Pluralistic Ignorance in the Bystander Effect

Informational Dynamics of Unresponsive Witnesses in Situations

calling for Intervention

Rasmus K. Rendsvig

Department of Media, Cognition and Communication

University of Copenhagen

Link to published version on SpringerLink

Abstract. The goal of the present paper is to construct a formal explication of

the pluralistic ignorance explanation of the bystander effect. The social dynamics
leading to inaction is presented, decomposed, and modeled using dynamic epis-
temic logic augmented with transition rules able to characterize agent behavior.
Three agent types are defined: First Responders who intervene given belief of ac-
cident; City Dwellers, capturing apathetic urban residents and Hesitators, who
observe others when in doubt, basing subsequent decision on social proof. It is
shown how groups of the latter may end in a state of pluralistic ignorance leading
to inaction. Sequential models for each agent type are specified, and their results
compared to empirical studies. It is concluded that only the Hesitator model pro-
duces reasonable results.

Keywords: bystander effect, pluralistic ignorance, social dynamics, social proof,

social influence, dynamic epistemic logic

1 Introduction: The Bystander Effect and Pluralistic Ignorance

On March 13, 1964, in Queens, New York, Catherine Susan Kitty Genovese was
raped and stabbed, the assailant fleeing multiple times during the ongoing assault that
resulted in Genoveses death. Multiple residents witnessed parts of the nearly one hour
long attack, without successfully intervening.
The foremost explanation put forth in the ensuing media coverage was apathy
among urban citizens. Through the pressures of city life, homo urbanis has lost his
sense of empathy for fellow man, all but grown indifferent to their quarrels (Latan and
Darley, 1970, 1968). This explanation struck social psychologists John M. Darley and
Bibb Latan as incorrect, whereupon they set out to provide an alternate explanation.
The first publication of experimental results is the classic (Latan and Darley, 1968),
which was soon followed by a vast amount of studies (see Latan and Nida (1981)
for a review). In these studies, focus has been moved from why urban citizens do not
help, to why people in groups are less prone to help. The collected data shows a robust
tendency, namely that the chance of help being offered diminishes as the number of
witnesses increases. This tendency will be referred to as the bystander effect.
An Explanation of the Intervention Process. One currently used text-book explana-
tion (see e.g. (Myers, 2012)) of the bystander effect stems from (Latan and Darley,
1970), and involves three steps which each bystander must go through before he or she
will intervene.1 First, a bystander must notice the event in question. With busy street life
involving traffic and pedestrians, the risk of overlooking a seizure is higher than on des-
olate streets. Where the problematic situation goes unnoticed, help will not be offered.
Second, if noticed, the bystander must interpret the event in order to decide whether an
emergency is occurring or not. In many cases, this will not prove to be a problem: situ-
ations involving car accidents or bleeding victims are seldom epistemically ambiguous.
However, a man slumped on a bench may provide an epistemic conundrum, as he may
be merely mumbling curses against the general youth, marooned following a too Sak
intense business lunch, or moaning in pain from the onset of a seizure. Such ambiguities
may be sought resolved by the acquisition of further information, readily present in the
form of social proof : if other bystanders are not showing signs of distress, the event will
be perceived as less critical and therefore ultimately bypassed. Third, in case the event
is interpreted as requiring help, the bystander has to gauge whether to take responsibil-
ity: when alone, there is no question as to who should intervene, but when gathered in
groups, diffusion of responsibility may arise. Such diffusion may be caused by uncer-
tainty as to whether we are among the best qualified to handle the situation, whether
others have already called for paramedics or are just about to act. When alone, the re-
sponsibility to intervene rests on one individual, but when in a group, the same pressure
is apparently distributed among all, thereby diminishing the chances that anyone will
act.

Pluralistic Ignorance. The goal of this paper is to model the social informational dy-
namics and decision procedures running the second of these steps, specifying condi-
tions under which a group of agents in an ambiguous situation may choose to seek
social proof in order to individually determine a correct course of action and the asso-
ciated consequences thereof, hereby providing a detailed explanation for (this part of)
the bystander effect. This narrower focus is taken as the second step of the bystander
effect explanation constitutes an interesting informational dynamics in its own right,
useful to the analysis of social situations in which neither distractions nor diffusion
play important roles.
The second step of the dynamics revolve around a belief state often referred to as
one of pluralistic ignorance: a situation in which everybody believes that everybody
else believes a given proposition/endorses a given norm, while no-one in fact believes
it/endorses it.2 Pluralistic ignorance has been put forth as a decisive factor in a plethora
of social situations, including the introduction of various unpopular norms such as col-
lege binge drinking (Prentice and Miller, 1993) and violent gang behavior (Bicchieri
and Fukui, 1999), the persistence of poor strategies in light of poor firm performance
(Westphal and Bednar, 2005) and lack of help seeking in class rooms (Miller and Mc-
1 For a comprehensive walk-through of this explanation and supportive data, refer to (Latan
and Darley, 1970).
2 (Halbesleben and Buckley, 2004) provides an illuminating overview of the history and devel-
opment of the term.
Farland, 1987); the allowance of ongoing mortgage deed merry-go-rounds in Denmark
during the financial upswing of 2007 (Hendricks and Rasmussen, 2012; Hansen et al.,
2013).
Pluralistic ignorance may cause individuals not to act for more than one reason.
One may be due to social inhibition you may not wish to be the only one raising your
hand to ask a question. Here, inaction results from vanity and social identity. There
may be doubt you may not wish to call the police if there is no cause for alarm.
In case of doubt, inaction follows from incorrect information processing. Both causes
may individually lead to inaction and may further co-occur. In the following, pluralistic
ignorance will be used to refer only to processes of the latter kind.

Structure of the Paper. In Section 2, an example of a social dynamics involving plural-

istic ignorance which lead to unfortunate inaction is presented, and an informal sketch
of the information processing involved is outlined, the structure of which is used as
a guideline for the formal representation. In Section 3, elements from dynamic epis-
temic logic (DEL) are presented. DEL is the primary modeling tool, used to represent
static epistemic states and belief revision in light of new information. Section 3 also
presents the modeling of the occurrence of the accident. In Section 4, it is shown how
agent types may be defined for DEL, hereby augmenting the framework with a notion
of choice agency. Three agent types relevant to the bystander effect are defined: First
Responders, a good samaritan type agent, who will choose to intervene in emergen-
cies if she believes one such is occurring; City Dwellers, capturing the apathetic urban
resident and Hesitators, who observes others when in doubt. In Section 5, it is shown
how Hesitators misinterpretation of other Hesitators choice to observe may lead to
a state of pluralistic ignorance. Coupled with Hesitators basing subsequent action on
social proof obtained through observation, it is further shown that this agent type will
choose to evade the scene of the accident, though they privately believe there is cause
for intervention. In Section 6, models run with each of the three agent types are com-
pared to empirical studies. It is concluded that the Hesitator model is the only of the
three producing reasonable results. In Section 7, we conclude.

2 From Ambiguity to Inaction

Consider the following example, inspired by (Halbesleben and Buckley, 2004). A firm
is performing poorly and this is caused by the currently implemented business strat-
egy. Every member of the board of directors has equal access to the relevant infor-
mation pertaining to strategic choice and firm performance, which to each member
strongly indicates that the prudent action is to change strategy. The available informa-
tion is, however, not conclusive: the possibility that a strategy change will leave the
firm performance-wise worse off cannot be ruled out. Hence, every board member is in
an epistemically ambiguous situation, where they privately believe intervention will be
fruitful, but all would prefer additional information before settling for a vote against the
status quo. As the voting situation arises, all seek such further information from their
peers, hoping that the votes of others will illuminate them. As all look to each other,
nobody initially raises an objection, which is interpreted by others as the choice that no
objection need be raised. Hereby, all conclude that their intelligent peers believe that
the status quo should not be changed. This perceived consensus among peers is then
seen as providing evidence to the conclusion that the currently implemented strategy
is in fact desirable. Having thus reflected, board members choose not to intervene in
the status quo, and the low firm performance continues another term or two. Though a
fictitious example, pluralistic ignorance does occur on corporate boards, often resulting
in poor strategic decisions (see (Westphal and Bednar, 2005)).
While the example differs in topic from a street accident, the two share common
information dynamics. The street incident revolves around the commonly unwanted
case of physical pain, the board meeting example focuses on the commonly unwanted
event of a sub-optimal status quo strategy. Both involve uncertainty about whether an
unwanted situation is the case or not, both require intervention in case the unwanted
situation indeed is occurring, and both allow for the gathering of further information by
observing peers.
Notice how the example implicitly utilizes two instances of pluralistic ignorance.
First, there is one instance of what may be called norm-based pluralistic ignorance:
though every member uses the decision rule if in doubt, seek further information, they
assume that others will follow a different rule, namely if in doubt, raise an objection.
This is an instance of pluralistic ignorance as everybody believes that everybody else
follows a given norm (here, a decision rule), while in fact no-one follows it. The second
instance of pluralistic ignorance is proposition-based: everybody ultimately believes
that everybody else believes that status quo is fine, while everybody privately believes
the strategy should be changed.

