring (e.g., ref. 12), but not to prescriptive norms (e.g., refs.
and provide a case study of high-level cognitive judgments rely- 18 and 19). Thus, for instance, we tend to think “It isn’t possible
ing on default implicit representations of possibility rather than for Lewis to open the safe” when it is highly unlikely he will guess
explicit deliberation. the combination (a descriptive feature). In contrast, we tend not
to think “It isn’t possible for Lewis to take the money” when it
modality | high-level cognition | morality | norms | possibility is sitting right in front of him, but belongs to somebody else (a
prescriptive feature).
Remarkably, however, circumstantial evidence indicates that
Human thought often involves alternative possibilities: how less explicit modal representations may be sensitive to both pre-things could have been, might be, or may turn out. In other scriptive and descriptive norms. That is, our default assessment
words, it involves modal cognition, a capacity to construct and of Lewis’s possible actions may exclude what would be wrong in
reason over sets of nonactual events. Past research demonstrates the same way that it excludes what would be unlikely.
a role for modal cognition in assessing causal responsibility (1– One source of evidence comes from studies of young children.
3), making moral judgments (4–6), inferring what others intend Whereas children, much like adults, say that events that violate
to communicate (7–9), and much more (10). physical laws cannot happen (e.g., eating lightning), young chil-
To date, research on modal cognition has not distinguished dren also say that events that violate moral or social norms can-
between implicit and explicit representations of possibility. In not happen (e.g., stealing candy) (20–24). Not only do they judge
fact, past methods have overwhelmingly asked how we explic- these kinds of immoral events are “impossible,” but also they
itly and reflectively represent and reason about different possi- even say that such events would require “magic” (18, 19).
bilities. We now know a great deal about how and when people Additionally, some contemporary models of high-level cog-
engage in explicit counterfactual reasoning or reasoning about nition in adults argue that the underlying representation of
what else would have happened if a given alternative event had
occurred (see ref. 11 for a recent review). There is also a small
but growing body of research on how human adults deliber- Significance
ate and decide whether a given event is possible or impossible
(12). In addition, there is an emerging literature on the neural As humans, we think not only about what is, but also what
substrates recruited when participants are instructed to engage could be. These representations of alternative possibilities
in episodic counterfactual reasoning or simulation of possible support many important cognitive functions, such as predict-
future events (13, 14). ing others’ future actions, assigning responsibility for past
This research has been important for building an understand- events, and making moral judgments. We perform many of
ing of the way that human adults explicitly represent and rea- these tasks quickly and effortlessly, which suggests access
son about particular possibilities. But are these same processes to an implicit, default assumption about what is possible.
recruited when humans make high-level judgments that are What are the default features of the possibilities that we
known to involve representations of possibilities? There is rea- consider? Remarkably, we find a default bias toward rep-
son for doubt. Causal and moral judgments, for example, are resenting immoral or irrational actions as being impossible.
often made quickly and effortlessly and appear early in human Although this bias is diminished upon deliberative reflection,
development (15, 16). Thus, these judgments are unlikely to it is the default judgments that appear to support higher-level
rely exclusively on explicit deliberation about, or simulation of, cognition.
alternative possibilities. Moreover, deliberative judgments about
possibility often provide a poor fit to the patterns of high-level Author contributions: J.P. and F.C. designed research; J.P. performed research; J.P. ana-
judgments they supposedly inform (see, e.g., ref. 17 on the dis- lyzed data; and J.P. and F.C. wrote the paper.
sociation between causal judgments and explicit counterfactual The authors declare no conflict of interest.
reasoning). This article is a PNAS Direct Submission.
In light of this, an intriguing possibility is that people rely on Freely available online through the PNAS open access option.
a different kind of modal representation—one that is rendered 1To whom correspondence should be addressed. Email: phillips01@g.harvard.edu.
quickly and automatically, and is based on a distinctive set of This article contains supporting information online at www.pnas.org/lookup/suppl/doi:10.
constraints. This could reflect a dissociable “system” or instead 1073/pnas.1619717114/-/DCSupplemental.
www.pnas.org/cgi/doi/10.1073/pnas.1619717114 PNAS | May 2, 2017 | vol. 114 | no. 18 | 4649–4654
PSYCHOLOGICAL AND
COGNITIVE SCIENCES
possibility must be sensitive to both descriptive and prescriptive not significantly affect participants’ judgments of the possibil-
norms. This feature has played a critical role, for example, in ity of ordinary events (z =−0.164, P =0.87) or improbable
understanding human causal reasoning (25, 26), modality in nat- events (z =−1.056, P =0.291). Although not predicted, partic-
ural language (27), and judgments of force and freedom (28), ipants more tended to judge that events that violated physics
among others (10). These models do not, however, provide evi- were impossible after deliberating (M =85.22, SD=17.44),
dence for how this modal representation is computed. than when forced to answer quickly (M =79.37, SD=15.82),
In sum, this evidence suggests that there may be a nondeliber- (z =−6.628, P < 0.001). Critically, however, we found that par-
ative and early-emerging form of modal cognition that incorpo- ticipants’ judgments of the possibility of immoral and irrational
rates information about both descriptive and prescriptive norms. events were affected by deliberation in the opposite direction.
According to this hypothesis, the default representation would They tended to judge that immoral events were impossible
support not only judgments of what is “possible,” but also diverse when they were not able to reflectively deliberate (M =37.30,
modal judgments concerning what “ought,” “should,” “might,” SD=30.21), but more judged them to be possible after delib-
“may,” or “could” be done. Moreover, the default representation erating (M =21.41, SD=31.26), (z =5.423, P < 0.001). Sim-
would serve as input to diverse high-level cognitive judgments. ilarly, they tended to judge that irrational events were impos-
We pursue these hypotheses in three sets of studies. sible more when they were not able to deliberate (M =42.10,
SD=25.89), than after deliberating (M =37.86, SD=30.89),
Study 1: Deliberative vs. Default Representations of (z =1.963, P < 0.05) (Fig. 1).
Possibility In short, we find evidence that default representations of what
To investigate whether there are differences in more and less is possible differ from more reflective representations in that
deliberative forms of modal cognition, we began by focusing prescriptive norms selectively constrain the default representa-
specifically on judgments of possibility, asking participants to tion of possibility. Moreover, two key contrasts suggest an espe-
make judgments of whether various events were possible or cially strong role for moral norms: Judgments of the possibility
impossible, while manipulating the amount of time available for of immoral events were more affected by deliberation than those
this judgment. These judgments were made in the context of of either improbable events χ2(1)= 28.035, P < 0.001, or irra-
some background information, such as i: tional events χ2(1)= 24.937, P < 0.001, despite the fact that all
of these events involved clear norm violations.
i) Josh is on the way to the airport to catch a flight for a hunting
safari in Africa. He leaves with plenty of time to make it there, Study 2: Does This Default Generalize Across Modal
but his car breaks down on the highway. Now Josh is sitting in Judgments?
his car near a busy intersection and knows he needs to get to
airport soon if he is going to catch his flight. Linguistically, different modal auxiliaries (e.g., might vs. ought)
are known to select different sets of possibilities (7, 8). All of
Participants read six different background contexts, and after these modal auxiliaries are alike, however, in that they select
each one they were shown a series of candidate events one at some set of contextually relevant possibilities. This is also true
a time. As each event was presented, participants pressed a in nonlinguistic cognition: We think differently about what a
key to indicate whether they thought that the event was pos- person could or might do than we do about what they should
sible or impossible. Crucially, participants either were forced or ought to do. In all of these cases though, these judgments
to make these judgments very quickly (≤1,500 ms) or were require us to reason over the set of events that are represented as
asked to reflectively deliberate on the possibility of each event being available in the context. Accordingly, we can ask whether
(≥1,500 ms). We used a total of 144 different events, which were distinct modal judgments are supported by distinct default sets
designed to fall into five categories: 48 ordinary events that did of possibilities or whether instead they share a common set of
not violate any norms, e.g., ii, a; 24 events that violated statisti- default possibilities and become differentiated only after addi-
cal norms, e.g., ii, b; 24 events that violated physical laws, e.g., tional processing.
ii, c; 24 events that violated moral rules, e.g., ii, d; and 24 events
that violated norms of rationality, e.g., ii, e. (The a priori catego-
rization of these events was confirmed in independent ratings of
the morality, probability, and rationality of each of the events; SI Ordinary Improbable Impossible Immoral Irrational
Text, Event Ratings and Fig. S1.)