Fig. 1. Flowchart of the dynamics leading to the bystander effect. Boxes with solid
lines represent epistemic states boxes with dotted lines indicate events.

The dynamics involved in the example may be decomposed into eleven elements
(see Fig. 1): six static states and five state-altering transitions. To start from scratch,
in the first state, nothing has happened (1). This is followed by the occurrence of an
event, epistemically ambiguous between being an accident (stabbing, onset of poor per-
formance) or nothing of consequence (dispute, reasonable performance) (2). This event
results in a second state, where everybody privately believes that an accident occurred,
while remaining ignorant about the beliefs of others (3). Based on this state, one may
choose to intervene (rush to help, object to strategic choice), may choose actively to
evade (ignore the stabbing, withhold objection), or may choose to seek further infor-
mation by observing the choice made by others. Crucially, the performed actions of
evasion and observation are here considered to be epistemically ambiguous: in seeking
further information, we do not want to flaunt our ignorance, so further observation is
made discretely. Given the ambiguity of the accident, observation is chosen and ex-
ecuted (4). It is claimed here that a crucial further element for the dynamics is the
resulting mis-perception of this choice when made by others: though we ourselves may
choose to observe, when we see others do the same, we consider it plausible that they
in fact chose to evade. Given this norm-based pluralistic ignorance, in the ensuing third
state (5), all still believe there was an accident, while believing that all others evaded.
To obtain information about the beliefs of others, their perceived actions must now be
interpreted (6): given that you evaded, what may I conclude about your beliefs pertain-
ing to the accident? Under the assumption that you are a reasonably decent person, only
that you believed there was none. Such interpretations conducted by all then results in
a fourth state (7) of proposition-based pluralistic ignorance: though we all believe there
was an accident, we also believe that no-one else believes so. Revising our beliefs in the
light of the obtained social proof (8), all conclude that no accident occurred (9). Given
a further chance to act (10), evasion will be the natural choice, leading to the final state
(11), where the accident in fact occurred, everybody believed so, but nevertheless chose
to evade it, due to the social information dynamics.

3 Plausibility Models for States and Actions

The sketch presented above suggests several ingredients required for a suitable model,
including propositional- and higher-order beliefs (beliefs about beliefs), belief change
in light of new information and agent action. To that end, dynamic epistemic logic3 with
updating by action models with postconditions is a suitable framework: All higher-order
beliefs are specified in relatively small models, factual change may be modeled using
postconditions, and the dynamics may be built step-by-step, allowing for a detailed
overview of each step. Step-by-step construction allows for easy replacement of single
modules, whereby alternative runs may be investigated. For each such run, the dynam-
ics terminate when either someone agent intervenes, or all agents choose to evade. The
dynamics presented concern a three agent case, with group size variations presented in
Section 6. Throughout, the same complete graph all see all social network structure is
assumed.

3 Technically, no logic is introduced; the dynamics are investigated using only model theory.
3.1 Statics
Multi-agent Plausibility Frames. Where A is a finite set of agents, a multi-agent plau-
sibility frame (MPF) is a structure S = (S, i )iA where S is a finite set of states and
each i is a well-preorder4 .
The idea behind plausibility frames is that such encode the knowledge and beliefs
of a group of agents, A, capturing which states each agent may tell apart, and how
plausible these states are relative to one another. If two states s,t are connected by i ,
then i cannot tell these states apart, but if s <i t (i.e. s i t and t 6i s), then i considers s
more plausible than t.5 Fig. 2 illustrates a simple plausibility frame F1 with two states,
s and t, and two agents, a and b. The arrow from t to s captures that s <b t, i.e. that b
cannot distinguish between s and t, but finds s strictly more plausible.6 Reflexive arrows
will only be drawn if they are the only arrows for a given agent. In Fig. 2, a cannot tell
s from s nor t from t.
a a
s t
b
Fig. 2. A simple, two agent plausibility frame, denoted F1

Indistinguishability Relation; Information and Plausibility Cells. Given an MPF S =

(S, i )iA , the indistinguishability relation for agent i is the equivalence relation i :=i
i . Further, the information cell of agent i at state s is Ki [s] = {t : s i t} and the plau-
sibility cell of agent i at state s is Bi [s] = Mini Ki [s] = {t Ki [s] : t i s , for all s Ki [s]}.
The plausibility cell Bi [s] contains the worlds the agent find most plausible from the in-
formation cell Ki [s] and represent the doxastic appearance (Baltag and Smets, 2008,
p. 25) of s to i.7 Notice that s i t means that s is at least as plausible as t for i.
In Fig. 2, it is the case that s a s and t a t, but not s a t. Hence Ka [s] = {s} and
Ka [t] = {t}. For agent b, s b s, t b t and s b t, entailing that Kb [s] = Kb [t] = {s,t}.
However, even if the actual state is t, agent b finds s more (most) plausible, so it is the
sole element in bs plausibility cell at t: Bb [t] = {s}.

Doxastic Propositions. When considering plausibility frames, sets of states may be

identified with propositions; e.g. the set {s} could be identified with the proposition
that intervention is desirable, and {t} with the same propositions negation. In this case,
Fig. 1 would represent a situation in which 2 finds it more plausible that intervention is
desirable than that it is not, while 1 would know whether or not this was the case.
More specifically, let a doxastic proposition (henceforth just proposition) be a map
P that assigns to every MPF S with state-space S a subset (P)S S.8 Denote true, ,
false, , and Boolean operations for arbitrary propositions P and Q by
4 Reflexive and transitive binary relation where every non-empty subset has a minimal element,
cf. (Baltag and Smets, 2008).
5 The relation i may therefore more appropriately be thought of as an implausibility relation,
where s i t is read t is as implausible as s, or more so.
6 A heuristic to aid recall is that s < t in form is similar to s t when looking at the arrowhead.
When s i t and t i s, arrowheads are omitted altogether.
7 The notation for information and plausibility cells are adopted from (Dgremont, 2010).
8 Parentheses will be omitted where no confusion should arise.
 ()S := S (P Q)S := PS QS
()S := 0/ (P Q)S := PS QS
(P)S := S\PS (P Q)S := (S\PS ) QS

Propositions with epistemic and doxastic modalities are given by

(Ki P)S := {s S : Ki [s] PS }

(Bi P)S := {s S : Bi [s] PS }

The Boolean case simply follows the immediate set-theoretic interpretation. Propo-
sitions with epistemic and doxastic operators represent statements of knowledge and
belief: Ki P / Bi P reads agent i knows / believes that P, and (Ki P)S / (Bi P)S are the sets
of states of these propositions given an MPF S. For knowledge, the definition entails
that s (Ki P)S iff for all t in is information cell relative to s, t is a P-state. Belief has
the same reading, but restricted to the plausibility cell for i at s.9

Example: King or Queen? Assume a situation with two players, a and b, where a has
one card in hand, being either a King or a Queen. Let Q be the doxastic proposition
a has a Queen on hand and K ditto for King. Over F1 , Fig. 1, set (Q)F1 = {s} and
(K)F1 = {t} (hence (Q)F1 = (K)F1 ). In this case, (Ka Q)F1 = {s}, (Kb Q) = 0/ and
(Bb Q)F1 = {s,t}. That is, in state s, agent a knows Q, in no state does b know Q, but
b believes Q in both states. The structure thus models a situation in which a knows
whether she holds a Queen, while b only has a belief this regarding. Whether this belief
is correct depends on which of the two states is the actual. Combined with such a
valuation of propositions, a plausibility frame is called a model. Denote F1 with the
described valuation SQ/K .

Epistemic Plausibility Models. Let a valuation set be a set of doxastic propositions,

considered the atomic propositions. An epistemic plausibility model (EPM) is an MP
frame together with a valuation set , denoted S = (S, i , )iA . For s PS , write
S, s |= P, and say that P is true or satisfied at state s in model S. A pointed EPM S =
(S, i , , s0 )iA is an EPM with a designated state s0 S, called the actual state. Where
s0 PS , write S |= P.