Reflective
ii) Is it possible or impossible for Josh to [(a)] hail a taxi at the Speeded
intersection, [(b)] fix his car by banging on it, [(c)] teleport 75
himself to the airport, [(d)] sneak onto public transportation,
or [(e)] sell his car for a ride to the airport?
This paradigm uses judgments of whether an event is possible
as an indicator of whether participants represent that event as
being available in the context. By changing the amount of time 50
participants have to deliberate, we ask whether (and how) par-
ticipants’ representation of the set of available events changes as
a result of more or less deliberation.
We first used a series of linear mixed-effects models to test
25
whether participants’ judgments of the possibility of the five
different kinds of events differed across time pressure/delay
conditions (SI Text, Statistical Approach). We observed an effect
of whether participants were reflectively deliberating, χ2(1)=
20.653, P < 0.001, and an effect of event-type χ2(4)= 860.48, 0
P < 0.001. Critically, however, these main effects were quali-
fied by a deliberation × event-type interaction, χ2(4)= 64.093, Fig. 1. Judgments of impossibility for five different kinds of events when
P < 0.001. We decomposed this interaction, using a series of participants made judgments after deliberating (solid bars) or without time
generalized linear models. These revealed that deliberation did to deliberate before responding (shaded bars). Error bars indicate ±1 SEM.
4650 | www.pnas.org/cgi/doi/10.1073/pnas.1619717114 Phillips and Cushman
% of Events Judged Impossible
We propose that there is a common mechanism responsible
for rapidly constructing a default representation of what is pos-
sible, which supports diverse cognitive functions. This proposal
predicts that any deviation from the default representation will
require additional processing, and thus the selection of different
sets of possibilities for different modal judgments is likely the
result of more deliberative cognition. This prediction is testable:
We can ask whether the correlation between judgments of what a
person could do, might do, ought to do, etc., is higher when peo-
ple do not have time to reflect (and thus rely more on a default
representation of possibility) and lower when people do have
time to reflect before making a modal judgment.
An obvious alternative prediction is that reducing the amount
of time that participants have to respond should increase the
noise observed in their responses. This predicts that modal judg-
ments will instead be less correlated with one another when
they have to be made extremely quickly (compared with when
they are made with time to reflect). We hypothesized that this
effect would be outweighed by the influence of a common default
modal representation.
Motivated by the results of study 1, we further hypothesized
Fig. 2. Correlation coefficient of each pair of modal questions, either when
that the default representation underlying diverse modal judg- participants were forced to answer quickly (left) or when they were given
ments would be sensitive to prescriptive norm violations. In other time to reflect before answering (right).
words, we hypothesized that under time pressure participants’
judgments of what a person could do, may do, might do, etc.,
would be sensitive to whether that action is immoral or irrational. culated the correlation between each pair of modal judgments,
This common sensitivity to prescriptive norms should thus be first focusing on events where descriptive norms were relevant
largely responsible for heightened similarity among the default and then, second, focusing on events where prescriptive norms
representations elicited by diverse modals. were relevant. This analysis allows us to ask whether it was the
Following the basic procedure introduced in study 1a, we col- presence of one of these kinds of norms that was responsible for
lected deliberative vs. nondeliberative judgments of what agents the difference in correlations when participants reflected before
may do (study 2a), might do (study 2b), could do (study 2c), answering.
ought to do (study 2d), or should do (study 2e). We also included Analyzing these correlations with linear mixed-effects models,
judgments of what it is possible for them to do (from study 1). we found an interaction effect between the type of norms that
These different modal terms are known to express different kinds were relevant and whether or not participants had time to reflect,
of modality (metaphysical, circumstantial, deontic), and the over- χ2(1)= 24.323, P < 0.001. Decomposing this interaction, we
all patterns of responses demonstrate that participants tracked first focused on modal judgments made in scenarios for which
these differences (SI Text, Differences Across Modal Auxiliaries descriptive norms were primarily relevant (e.g., winning the lot-
and Fig. S2). tery). Here, we observed a modest decrease in the correlation
Our analyses treat the 144 different events as the primary unit between the modal judgments when participants reflected before
of analysis. For each of the six modal judgments, we calculated answering (Mr =0.873, SDr =0.055) from when they answered
the proportion of the time that the modal judgment was accepted before reflecting (Mr =0.908, SDr =0.042), t(14)= 5.336, P <
vs. rejected for each event, doing so independently for trials when 0.001, d =0.724. Next, we focused on judgments made in sce-
participants had time to reflect and when they did not. These narios for which prescriptive norms were relevant (e.g., theft)
scores reflect the likelihood that an event will be included in and found a much larger decrease in the correlations between
the set of possibilities relevant to each type of modal concept the modal judgments made when participants reflected before
(e.g., what should be done, what could be done, and so on). We answering (Mr =0.278, SDr =0.265), compared with when
then calculated the correlation between the events’ representa- they answered before reflecting (Mr =0.628, SDr =0.145),
tions for each possible pair of different modal judgments, both t(14)= 8.841, P < 0.001, d =1.642 (Fig. 3). A corollary finding
when participants were given time to reflect before answering is that deontic modals show less of a difference between delibera-
and when they were not. tive and reflective responses than metaphysical or circumstantial
As hypothesized, modal judgments were significantly more modals (SI Text, Within-modal correlations).
correlated when participants did not have time to reflect (Mr = Together, these analyses indicate that prescriptive norms con-
0.893, SDr =0.060) than when they did (Mr =0.830, SDr = stitute a principal dimension differentiating specific modal con-
0.101), t(14)= 5.131, P < 0.001, d =0.754 (Fig. 2). This finding cepts from their common default. In contrast, the use of descrip-
supports the view that participants relied on a common default tive norms appears to be relatively more consistent across
representation of the set of available events when answering distinct modal concepts, even upon reflection. This pattern is
quickly. By contrast, when participants reflected on each judg- most evident for moral norms. Focusing specifically on the
ment, their responses were more strongly dictated by unique fea- immoral events, which clearly distinguish different kinds of
tures of each modal concept and thus became less correlated with modal reasoning, we created a correlation matrix of all of the
one another. different judgments. This approach revealed that all of the dif-
We next asked whether a principal dimension on which default ferent modal judgments were highly correlated when partic-
judgments became more similar was sensitivity to prescriptive ipants were not able to reflect (0.558≤ r ≤ 0.919), but were
norms. To do this, we divided the events that participants judged not when participants took time to reflect (−0.016≤ r ≤ 0.572)
into two categories: those for which descriptive norms were (Fig. 4). A corollary finding is that participants’ nonreflective
primarily relevant (the improbable and physically impossible modal judgments were all highly correlated with reflective judg-
events) and those for which prescriptive norms were primarily ments of what agents ought to do (rMean =0.465), but not other
relevant (the irrational and immoral events). We then again cal- reflective judgments (−0.091< rMean > 0.264). Put simply, when
Phillips and Cushman PNAS | May 2, 2017 | vol. 114 | no. 18 | 4651
PSYCHOLOGICAL AND
COGNITIVE SCIENCES
the ordinary, immoral, irrational, or improbable events used pre-
viously (ii, a–e), as in iii:
iii) Josh calls his father who lives a few states away and tells
him about his problem. Not really knowing how to help him,
Josh’s father makes a suggestion. His father says Josh could
sneak onto public transportation.
Participants were then told that, notwithstanding this advice, the
agent decided to pursue some specified alternative course of
action, as in iv:
iv) Josh ignores his father’s suggestion and decides to book the
next available flight, even though it is quite expensive.