EPMs and Kripke Models. Where epistemic plausibility frames are special instances
of Kripke frames (see e.g. (Blackburn et al., 2001)), epistemic plausibility models are
not special instances of Kripke models.10 However, every EPM S gives rise to a Kripke
model MK . First, let be where the functional nature of each doxastic proposition is
ignored. may then be treated as a set of atomic proposition symbols. Second, define
a valuation mapkk : P(S), assigning to the elements of a set of states from
9 The reason for using this atypical definition of propositions is that it allows us to speak about
the same proposition across multiple models. This is practical as model transformations will
play a large role in the latter. A Kripke model valuation may easily be extracted from set of
doxastic atomic propositions; see (Baltag and Smets, 2008) for details.
10 As opposed to the terminology of (Baltag and Smets, 2008; van Benthem, 2007; Demey, 2011)
where epistemic plausibility models are Kripke models.
the state-space of the underlying frame. Simply let MK and S be based on the same
frame, and let the valuation map kk for MK be given by kPk := PS , for all P .
The alternative definition of EPMs is used as it is natural when dealing with doxastic
propositions rather than a syntactically specified language.

Relevant Propositions and the Initial State. To model the second step of the bystander
effect dynamics for three agents, ten atomic propositions are required. First, use A to
denote that an accident has occurred. This is the basic fact about which the agents must
establish a belief. Second, each agent i A = {a, b, c} must choose one of three actions:
either to intervene, Ii , to observe, Oi , or to evade the scene, Ei . The set of these atoms
is denoted . As the model constructed is temporally simple, the propositions are best
read as agent i has intervened / observed / evaded. It is assumed that no agent can
perform two actions simultaneously, i.e. that Ii Oi = Ii Ei = Oi Ei = 0. / Denote the
set of all doxastic propositions obtainable from and the above construction rules by
Prop .
The initial state (where nothing has happened) may now be represented by the EPM
S0 , Fig. 3.

A Fig. 3. The initial state where nothing

has happened; all atoms are set to false:
s0
(A)S0 = (Ii )S0 = (Oi )S0 = (Ei )S0 = 0/

In this simple state, every agents knows exactly what has transpired so far: nothing. All
propositions are false, everybody knows this, which is again known byVall, etc. Among
others, it is the case that S0 , s0 |= (A I1 O2 E3 ) iA (Ki (A jA K j A)).
V

The end conditions of runs mentioned on page 5 may now formally be specified:
W
identify the end of a run with any EPM satisfying at its actual state either iA Ii or
V
iA Ei , capturing respectively that at least one agent intervenes, or all evade.

3.2 Changing Models: Action Models and Action-Priority Update.

To capture factual and informational changes that occur due to events, a static epistemic
plausibility model may be transformed using an action model, capturing the factual and
epistemic representation of the event, and the action-priority update product. The guid-
ing idea is that an action model encodes the belief and knowledge agents have about
an ongoing event, the information from which is combined with the static model by
taking the two models product: the result is a new static model in which the agents
new information takes priority over that of the previous static model. The present for-
mulation rests on (Baltag and Smets, 2008), with the addition of postconditions, as used
in (van Ditmarsch and Kooi, 2008; Bolander and Birkegaard, 2011). The latter allows
action models to not only change the knowledge and belief of agents, but also effectuate
ontic11 changes, needed when the environment or agents perform actions.
11 Ontic facts are all non-doxastic facts, i.e. propositions that do not contain belief or knowledge
operators.
Action Plausibility Models. A (pointed) action plausibility model (APM)

E = ( , i , pre, post, 0 )iA

is an MP frame ( , i )iA augmented with a precondition map, pre : Prop and

a postcondition map post : Prop such that post( ) = where {, } or
= ni=1 i with i {P, P : P }. Finally, 0 is the actual event.
V

Just as every world in an EPM represents a possible state of affairs, specified by

the worlds true propositions, so every action in an APM represents a possible change.
What change is specified by the pre- and postconditions; preconditions determine what
is required for the given action to take place, i.e. what conditions a world must satisfy
for an action to be executable in that world, and postconditions what factual change the
action brings about.

Example: King or Queen?, cont. Continuing the example above, let now a play her
card face down. Assuming that a knows which cards she is playing, this situation may
again be presented by the MPF F1 in Fig. 1, with s representing the action a plays a
Queen and t a plays a King. The precondition for s, that a is playing a Queen, is Q,
that a has a Queen on hand, and vice versa for t and K. Hence pre(s) = Q and pre(t) =
K. Following the play, a will either no longer have a Queen on hand, or no longer have
a King. Hence, post(s) = Q and post(t) = K. With these pre- and postconditions,
F1 is an action model, call it EQ/K , representing the event where a is certain that she
is playing a Queen, while b is uninformed about which of the two plays is the actual,
finding it more plausible that a plays Q.

Doxastic Programs. Where EQ/K represents a situation where a plays her card face
down, a show-and-tell play by a is captured by the two strict subsets of the model, i.e.
by the doxastic programs Q = {s} and K = {t}. A doxastic program is the action
model equivalent of a proposition, i.e. a subset of all actions in the models event space:
. Over EQ/K , the program Q captures the event where a plays Queen and b sees
this, and K the same for a playing King. In the ensuing, it will be assumed that doxastic
programs contain the actual action.

An Accident Occurs. What is a suitable APM capturing both the factual change that the
accident occurs, as well as a, b and cs information about this? Given that the accident
in fact occurs, it is clear that the actual event 0 of the model must change the truth
value of A from false in S0 to true in the ensuing EPM S1 . Further, no agents perform
actions during the event, so post(0 ) = A.
Focusing on a, then how does she perceive the occurrence of accident? As [m]ost
emergencies are, or at least begin as, ambiguous events (Latan and Darley, 1968,
p. 216) a will at least be uncertain regarding whether it occurs or not, and therefore
considers an alternative event 0 with post(0 ) = possible.12 Moreover, ex hypothesi,
her perception of the event indicates that in fact A, so 0 a 0 . How does a perceive
12 The postcondition leaves all atomic propositions as they were in the previous model. This is
specified by the action priority update product below.
that b and c perceive the accident? Not being telepathic, a cannot tell, and she considers
it possible that both, neither, or either of b and c perceive the event as she does. It
is assumed, though, that a perceives b and c during the event as forming an opinion
about whether or not A. This assumption is made for two reasons: 1) it produces a
smaller model, and 2) pertaining to social proof, only agents perceived as informed are
interesting from as point of view. Variations to this assumption would be interesting,
but are not dealt with here. Finally, a must consider it possible that b and c are wrong
about the way a perceives the event. Taking this into consideration, Fig. 4 illustrates as
doxastic perception of the accident.

0
a abc a
Fig. 4. Agent as perception of the accident, 0 . All k
1 0 3
actions have post(k ) = A; for all k , post(k ) = .
ab c a a ac b Some links are gray only for presentation; they mirror
1 2 3 the labels above. Diagonal links implied by transitivity
a bc are omitted; so are many links for b and c (see Fig. 5).
2
Agent a considers it more plausible that an accident is occuring, but is (at this point) agnostic
about the beliefs of her peers; she finds it possible they all agree (0 , 0 ), that she agrees with
only b (1 ,1 ) or with only c (3 ,3 ), or that both b and c perceive the ongoing event as non-
hazardous (2 ,2 ). Finally, she cannot rule out that b and c both find the event unproblematic
and that they perceive a as doing the same (7 ,7 ).

Assuming that b and c perceive the accident in an identical manner, the model in Fig.
3 may be suitably duplicated and combined, resulting in the joint model E0 , Fig. 5.
Notice that b and cs perceptions are identical to as. The only states not obtained from
a duplication of Fig. 3 are 7 and 7 . In fact, no one considers these possible, but neither
can anyone rule out that others entertain them.
 5
ab bc a ac

6 bc 5 4
b ac b c c ab

bc 6 0 bc 4
ac ac abc ab ab

1 a 0 3
ab c ac ab ac b

1 2 3
a bc

bc
b 2

7
abc

7
Fig. 5. The EPM denoted E0 , representing the joint perception of the occurrence of an accident
for three agents a, b and c. All k actions have post(k ) = A; for all k , post(k ) = . Grey
links are only for presentation; they mirror the labels above. Links implied by transitivity are
omitted, e.g. between 6 and 4 for a. In essence, E0 captures 1) that the accident in fact occurs,
2) that this is ambiguous for every agent, who all find it more plausible that it does occur and
3) that no agents learns anything about others perception of the event, except that all have an
opinion as to whether or not the accident occurred.

To incorporate the (new) information from an action model or a doxastic program

in an EPM, the two must be combined. A natural procedure for doing so is the action-
priority update product (Baltag and Smets, 2008).

Action-Priority Update Product. The action-priority update is a binary operation

with first argument an EPM S with relations i and second argument a doxastic program
over some APM E with action space and relations i . The APU product is an
EPM
S = (S , i , , (s0 , 0 ))
where the updated state space is S = {(s, ) S : S, s |= pre( )}; each updated
pre-order i is given by (s, ) i (t, ) iff either i and s i t, or else i and
s i t;13 the valuation set is identical to , with the requirement that for every atom
P ,

PS = {(s, ) : s PS and post( ) 6|= P} {(s, ) : post( ) |= P}

for states (s, ) S . Finally, (s0 , 0 ) is the new actual world.