Within each context, the agent was always described as pursuing
the same course of action, regardless of the advice given.
After reading all of this information, participants were asked
whether they agreed or disagreed that the agent was forced to do
the action he or she actually pursued, as in v:
v) Josh was forced to book the next available flight.
Following previous work on judgments of force (28, 32, 33),
Fig. 3. Correlation coefficient for each pair of modal judgments either participants should judge the agent to have been forced only if
for items where descriptive norms were primarily relevant (Left) or for they believe that there were no other options available to the
items where prescriptive norms were primarily relevant (Right). Correlations agent—in other words, if the set of relevant alternative possibil-
between modal judgments made without time to reflect are plotted on the ities was empty. Thus, in the context of the specific advice given
left-hand side of each panel; correlations between modal judgments made to the agent, participants should agree that the agent was forced
after reflecting are plotted on the right-hand side of each panel.
only if they did not represent the proposed option as an available
possibility.
answering quickly, modal judgments become more ought-like Crucially, in all cases, the judgments of force were made with
(box in Fig. 4). unlimited time available. Thus, our question is whether deliber-
In summary, we have demonstrated thus far that default modal ative judgments of force for each scenario are better predicted
judgments show a common sensitivity to both descriptive and by deliberative or speeded judgments of possibility for the same
prescriptive norms. In contrast, reflective modal judgments can scenario in study 1. We hypothesized that force judgments would
deviate away from this default, allowing for some kinds of modal be best predicted by speeded possibility judgments, reflecting a
cognition (e.g., thoughts of what could happen) to focus primar- dominant influence of the default representation of the set of
ily on what is physically possible, without regard to immorality or available events.
irrationality. To analyze these data, we computed the average agree-ment that the agent was forced to do a given action when the
Study 3: Default Modal Representations in High-Level
Cognition
Existing evidence suggests that the modal representations used
in other domains of high-level cognition—for instance causal
attribution, language comprehension, and moral judgment—
are similar to the default representations identified in studies
1 and 2. For instance, causal judgments of an event are sensitive
to the moral status of alternative events (e.g., refs. 25 and 29),
which is consistent with the demonstrated role of prescriptive
norms in default modal representations. This current evidence
is indirect, however, and the role of alternative events in causal
cognition remains controversial (e.g., ref. 30). Accordingly, study
3 provides a direct test of the hypothesis that the default rep-
resentations of the set of events available in a given context are
exported to other cognitive domains.
To do this, we used judgments of “force” as a case study. This
is an ideal test because empirical (28, 31) and theoretical (32, 33)
studies concur that judgments of force rely on representations of
alternative possibilities. Put simply, if an agent was “forced” to
act in some way, this implies that there was no relevant alterna-
tive; and if he or she was not forced, this implies that some alter-
native existed. We therefore asked, when people decide whether
an agent was forced to do something, do they rely on a set of pos-
sibilities more like a deliberative set or instead on the implicit set
that we identified in study 1?
Participants in this study once again read the six different back- Fig. 4. Graphical depiction of the correlation matrix for speeded and
ground contexts used in the previous studies. However, in this reflective modal judgments of events that involved immoral actions. Squares
case, the agent asked another person for advice about what to do. within the black box (bottom left) depict correlations of speeded modal
This person then directly suggested that the agent pursue one of judgments with reflective judgments of ought.
4652 | www.pnas.org/cgi/doi/10.1073/pnas.1619717114 Phillips and Cushman
alternative proposed was one of the events used in the previ- mark of this default representation is that it tends to exclude
ous studies. We then first asked whether participants’ nonreflec- immoral actions. First, we found that time pressure makes peo-
tive judgments were predictive of judgments of force in a sim- ple less likely to judge it possible to act immorally or, to a lesser
ple linear model and found that they were, F (1, 118)= 46.12, extent, irrationally. Second, we found that this effect is reflected
P < 0.001, η2 =0.281 (Fig. 5). Next, we asked whether reflective across diverse modal concepts: what a person ought to do, may
or nonreflective judgments of possibility were a better predic- do, could do, should do, or might do. These judgments appear
tor of judgments of force by comparing a series of linear mixed- to share a common default basis and become more differenti-
effects models. We found that including reflective judgments of ated only after deliberative processing. Finally, as a case study of
possibility did not significantly improve a model that already the role of implicit modal cognition in reflective high-level judg-
included nonreflective judgments of possibility, χ2(1)= 0.030, ments, we found that default modal judgments are better predic-
P =0.863. In contrast, including nonreflective judgments of pos- tors than reflective modal judgments of people’s decisions about
sibility did significantly improve a model that already included whether a person was forced to act in a particular way. A medi-
reflective judgments of possibility, χ2(1)= 27.640, P < 0.001. ation analysis indicates that, to the extent that we are forced to
Thus, above and beyond reflective judgments of what is possible, act morally, it is because our default representation of what is
the default representation of possibility is predictive of delibera- possible tends to exclude immoral actions.
tive judgments of force. Although all forms of modal cognition involve constructing
We next asked whether morality had an effect on judgments and reasoning over sets of possibilities, it is remarkable that they
of force [as found in previous work (28, 31)] and whether this apparently begin with a common default template. By analogy,
effect was mediated by nonreflective judgments of possibility. all forms of social cognition—whether about in-groups or out-
We found that this was the case: Participants judged agents groups, cooperators or competitors—involve reasoning about
to be more forced when the proposed alternative was immoral sets of other people, but it would be remarkable to discover that
vs. ordinary, t(70)= 7.309, P < 0.001, d =1.827. Moreover, this under time pressure, all forms of social cognition begin with a
effect was mediated by nonreflective judgments of possibility, representation of the very same individuals. In modal cognition,
95%CI [0.018, 0.542], P =0.04. A similar pattern of results much like in social cognition, one might have initially thought
was also observed for events that involved violations of ratio- that even rapid categorizations of what is possible for a person to
nal norms (t(70)= 5.462, P < 0.001, d =1.366; 95%CI [0.045, do and what a person ought to do would depend on separate pro-
0.864], P =0.04). These results fit well with ongoing work that cesses that pick out separate events. After all, most people would
offers a semantics for force according to which prescriptive say that it is possible to run a red light at an empty intersection,
norms impact the set of actions available to an agent (34). (See but not that one ought to. In contrast to this intuition, our find-
SI Text, Study 3b: Can Inferences About Agents’ Desires Explain the ings suggest that it takes time to realize that it is possible to run
Impact of Norms on Judgments of Force? for a test of an alterna- a red light. We begin with a default representation of possibil-
tive semantics for force that proposes to account for the impact ity that tends to exclude this action, along with other immoral or
of norms through inferences about the agents’ desires.) irrational acts.
The purpose of this default mechanism is a key area for further
Discussion study. One appealing possibility is that the default representation
We find evidence for a default representation of what is possi- serves the function of proposing actual candidate actions during
ble, i.e., a set of events, specific to a context, that constitutes a decision making—in other words, it helps us to construct a choice
starting point for understanding nonactual alternatives. A hall- set (35). It is natural to suppose that screening out immoral or
irrational actions from one’s own decision making would tend to
be adaptive.
Importantly, the default representation we have uncovered
bearsastrikingresemblance to thepatternofpossibility judgments
observed early in human development (18, 19, 21, 24). In both
cases, judgments of possibility are sensitive not only to descrip-
tive norms but also to prescriptive ones. The similarity of these
patterns presents the intriguing possibility that the modal repre-
sentation observed early in development is retained in adult cog-
nition alongside a separate later-developing capacity for delibera-
tively reasoning about possibilities. Future work should continue
to explore this connection directly and ask which other factors
serve to constrain both of these representations of possibilities.
Finally, it is worth noting a potential connection between our
research and the widespread, surprising presence of moral norms
in disparate corners of the human mind: e.g., in causal reason-
ing and mental state attribution (10, 25, 26, 29, 36). These pro-
cesses also involve reasoning over a set of possibilities. Given the
present results, an exciting possibility is that morality shapes how
we think about many things because it constrains the very possi-
bilities that come to mind.