13 i is from E and i from S. i denotes ( i and not i ), i denotes ( i
and i ).
 The APU product gives priority to new information encoded in over the old be-
liefs from S by the anti-lexicographic specification of i that gives priority to the
APM plausibility relation i . The definition further clarifies the role of pre- and post-
conditions; if a world does not satisfy the preconditions of an action, then the given
state-action pair does not survive the update, and if postconditions are specified, these
override earlier ontic facts, else leave all as was.14

Example, concl.: King or Queen? The factual and doxastic consequences of as play
of the Queen is calculated by finding the APU product of SQ/K and EQ/K . The result is
(again) an EPM SQ/K EQ/K with underlying frame F1 , with state space {(s, s), (t,t)}.
The state (s, s) survives as s from SQ/K satisfies the preconditions of s from EQ/K ,
while (s,t) does not, as pre(t) = K while (K)SQ/K = {t}. Further, in the actual state, Q
changes truth value from true in s SQ/K to false in (s, s) SQ/K EQ/K as a result
of the postconditions of s EQ/K . Finally, the information of the agents have changed:
e.g., b now knows that a does not hold a Queen ((s, s) |= Kb Q).

Updating with the Accident. Updating the simple initial state model S0 with accident
event model E0 results in a state of great uncertainty. In summary, every agent believes
that A: an accident has occurred (though no-one knows), all know that their peers either
believe or disbelieve A (none are indifferent between A and A states), and all consider
it possible that the others consider it possible that all believe there is no accident.
Formally, updating the simple structure of S0 with E0 produces the EPM S1 :=
S0 E0 which shares frame with E0 (Fig. 5) and has (s0 , 0 ) as actual state. In S1 ,
for all i {0, ..., 7}, (s0 , i ) AS1 , and (s, i ) (A)S1 . Among others, the following
doxastic propositions are true at (s0 , 0 ): A, iA Bi A, Ka (Bb A Bb A), Ka Bb A.
V

Based on this second static state, the agents must make their first decision, as spec-
ified by their decision rules, which jointly determine the type of agent they are.

4 Decisions: Agent Behavior Characterized by Transition Rules

Though agent action may be represented in the introduced DEL framework using suit-
able atoms and postconditions, the notion of agency in DEL is purely doxastic. To
move from only believing agents to acting agents, a richer framework is called for.
One possibility would be to introduce a game- or pay-off structure in parallel to the
DEL framework or embed the entire dynamics modeled in a temporally extended game
tree, whereby actions could be made rationally, based on utility maximization at end
nodes. A drawback to this method is the large models required: every branch must be
fully specified before decisions may follow. Further, considering all possible branches
is a cognitively complex task, making the approach empirically unrealistic.
Instead, an alternative approach involves utilizing rule of thumb decisions, brute-
forced by the current beliefs of agents. This method, detailed below, forfeits rational
decisions, but overcomes the two drawbacks of the game theoretic approach by letting
14 The definition is based on (Baltag and Smets, 2008) for the anti-lexicographic order, adding
postconditions from (van Ditmarsch and Kooi, 2008; Bolander and Birkegaard, 2011).
choice be dictated locally by current beliefs. Given an EPM, a set of doxastic programs
provides a multitude of possible updates. In modeling a dynamic process, the modeler
must choose which model is suitable for the next update, based on no strict directions
from the to-be-updated EPM. However, environment or agent behavior will often be
seen as dictated to some degree by facts or beliefs from the current EPM, thus used as
a guideline. To incorporate the next action model choice in a formal manner, transition
rules are introduced, locally specifying the next update as a function of the current
EPM.
Transition rules are used to characterize agent behavior. Each behavior is specified
by a set of transition rules, each with a trigger condition and a goal formula. If an
EPM satisfies some trigger conditions, the ensuing EPM must satisfy the matching goal
formulas. An APM that ensures that the goals are obtained then satisfies, or solves, the
rules, and is seen as a possible choice for the agent in question. Hereby an EPM, a set
of behavior-governing rules and a set of APMs jointly specify the transition to the next
EPM.

Transition Rules. A transition rule T is an expression [X] where , Prop .

Call the trigger and the goal. If EPM (S, s0 ) satisfies the trigger of a transition rule
T , T is said to be active in S (else inactive).
Specified below, transition rules may be used to choose the next update based on
local conditions of the current EPM. E.g., updates by the environment may be spec-
ified using atoms in the trigger. To exemplify, let R and W be atoms with resp. read-
ings it rains and the street is wet, then the transition rule T1 = R [X]W reads if
it rains, then the next update must be such that after it, the street is wet. Transition
rules may also be used as agent decision rules for factual change, using Bi /Ki -
formulas as triggers and suitable formulas as effects. E.g., the set of transition rules
{Bi R [X]Ui , Bi R [X]Ui } may be used to specify agent behavior relative to rain:
if i believes it rains, then next i will have an umbrella, and if i believes it does not
rain, then next i will not have an umbrella. Used thus, transition rules are akin to the
programs and knowledge-based programs of (Fagin et al., 1995), here tailored to the
DEL framework. They further instantiate one-step epistemic planning problems, in the
terminology of (Bolander and Birkegaard, 2011).

Dynamic Modalities. Note that transition rules are not doxastic propositions: the
modality [X] has no interpretation, and construed as a formula, T 1 has no truth con-
ditions. Instead, transition rules are prescriptions for choosing the next action model.
The choice of model is made by implementing a transition rule over an EPM S and a
set G of doxastic programs over one or more APMs using dynamic modalities.
For any program over APM E, [ ] is a dynamic modality, and the doxastic propo-
sition [ ] is given by

([ ])S := {s S : , if (s, ) S then (s, ) S }.

That is, a state s from S is a [ ]-state iff every resolution of over s is a -world in
S . [ ]-modalities are natural when is desired common knowledge among A.
 Further, where a is doxastic program, let [ ]i be given by

([ ]i )S := {s S : , if (s, ) S Ki [(s0 , 0 )] then (s, ) S }.

That is, a state s from S is a [ ]i -state iff every resolution of over s that is included
in is information cell relative to the actual world (s0 , 0 ) in S is a -world in
S . Hence S, s |= [ ]i iff S , (s0 , 0 ) |= Ki .
The [ ]i -modalities are natural when transition rules prescribe prescribe agent choices,
as they ensure that the performing agent knows her choice following the action, while
allowing others to be unaware of the choice made.

Solutions and Next APM Choice. A set of transition rules dictates the choice for the
next APM by finding the transition rule(s)s solution. A solution to T = [X] over
pointed EPM (S, s) is a doxastic program such that S, s |= [ ]. is a solution
to the set T= {T1 , ..., Tn } with Tk = k [X]k over (S, s) if S, s |= n1 (k [ ]k ),
V

i.e. if is a solution to all Ti over (S, s) simultaneously.15 Finally, a set of doxastic

programs G is a solution to T over S iff for every t of S, there is a G such that is
a solution to T over (S,t).16
If G is a solution to T over S, then given a state from S, the transition rules in T will
specify one (or more) programs from G as the next choice. A deterministic choice will
be made if G is selected suitably, in the sense that it contains a unique for each s. In
the ensuing, solution sets will be chosen thus.

Example: Looping System. Consider the very simple system, consisting of an EPM
S with s0 PS , and APM E with pre(0 ) = P, post(0 ) = P, and the set T= {T0 , T1 }
of transition rules:

T0 = P [X]P s 0 1
S: 0 E:
T1 = P [X]P P hP; Pi hP; Pi

With 0 = {0 } and 1 = {1 }, G = {0 ,1 } is a solution to T over S. For T1 , S, s0 |=

P [0 ]P as s0 6 (P)S . For T0 , it is easy to check that S , (s0 , 0 ) |= P, entailing
that S, s0 |= [0 ]P. As 0 is unique, this is chosen as next update. It should be easy to
see that G is also a solution to T over S 0 , where 1 is chosen. Further re-application
of T loops the system.17

Three Agent Types. Transition rules may be used to provide general characterizations
of agent behavior determined by belief. Rules with a doxastic trigger will be referred
to as decision rules, by sets of which an abundance of possible agent types may be
defined. Of interest are the following three, corresponding to three types of human
behavior relevant to the bystander effect.
15 Note the analogy with numerical equations; for both 2 + x = 5 and {2 + x = 5, 4 + x = 7}, x = 3
is the (unique) solution.
16 The definition is altered to suit transition rules using [X]i modalities by suitable replacing
[X] with [X]i and [ ] with [ ]i throughout.
17 For a definition of system, see (Rendsvig, 2013a).
 First Responder: City Dweller: Hesitator:
Bi A [X]i Ii Bi A [X]i Ei Ki A [X]i Ii
Bi A [X]i Ei Bi A [X]i Ei Bi A Ki A [X]i Oi
Bi A [X]i Ei
Table 1. Decision rules specifying three agent types. Denote by F1i and F2i resp. the upper and
lower first responder rule indexed for i, and set Fi := {F1i , F2i } and treat C1i , C2i , H1i , H2i , H3i ,
Ci and Hi in a similar manner.