Materials and Methods
Participants. All participants were recruited from Amazon Mechanical Turk
(www.mturk.com) through TurkPrime (www.turkprime.com), which was
used to prevent repeat participation across the studies. Sample sizes and
Fig. 5. Relationship between the implicit representation of possibility and demographic information are as follows: study 1a (judgments of possibility),
high-level judgments of whether or not an agent was forced to do a given 498 participants (Mage = 33.85, SDage = 10.82, 233 females); study 1b (event
action in six different contexts. Shaded areas represent 95% confidence ratings), 61 participants (Mage = 31.53, SDage = 7.87, 25 females); study 2a
intervals. (ought), 301 participants (Mage = 33.40, SDage = 9.87, 147 females); study 2b
Phillips and Cushman PNAS | May 2, 2017 | vol. 114 | no. 18 | 4653
PSYCHOLOGICAL AND
COGNITIVE SCIENCES
(might), 201 participants (Mage = 33.79, SDage = 10.20, 92 females); study 2c implicitModality. Studies 1 a and b and 2 a–e were conducted through Test-
(could), 304 participants (Mage = 34.87, SDage = 11.32, 144 females); study able (www.testable.org) and study 3 a and b were conducted in Qualtrics.
2d (may), 200 participants (Mage = 32.80, SDage = 9.21, 86 females); study 2e Stable links to experiments are as follows: study 1 a and b, (possibility)
(should), 297 participants (Mage = 34.05, SDage = 9.86, 153 females); study testable.org/t/3111b018 and (event ratings) testable.org/t/31041bbc; study
3 (force judgments), 400 participants (Mage = 38.05, SDage = 12.73, 218 2 a–e, (ought) testable.org/t/31092d3e, (might) testable.org/t/3120b96c,
females); and study 3b (desire judgments), 400 participants (Mage = 36.61, (could) testable.org/t/31bbed30, (may) testable.org/t/31f78b37, and (should)
SDage = 11.80, 200 females). These studies were approved by the Harvard testable.org/t/317a2d7c; and study 3 a and b, (force) goo.gl/6POVuY and
University Institutional Review Board, IRB14-2016, and informed consent (desire) https://goo.gl/30OjG1.
was acquired from all participants.
Access to Data, Analysis Code, and Materials. All materials used to conduct ACKNOWLEDGMENTS. This research was supported by Grant N00014-14-1-
these studies and analyze the results are available at github.com/phillipsjs/ 0800 from the Office of Naval Research.
1. Lewis D (1973) Counterfactuals (Blackwell Publishing, Oxford). 21. Kalish C (1998) Reasons and causes: Children’s understanding of conformity to social
2. Pearl J (2009) Causality (Cambridge Univ Press, New York). rules and physical laws. Child Dev 69:706–720.
3. Gerstenberg T, Tenenbaum JB (2017) Intuitive theories. Oxford Handbook of Causal 22. Kushnir T, Gopnik A, Chernyak N, Seiver E, Wellman HM (2015) Developing intuitions
Reasoning, ed Waldmannn M (Oxford Univ Press, New York), pp 515–538. about free will between ages four and six. Cognition 138:79–101.
4. Kane R, ed (2011) The Oxford Handbook of Free Will (Oxford Univ Press, New York). 23. Shtulman A (2009) The development of possibility judgment within and across
5. Hart HLA, Honoré T (1985) Causation in the Law (Oxford Univ Press, New York). domains. Cogn Dev 24:293–309.
6. Lagnado DA, Gerstenberg T, Zultan R (2013) Causal responsibility and counterfactuals. 24. Shtulman A, Carey S (2007) Improbable or impossible? How children reason about the
Cogn Sci 47:1036–1073. possibility of extraordinary events. Child Dev 78:1015–1032.