The First Responder will intervene if she believes there is an accident, otherwise not.
First Responders thus reflect the normally expected, but not witnessed, behavior in re-
lation to emergencies. A City Dweller will evade the scene no matter what his beliefs,
hereby reflecting the medias grim picture of the apathetic urban citizen, ignoring the
murder of Kitty Genovese. Finally, a Hesitator will choose to observe if she believes but
does not know that there is an accident, and will else evade. Hereby the Hesitator rules
capture (part of, see below) the behavior used as explanation for the bystander effect
(e.g. by Latan and Darley (1968) when they write it is likely that an individual by-
stander will be considerably influenced by the decisions he perceives other bystanders
to be taking (p. 216)). Presently, focus will be on Hesitators, with comments on First
Responders and City Dwellers. The latter two are subjects of Section 6.

Possible Choices. To implement either of the rule sets, a suitable set of doxastic pro-
grams for X to range over must be specified. It seems natural to assume that when an
agent is intervening, then this is epistemically unambiguous for all agents. When b and
c see a choose either to observe or evade, it seems more plausible that they cannot tell
these actions apart, as neither action has an observable, distinguishing mark.18 It is as-
sumed that agents find it more plausible that others evade than that they observe.19 In
sum, these consideration give rise to the APM E1i of Fig. 6.

i : i : Fig. 6. The APM E1i , representing the three

A i i
moves available to i as well as the doxastic per-
Ii Oi Ei
A\{i} ception of these for the remaining agents. State
names specify postconditions; all preconditions are . If i chooses to intervene, i will be next
APM choice, whereas if i chooses to either observe or evade, i will be used, with actual event
resp. Oi or Ei .

18 Unless evading entails leaving the scene or observation is performed in a non-discrete manner.
Often this is not the case, though: Among American males it is considered desirable to appear
poised and collected in times of stress. ... If each member of a group is, at the same time,
trying to appear calm and also looking around at the other members to gauge their reactions,
all members may be led (or misled) by each other to define the situation as less critical than
they would if alone. Until someone [intervenes], each person only sees other nonresponding
bystanders, and ... is likely to be influenced not to act himself. (Latan and Darley, 1968, p.
216); ... Apparent passitivity and lack of concern on the part of other bystanders may indicate
that they feel the emergency is not serious, but it may simply mean that they have not yet had
time to work out their own own interpretation or even that they are assuming a bland exterior
to hide their inner uncertainty and concern. (Latan and Rodin, 1969, p. 199).
19 The second quote in the previous note seems to indicate the plausibility of this assumption.
E1i does not facilitate simultaneous choice, in the sense that it does not contain a solu-
tion to e.g. {H1a , H1b } over S1 . Combining, however, a copy of E1i for each of a, b and
c while respecting the doxastic links in an intuitive way may easily be done. Specifi-
cally, a combined APM E1 may be obtained by taking the reflexive, transitive closure of
the Cartesian graph product E1a E1b E1c (see e.g. (Hammack et al., 2011) for defi-
nition) and specifying pre- and postconditions as follows: for (s,t, u) E1a E1b E1c ,
let pre(s,t, u)E1 = and post(s,t, u) = post(s)E1a post(t)E1b post(u)E1c . The re-
sulting APM E1 has eight mutually disconnected components of four types, see Fig.
7.
EEE

jk ij
IEE
i ik
ik ij
EII OEE j EEO
k EOE i
A
jk IOE IEO ij ik
III i jk
OII ij OEO k EOO
ik
i OOE j
i
IOO ij jk
ik

OOO

Type I (1x) Type II (3x) Type III (3x) Type IV (1x)

Fig. 7. The APM E1 , representing simultaneous move programs, omitting (most) reflexive and
transitive arrows; states are labeled with postconditions, with IOE representing h; Ii O j Ek i,
etc.; all preconditions are . Notice the rise in dimensions; Type I is a point, Type IV is a cube.

In Fig. 7, the Type I model is obtained by the Cartesian product of i ,j and k ; Type
II from j ,k and i ; Type III from i and j , k and Type IV from i , j , k . Notice
that reflexive and transitive closure is required to ensure that all i s are pre-orders.
Euclidean closure is not required, but is doxastically reasonable: in e.g. the Type III
component, i should not be able to distinguish between IOE and IEO, nor consider
either more plausible.
Doxastic programs over E1 identical to each of the four sub-model types give rise
to the desired solution space E . Let E be the set consisting of all pointed E1 models,
and let E be the set of all doxastic programs over models in E such that each E
contains exactly one of the type I-IV sub-models. Then E contains a unique solution to
every combination of the three agent types. (recall, def. of solution set: for every s S.)
Over S1 and E , the Type I program is the unique solution to F1 F2 F3 ; the
Type II program with i = 2 and 0 = EII is the unique solution to C1 F2 F3 with
0 = OII, it is the unique solution to H1 F2 F3 ; the Type III program with i = 1, j =
2, k = 3 and 0 = IEO is unique solution to F1 C2 H3 ; the Type IV program with
0 = OOO is unique solution to H1 H2 H3 . Interestingly, the Type IV program with
actual state OOO reflects an implicit norm-based pluralistic ignorance: though every
agent is observing, each perceive the situation as one where they are the only ones
doing it.

Choosing to Observe. Given the above, the system based on S1 with H1 H2 H3

as rules over E1 dictates the type IV program with 0 = OOO as next APM choice.
That is, every agent chooses to observe. The ensuing EPM S2 := S1 IV contains 128
states, but is easily described. Take IV and replace every state with a complete copy of
S1 s frame, connect two states from two different copies ((s, 1 ) i (s , 2 )) iff s = s
and 1 i 2 , and finally take the reflexive-transitive closure. The new actual world is
((s0 , 0 ), OOO), satisfying
^ ^
A, Oa Ob Oc and (Ki Oi Bi Ei ).
iA iA\{i}

The latter captures the post-factual effects of the mentioned norm-based pluralistic ig-
norance of IV : all have the belief that they individually were the only ones to observe,
while all others evaded. Importantly, had EPM E1i on page 15 been such that agents
perceived actions according to their own decision rules, i.e. found observation more
plausible than evasion in others, all would have correct beliefs about the actions of
others.

5 Action Interpretation and Social Proof

Albeit all agents have formed the belief that their co-witnesses evaded, none has formed
any beliefs rationalizing these choices: All are still doxastically indifferent between
whether the others believe A or A. Neither does any agent have means of deducing
others beliefs, given the introduced formal framework. Such a deduction would re-
quire e.g. the ability to rationalize by forward induction, which requires information
from both past play and future possibilities (?) couched in a game framework repre-
senting preferences, rationality, etc. (see e.g. (Rendsvig and Hendricks, 2013) for an
implementation). Though structures akin to game trees may be defined using EPMs,
APMs and protocols (van Benthem et al., 2007; Dgremont, 2010), a simpler, more
superficial construct may be used to facilitate the reasoning. The suggested approach
utilizes an inverse version of decision rules, brute forcing conclusions about belief
from observations about action.
In making decisions, our beliefs about the relevant state of affairs dictate our ac-
tion, up to error and human factors. Hence the route from beliefs to actions is often
functional. As this function will often not be injective, moving from actions to beliefs
is not as straightforward, since multiple different belief states may result in the same
action. Having to provide a rationalization of a given action will therefore often include
abductive reasoning. An abductive hypothesis to rationalize an action allows inferring
as explanation of the observed action a previous belief state of the acting agent. Below,
such hypotheses are called interpretation rules.

Interpretation Rules. An interpretation rule is a doxastic proposition [S]Bi ,

with called the basis and the content, with the underlying idea that on the basis of
an action (e.g. Ei ), agents may deduce something about the content of is beliefs (e.g.
that Bi A).
Doxastic propositions involving the modality of the consequent are given by

([S])S := {s S : s S such that s s and (S, s) |= },

where S is the domain of S, and where s s means that s is a predecessor20 of s . Hence
[S] is true in (S , s ) just in case s s predecessor in S was a -world. The modality is
included to respect the temporal aspect introduced by updates, and S is to be substituted
with the EPM based on which i made the choice in question.
A set of interpretation rules may in general be implemented using an APM where
the preconditions of each state is a conjunction of interpretation rules with different
bases with a conjunct for each action to be interpreted. Hereby each state represents a
different hypothesis regarding the acting agents type, i.e. how the agent made decisions.
The plausibility order then specifies the abductive hierarchy of such hypotheses.21
To simplify, agents are given only one hypothesis about types, the hypothesis also
being correct in the sense that the interpretation rules are (close to) the converse of the
transition rules that are in fact applied. Hence the interpretation rule model APM E2i j
that determines how agents A\{ j} interprets the actions of j is a one state model. Let
j E2 j and set

pre( j ) := E j [S1 ]B j A
I j [S1 ]K j A
O j [S1 ]B j A K j A

Applying such rules for all agents may be done by sequential application of E2 j for
each j {a, b, c} on S2 (and the resulting models).22 Call the APU product S3 .

Belief-based Pluralistic Ignorance. By application of interpretation rules, agents es-

tablish beliefs about each others previous beliefs, but even though the applied interpreta-
tion rules were correct, the obtained beliefs are wrong: S1 |= Bb A, but S3 |= Ba [S1 ]Bb A.
This is a direct consequence of the mis-perception of actions occurring in E1 .23
In the actual world of S3 , the Vagents are in a state of belief-based pluralistic ig-
norance with respect to A: S3 |= iA (Bi A jA\{i} Bi B j A), cf. the definition in
V

(Hansen, 2011, ch. 6). To see this, notice what happens when a interprets the ac-
tions of b and c over S2 . In S2 , the most plausible copy of S1 (Fig. 4) is the one in
which all states satisfy Oa Eb Ec (Fig. 7). Of these 16 states, only 4 satisfy both
pre(b ) and pre(c ), namely the successors of 2 , 2 , 7 and 7 , and only the first is
in as plausibility cell relative to (s0 )S2 . As all other states in this S1 -copy are deleted
upon update with E2b ; E2c , it follows that as plausibility cell Ba [(s0 )S3 ] contains only
the state ((((s0 , 2 ), OEE), b ), c ), which in turns satisfies A Bb A Bc A. Hence
20 When constructing APU products, a state in the product model is an ordered pair (s, ) of a
state s and and action . In this pair, s may again be such a pair. Say that a predecessor of s is
any s that occurs in any of the ordered pairs of s , including s itself.
21 It is possible to give agents a choice of interpretation by invoking transition rules with interpre-
tation rules as possible solutions. In the present, agents are given no choice of interpretation,
and this construction is consequently skipped for simplicity.
22 Each E2 j functions as a truthful public announcement of pre(), for which the order of an-
nouncements does not matter (Baltag and Smets, 2009): states are deleted, the remaining or-
derings staying as previous. Deleting simultaneously or in some sequence makes no difference.
23 Though time has passed, beliefs have not changed, and this is known to all:
S3 |= iA Ki ( jA [S1 ]B j B j ) for {A, A}.
V V
S3 |= Ba (A Bb A Bc A). Analogous reasoning for b and c shows that (s0 )S3 is a
state of pluralistic ignorance w.r.t. A.
Again, importantly, had EPM E1i been defined so that agents considered obser-
vation more plausible than evasion, this state of pluralistic ignorance would not have
arisen.

Social Proof. In the portrayal of the bystander effect, witnesses alter their beliefs fol-
lowing their mutual act of orientation, and in the light of the newly obtained information
that no one else believes that there is cause for alarm, concludes that no intervention is
required. To represent the revised beliefs of agents24 , introduce a new operator SBi|G ,
representing the beliefs of agent i when socially influenced by her beliefs about the
beliefs of agents from group G. SBi|G is defined using simple majority voting with a
self-bias tie-breaking rule: let

s (SBi|G )S iff + |{ j G : s (Bi B j )S }| > + |{ j G : s (Bi B j )S }|

with tie-breaking parameters , given by

( (
if s (Bi )S if s (Bi )S
= =
0 else 0 else

This definition leaves agent is social beliefs w.r.t. undetermined (i.e. (SBi|G
SBi|G )) iff both i is agnostic whether and there is no strict majority on the matter.
Applying the notion of social belief to A in S3 , it is easily seen that S3 |= iA SBi|A A.
V

That is, upon incorporating social proof, all agents socially believe, contrary to their
private beliefs, that no accident occurred.

Action under Influence. Notice that none of the three agent types introduced so far
will change their action if presented again prompted to intervene, observe or evade.
First Responders will again intervene, City Dwellers will again choose to evade, and
Hesitators will again, irrespective of social proof, choose to observe.
To make Hesitators pay heed to the observation they chose to make, their decision
rules are changed (in the ensuing section, a fusion of the two types is defined). Let an
influenced agent act in accordance with the following rules:

Influenced:
SBi|G A [X]i Ii
SBi|G A [X]i Ei

Note that an Influenced agent acts like a First Responder who bases her actions on
social beliefs.
24 Strictly speaking, in the present model agents do not revise their beliefs. An additional operator
is instead introduced to facilitate comparison with private beliefs. A belief revision policy may
easily be defined using decision rules to the effect that agents update their beliefs under the
suitable circumstances, see (Rendsvig, 2013b).
 A Hesitator-now-turned-Influenced presented with the choice to intervene, observe
or evade (as given by E ) will choose to evade. More precisely, if a, b and c are Influ-
enced, the unique next APM choice will be the Type IV program with 0 = EEE. The
actual world in the ensuing EPM S4 , the final step of the model, will then satisfy
^ ^ ^
A Bi A SBi|G A Ei .
iA iA iA

The last conjunct is an (unfortunate) end condition, as specified on page 8. Hereby,

informational dynamics leading to an observable bystander effect has been modeled.

6 Comparison to Empirical Studies

The presented sequence of models, transition rules and updates conjoined captures im-
portant informational aspects of the observable bystander effect, given that the model
is accepted. As presented, the sequence may be regarded as one possible execution of
a broader, implicit system. Other runs of this system may be constructed by varying
parameters relevant to the bystander effect.
In this section, the effect of changing two parameters will be presented. The first
change is of agent types, where it is seen that for non-mixed groups, both City Dwellers
and (Influenced) Hesitators will produce the observable bystander effect. The second
variation is group size, and it is shown that of non-mixed populations, only (Influenced)
Hesitator behavior varies as a function of group size.
Let us briefly outline the implicit system before changing parameters. The system
has initial state S0 on page 8, where everybody knows nothing has happened and end
W V
conditions either iA Ii or iA Ei . S0 is updated with the occurrence of the accident,
E0 on page 9 resulting in S1 on page 12, where all believe an accident has occurred,
while having no information about others beliefs. Apart from adding further agents to
the population, these steps will remain fixed.25 Next, agents make a first decision over
E and S1 is updated with the next APM choice.26 Depending on agent types, the run
might end at S2 . If not, the interpretation rule model on page 18 is applied for all agents,
and a second decision is made based on the outcome, possibly involving the aggregation
of the perceived beliefs of others. Again, if the system does not satisfy one of the end
conditions, it will continue, in which case the interpretation rule model is re-applied
(suitably altered to accommodate the temporal shift), followed by decisions, etc.
In the run described in the previous sections, two different agent types were used.
For the first choice made, agents were assumed to be Hesitators, making them choose
to observe. For their second choice, they were assumed to be Influenced, making them
act on their social beliefs. 27 To facilitate comparison of models, this mixed type may
be properly defined as Influenced Hesitators:
25 Concerning E0 , it should be obvious how the APM must be altered to include further agents,
while maintaining complete higher-order ignorance.
26 Again, it should be obvious how E may be altered to accommodate for a larger population.
27 The shift was made to ease the exposition. Influenced agents require the notion of social be-
liefs, not necessary for Hesitators first choice.
City Dweller: Hesitator: Influenced Hesitator:
Bi A [X]i Ei Ki A [X]i Ii Ki A [X]i Ii
Bi A [X]i Ei Bi A [X]i Ei Bi A [X]i Ei
Bi A Ki A [X]i Oi Ki A (SBi|G A SBi|G A) [X]i Oi
First Responder: Oi SBi|G A [X]i Ii
Bi A [X]i Ii Oi SBi|G A [X]i Ei
Bi A [X]i Ei

Table 2. Re-specification of agent types.

Notice that Influenced Hesitators behave as a mixture of Hesitators (first three rules)
and First Responders (last two), but who take social proof into account. Notice the
difference between third rule for Hesitators and the same for Influenced Hesitators. The
latter requires that Influenced Hesitators have undetermined social beliefs before they
choose to observe. The altered First Responder rules (rows four and five) capture that
if the agents has observed and have determined social beliefs, observation gives way to
intervention or evasion.28

Varying Types and Group Size. The table below summarizes the end conditions and the
EPM number where they arise relative to agent types and group size. For end conditions,
W V
Ik and Ek represent iA Ii and iA Ei respectively being satisfied in EPM Sk , with k
rising as described in the previous sections. means that no end conditions are met.