7. Portner P (2009) Modality (OUP, Oxford). 25. Halpern JY, Hitchcock C (2015) Graded causation and defaults. Br J Philos Sci 66:413–
8. Kratzer A (2012) Modals and Conditionals: New and Revised Perspectives (Oxford Univ 457.
Press, New York), Vol 36. 26. Kominsky JF, Phillips J, Gerstenberg T, Lagnado D, Knobe J (2015) Causal superseding.
9. Stalnaker R (2002) Common ground. Linguist Philos 25:701–721. Cognition 137:196–209.
10. Phillips J, Luguri J, Knobe J (2015) Unifying morality’s influence on non-moral judg- 27. Knobe J, Szabó ZG (2013) Modals with a taste of the deontic. Semant Pragmat 6:
ments: The relevance of alternative possibilities. Cognition 145:30–42. 1–42.
11. Byrne RM (2016) Counterfactual thought: From conditional reasoning to moral judg- 28. Young L, Phillips J (2011) The paradox of moral focus. Cognition 119:166–
ment. Annu Rev Psychol 67:135–157. 178.
12. Shtulman A, Tong L (2013) Cognitive parallels between moral judgment and modal 29. Hitchcock C, Knobe J (2009) Cause and norm. J Philos 106:587–612.
judgment. Psychon Bull Rev 20:1327–1335. 30. Samland J, Waldmann MR (2016) How prescriptive norms influence causal inferences.
13. Schacter DL, Benoit RG, De Brigard F, Szpunar KK (2015) Episodic future thinking Cognition 156:164–176.
and episodic counterfactual thinking: Intersections between memory and decisions. 31. Phillips J, Knobe J (2009) Moral judgments and intuitions about freedom. Psychol Inq
Neurobiol Learn Mem 117:14–21. 20:30–36.
14. De Brigard F, Spreng RN, Mitchell JP, Schacter DL (2015) Neural activity associated with 32. Berlin I (1970) Four Essays on Liberty (Oxford Univ Press, New York), Vol 969.
self, other, and object-based counterfactual thinking. NeuroImage 109:12–26. 33. Aristotle (2002) Nicomachean Ethics (Oxford Univ Press, New York).
15. Hamlin JK, Wynn K, Bloom P (2007) Social evaluation by preverbal infants. Nature 34. Mandelkern M, Phillips J (2017) Force: Topicalization, context-sensitivity, and morality.
450:557–559. Available at https://osf.io/9shca/. Accessed April 6, 2017.
16. Buon M, Jacob P, Loissel E, Dupoux E (2013) A non-mentalistic cause-based heuristic 35. Ben-Akiva M, Boccara B (1995) Discrete choice models with latent choice sets. Int J
in human social evaluations. Cognition 126:149–155. Res Mark 12:9–24.
17. Mandel DR, Lehman DR (1996) Counterfactual thinking and ascriptions of cause and 36. Knobe J (2003) Intentional action in folk psychology: An experimental investigation.
preventability. J Pers Soc Psychol 71:450–463. Philos Psychol 16:309–325.
18. Phillips J, Bloom P (2017) Do children believe immoral events are magical? Available 37. Jaeger TF (2008) Categorical data analysis: Away from ANOVAS (transformation or
at https://osf.io/en7ut/. Accessed April 6, 2017. not) and towards logit mixed models. J Mem Lang 59:434–446.
19. Shtulman A, Phillips J (2017) Differentiating could from should: Developmental 38. Barr DJ, Levy R, Scheepers C, Tily HJ (2013) Random effects structure for confirmatory
changes in modal cognition. Available at https://osf.io/d4xs5/. Accessed April 6, 2017. hypothesis testing: Keep it maximal. J Mem Lang 68:255–278.
20. Browne CA, Woolley JD (2004) Preschoolers’ magical explanations for violations of 39. Bates D, Maechler M, Bolker B, Walker S (2015) Fitting linear mixed-effects models
physical, social, and mental laws. J Cogn Dev 5:239–260. using lme4. J Stat Softw 67(1):1–48.
4654 | www.pnas.org/cgi/doi/10.1073/pnas.1619717114 Phillips and Cushman