1 2 3 k4
FR I2 I2 I2 I2
CD E2 E2 E2 E2
H
IH I3 I3 E3 E3

Table 2. Summary of end conditions and times as a function of agent type and
group size.

In words, for any group size, First Responders will intervene in S2 , i.e. immediately
following the accident. At the same time, City Dwellers will evade the scene, no mat-
ter the group size. Simple Hesitators, as defined in Table 2, will never reach an end
conclusion, as they will never come to either believe there is no accident, or know that
there is one. That all these three agent types actions are invariant over group size is
due to their inherently non-social decision rules. The social, or Influenced, Hesitators
will however change their behavior according to group size: they will intervene imme-
diately if the group size is small enough for their private belief not to be overridden by
social proof. If the group size is 3 or above, Influenced Hesitators will conclude, by the
mis-perception of others choice to observe as an act of evasion and the resulting state
28 The requirement that i must have observed before acting on social beliefs ensures that agents
do not intervene immediately after seeing the accident (a private belief that A would imply that
SBi|G A, as agents then hold no beliefs regarding others beliefs).
of pluralistic ignorance, that enough agents believe that no accident occurred for them-
selves to be socially convinced that this is the case. Consequently, they will choose to
evade in S3 for any group size of 3 or above.
At group size 2 these agents still decide to intervene because they use their own be-
lief to break the tie between what they perceive as an even split on whether an accident
is happening.29

Comparison to Empirical Studies. Running the system with each of the four agent
types may be considered as providing four different models of the bystander effect,
each of which may be compared to empirical results to evaluate consistency with data.
Table 2 allows for only a simple comparison, checking whether end conditions as a
function of group size correlates properly with the observed.
A wide variety of studies have been performed on the inhibiting effect of the pres-
ence of others in situation requiring intervention (see e.g. (Latan and Nida, 1981) for a
meta-study). Many of these have different, more specific foci, e.g. the role of diffusion
of responsibility, friendship, gender, and more. As the focus of this paper is the second
step of the bystander effect, only studies on the effect of social proof on the perception
of the accident are relevant. Alas, no one such has been found that provides suitable
data for all group sizes. Comparison can therefore only be made using multiple studies
invoking different experimental settings.
Inherently, the presented models are deterministic, while experimental data provides
information about the percentage of individuals who intervene. Given this, data will not
be correctly matched. To evaluate tendency of correctness, acceptance or rejection of
models are therefore based not on strict consistency, but on the weak requirement that
a model must correctly match the binary experimental end conditions in strictly more
than 50% of cases.

Smoky Room and the Rejection of FR, CD and H. The classic smoky room exper-
iment of (Latan and Darley, 1968), specifically designed to test the hypothesis of the
second step of the explanation of the bystander effect (see p. 2), has served as a strong
guide for the construction of the models, and provides data which allows the rejection
of three of them. In the study, groups of size 1 or 3 where sat in a waiting room, com-
pleting questionnaires. The groups of size 3 either consisted of 1 individual naive to
the experiment and 2 of the experimenters confederates, or 3 naive subjects. While
completing the questionnaire, smoke was introduced to the room through a visible vent,
ambiguously indicating either an emergency (e.g. fire) or not (e.g. steam). As the possi-
ble accident will have dire consequences for the subjects themselves, the experimenters
assumed that no diffusion of responsibility arose.
The experiment was stopped when either one agent intervened, or after six minutes
of smoke introduction and questionnaire completion, at which point smoke was heavy.
W V
Compared to the model, these end conditions are identified with iA Ii and iA Ei ,
respectively.
Of the subjects that were alone, 75% reported the smoke, a number high enough
to warrant the rejection of the City Dweller model, which would have it that all evade.
29 Cf. the tie-breaking rule used in the definition of social beliefs.
Likewise, the Hesitator model is rejected, as it would have it that individuals would
continuously observe (instead of completing their questionnaires). Both the First Re-
sponder and Influenced Hesitator models score better than 50%.
With 2 confederates in the room, only 1 in 10 naive subjects intervened. With 75%
of individuals intervening when alone, it should be expected that 98% of groups of size
3 with three naive subjects would intervene if individuals acted independently,30 but
only 38% of the 8 groups did so.
The First Responder model is 10% correct in the confederates condition31 and 38%
in the naive group condition, hereby falling below the 50% mark. The Influence Hes-
itators model does better: it is 90% correct of the confederates condition32 and 62%
correct in the naive group condition. Hence, it fairs better than 50% overall.
Interestingly, the response time for intervention in the three naive subjects case was
considerably longer than the single subject case, indicating 1) that individuals in groups
did pay heed to social proof before acting, and 2) that in many cases (38%), social proof
from only two peers was not enough to preclude intervention.

Is One Additional Witness Enough for Intervention Inhibition? The smoky room
study only compares groups of size 1 and 3, whereby it does not supply sufficient data to
evaluate the IH model for group sizes of 2 or above 3. To evaluate the model for group
size of 2, another classic experiment may be consulted, namely the lady in distress
case of (Latan and Rodin, 1969). Three conditions where tested in this experiment:
with a lone, naive subject, with one naive subject and one confederate, and with two
naive subjects, again filling out questionnaires. In a simulated accident, the female in-
terviewer faked a fall in an easily accessible and audible adjacent room. The fall was
indicated by a loud crash, a scream and subsequent moaning of complaints and hurt.
Contrary to the smoky room, this accident is ambiguous between either a serious acci-
dent (e.g. broken leg) or a not-so-serious one (weakly sprained ankle).
In the first condition, 70% of the alone subjects intervened, with a strong drop to 7%
when an inactive confederate was introduced. With two naive subjects, 91% of groups
would be expected to intervene if subjects acted independently, whereas only 40% of
such groups in fact did so.
These percentages strongly contradict the IH model for group size 2, as the expec-
tation is that neither of the two agents would be sufficiently influenced by each other to
not act on their private beliefs. Both would therefore intervene following observation.
This makes the model incorrect in 60% of cases, making it worse than a random bet.
Partly, the model may misfire as the experiment does not conform to the plural-
istic ignorance explanation of the bystander effect. Specifically, the experiment does
not preclude the possibility of a mix of social proof and diffusion of responsibility ef-
fects, given that the accident in question did not put the subjects in faked danger. An
experiment precluding diffusion effects may be conjectured to show a higher degree of
intervention, yielding a better fit.
30 See (Latan and Darley, 1968) for calculation of hypothetical baseline based on the alone
condition.
31 In a mixed population model, using City Dweller agents for the two confederates.
32 Again in a mixed population model, using City Dweller agents for the two confederates.
 The meta-study (Latan and Nida, 1981) strongly indicates that determining social
influence occurs in groups of size 2. Summarizing 33 studies with face-to-face inter-
action, the effective individual probability of helping was 50%, with an effective indi-
vidual response rate in groups only 22%. Most of these studies involved groups of 2.33
Hence, it seems that the model misfires when it comes to groups of size 2.
The obvious parameter to tweak for a better fit is the self-biased majority voting
definition of social beliefs, which does not put enough weight on the other in the 2
person case. Changing this to one favoring the perceived beliefs of the other would
yield a better model for the 2 subject case, while it would not alter the results for the
group size 3 case. Table 3 summarizes the effect of the Influenced Hesitator model run
using a tie-breaking rule favoring the opposite belief of ones own.34

1 2 3 k4
IH I3 E3 E3 E3

Table 3. Summary of end conditions as a function of agent type and group size
for others-biased social beliefs.

Though this model does not fair very well on the data from the smoky room and
lady in distress studies, it does at least do better than a random bet.

7 Conclusions

Accepting the individual modeling steps as reasonable explications of the reasoning

steps occurring in bystander effect-like scenarios, the constructed dynamics allows for
a number of conclusions about such phenomena:

The sub-optimal choice for all to evade is not a consequence of apathetic agents:
City Dweller evasion is not influenced by group size.
The sub-optimal choice for all to evade is a direct result of considering social proof
in a state of proposition-based pluralistic ignorance: Influenced Hesitators with cor-
rect beliefs about their peers beliefs would choose to intervene.
Proposition-based pluralistic ignorance arises due to norm-based pluralistic igno-
rance: that all agents assume others are evading when they themselves are observing
is a necessary condition for the state of proposition-based pluralistic ignorance to
arise.
Subjects do not incorporate social proof by self-biased majority voting, but rather
the opposite.
33 How well these individual studies conform to the pluralistic ignorance explanation of the by-
stander effect has not been checked.
34 I.e., interchanging the , tie-breaking parameters. Alternatively, social beliefs could be de-
fined by weighing others perceived beliefs higher than ones own, or by moving to a thresh-
old rule requiring e.g. perceived agreement with all peers as done in (Seligman et al., 2013;
Christoff and Hansen, 2013)).
Several venues further for both formal and empirical research present themselves. As
no pure agent type group fits data very well, two possibilities are worth investigating.
First, how well will models with mixed groups perform? That not all subjects chose to
intervene in the single agent case seems to indicate that at least some behave as City
Dwellers; that some chose to intervene in the three agent case indicates that some act
as First Responders. With suitable proportions of each agent type, a model may be
produced which will match data more closely with an average of end conditions of runs
based on random picks from the mixed population. To fit both population mix and the
social influence parameter, a data set from a large-scale smoky room-style study would
be required.
Finding implementable resolution strategies for the pluralistic ignorance state could
be of benefit, if these turn out to work in practice. Some such have been suggested
in the literature; in the study of Schroeder and Prentice (1998), information on the
subject diminished the alcohol consumption among college students. How information
should provoke changes in agent type in the present framework is an open question. A
shorter term strategy for obtaining help is suggested by Cialdini (2007): Single out an
individual and ask only her for help. If an agent is singled out, the inaction of others
will no longer be eligible as a source of information about the event. Hence the agent
is forced to act on her private beliefs, in which case both Influenced Hesitator models
predicts intervention. Of formal studies, Proietti and Olsson (2013) show how a state
of pluralistic ignorance state may be dissolved by a series of announcements of private
beliefs heard by matrix neighbors. Specifying an agent type replicating the behavior
and varying only the network structure of the model might provide further insights into
the fragility of the phenomenon.
For a complete model of the bystander effect, both the first (noticing the event)
and third step (assuming responsibility) of the explanation provided in the introduction
must be modeled. The former may rest less on information processing than features
of physical space: as more people are present, less may notice the event e.g. due to
obscured line of sight. Modeling the third step may require a more expressive logical
framework in which beliefs regarding agent types may be held: if all agents falsely
believe a First Responder is present, all may believe that intervention is required while
no-one will take action.

Acknowledgements. The author would like to thank the editors for organizing the
CPH-LU workshops on social epistemology, as well as the participants of said work-
shops for valuable comments and discussion. The work has benefited especially from
discussions with Henrik Boensvang and Vincent F. Hendricks. Carlo Proietti is thanked
for his corrections to the manuscript. Finally, a warm Thank You to the two anonymous
reviewers: the comments, criticisms and correction provided by your thorough reading
of the original submission have been invaluable.
 Bibliography

Baltag, A. and Smets, S. (2008). A Qualitative Theory of Dynamic Interactive Belief

Revision, in G. Bonanno, W. van der Hoek and M. Wooldridge (eds), Logic and the
Foundations of Game and Decision Theory (LOFT 7), number Loft 7 in Texts in Logic
and Games, Vol. 3, Amsterdam University Press, pp. 958.
Baltag, A. and Smets, S. (2009). Talking your way into agreement: Belief merge
by persuasive communication, in M. Baldoni, C. Baroglio, J. Bentahar, G. Boella,
M. Cossentino, M. Dastani, B. Dunin-Keplicz, G. Fortino, M.-P. Gleizes, J. a. Leite,
V. Mascardi, J. Padget, J. Pavn, A. Polleres, A. E. F. Seghrouchni, P. Torroni and
R. Verbrugge (eds), Proceedings of the Second Multi- Agent Logics, Languages, and
Organisations Federated Workshops, CEUR Workshop Proceedings, Vol. 494, Turin,
Italy, pp. 129141.
Bicchieri, C. and Fukui, Y. (1999). The great illusion: ignorance, informational cas-
cades, and the persistence of unpopular norms, Business Ethics Quarterly 9(1): 127
155.
Blackburn, P., de Rijke, M. and Venema, Y. (2001). Modal Logic, Cambridge University
Press.
Bolander, T. and Birkegaard, M. (2011). Epistemic planning for single- and multi-agent
systems, Journal of Applied Non-Classical Logics 21(1): 934.
Christoff, Z. and Hansen, J. U. (2013). A Two-Tiered Formalization of Social Influence,
in D. Grossi and O. Roy (eds), Logic, Rationality, and Interaction, Lecture Notes in
Computer Science, Springer, pp. 6881.
Cialdini, R. (2007). Influence: The Psychology of Persuasion, HarperCollins Publishers.
Dgremont, C. (2010). The Temporal Mind: Observations on the logic of belief change
in interactive systems, PhD thesis, University of Amsterdam.
Demey, L. (2011). Some remarks on the model theory of epistemic plausibility models,
Journal of Applied Non-Classical Logics 21(3-4): 375395.
Fagin, R., Halpern, J. Y., Moses, Y. and Vardi, M. Y. (1995). Reasoning About Knowl-
edge, The MIT Press.
Halbesleben, J. R. and Buckley, M. R. (2004). Pluralistic ignorance: historical develop-
ment and organizational applications, Management Decision 42(1): 126138.
Hammack, R., Imrich, W. and Klavzar, S. (2011). Handbook of Product Graphs, Dis-
crete Mathematics and its Applications, 2 edn, CRC Press.
Hansen, J. U. (2011). A logic toolbox for modeling knowledge and information in multi-
agent systems and social epistemology by, PhD thesis, Roskilde University.
Hansen, P. G., Hendricks, V. F. and Rendsvig, R. K. (2013). Infostorms, Metaphiloso-
phy 44(3): 301326.
Hendricks, V. F. and Rasmussen, J. L. (2012). Nedtur! Finanskrisen forstet filosofisk,
Gyldendal Business, Copenhagen.
Latan, B. and Darley, J. M. (1968). Group inhibition of bystander intervention in
emergencies., Journal of Personality and Social Psychology 10(3): 21521.
Latan, B. and Darley, J. M. (1970). The Unresponsive Bystander: Why Doesnt He
Help?, Appleten-Century-Crofts, New York.
Latan, B. and Nida, S. (1981). Ten Years of Research on Group Size and Helping,
Psychological Bulletin 89(2): 308324.
Latan, B. and Rodin, J. (1969). A Lady in Distress: Inhibiting Effects of Friends
and Strangers on Bystander Intervention, Journal of Experimental Social Psychology
5: 189202.
Miller, D. and McFarland, C. (1987). Pluralistic ignorance: when similarity is inter-
preted as dissimilarity, Journal of Personality and Social Psychology 53: 298305.
Myers, D. G. (2012). Social Psychology, 11 edn, McGraw-Hill.
Prentice, D. and Miller, D. (1993). Pluralistic ignorance and alcohol use on campus:
some consequences of misperceiving the social norm, Journal of Personality and
Social Psychology 64: 24356.
Proietti, C. and Olsson, E. J. (2013). A DDL Approach to Pluralistic Ignorance and
Collective Belief, Journal of Philosophical Logic March: 117.
Rendsvig, R. K. (2013a). Aggregated Beliefs and Informational Cascades, in D. Grossi
and O. Roy (eds), Logic, Rationality, and Interaction, Lecture Notes in Computer
Science, Springer, pp. 337341.
Rendsvig, R. K. (2013b). Look and Learn: Extracting Information from Actions of
Others, [Link] .
Rendsvig, R. K. and Hendricks, V. F. (2013). Social Proof in Extensive Games, under
review .
Schroeder, C. M. and Prentice, D. A. (1998). Exposing Pluralistic Ignorance to Re-
duce Alcohol Use Among College Students, Journal of Applied Social Psychology
28(23): 21502180.
Seligman, J., Girard, P. and Liu, F. (2013). Logical dynamics of belief change in the
community, Under submission .
van Benthem, J. (2007). Dynamic logic for belief revision, Journal of Applied Non-
Classical Logics 17(2): 129155.
van Benthem, J., Gerbrandy, J. and Pacuit, E. (2007). Merging Frameworks for Inter-
action : DEL and ETL, in D. Samet (ed.), Proceedings of TARK 2007, pp. 7281.
van Ditmarsch, H. and Kooi, B. (2008). Semantic Results for Ontic and Epistemic
Change, in G. Bonanno, W. van der Hoek and M. Wooldridge (eds), Logic and the
Foundations of Game and Decision Theory (LOFT 7), number Loft 7 in Texts in Logic
and Games, Vol. 3, Amsterdam University Press, pp. 87117.
Westphal, J. D. and Bednar, M. K. (2005). Pluralisitc Ignorance in Corporate Boards
and Firms Strategic Persistence in Response to Low Firm Performance, Administra-
tive Science Quarterly 50: 262298.