Collective Action And The Evolution Of Social Norms Essay


People often ignore material costs they incur when following existing social norms. Some individuals and groups are often willing to pay extremely high costs to enact, defend, or promulgate specific values and norms that they consider important. Such behaviors, often decreasing biological fitness, represent an evolutionary puzzle. We study theoretically the evolutionary origins of human capacity to internalize and follow social norms. We focus on two general types of collective actions our ancestors were regularly involved in: cooperation to overcome nature’s challenges and conflicts with neighboring groups. We show that norm internalization evolves under a wide range of conditions, making cooperation “instinctive.” We make testable predictions about individual and group behavior.


Human behavior is strongly affected by culturally transmitted norms and values. Certain norms are internalized (i.e., acting according to a norm becomes an end in itself rather than merely a tool in achieving certain goals or avoiding social sanctions). Humans’ capacity to internalize norms likely evolved in our ancestors to simplify solving certain challenges—including social ones. Here we study theoretically the evolutionary origins of the capacity to internalize norms. In our models, individuals can choose to participate in collective actions as well as punish free riders. In making their decisions, individuals attempt to maximize a utility function in which normative values are initially irrelevant but play an increasingly important role if the ability to internalize norms emerges. Using agent-based simulations, we show that norm internalization evolves under a wide range of conditions so that cooperation becomes “instinctive.” Norm internalization evolves much more easily and has much larger effects on behavior if groups promote peer punishment of free riders. Promoting only participation in collective actions is not effective. Typically, intermediate levels of norm internalization are most frequent but there are also cases with relatively small frequencies of “oversocialized” individuals willing to make extreme sacrifices for their groups no matter material costs, as well as “undersocialized” individuals completely immune to social norms. Evolving the ability to internalize norms was likely a crucial step on the path to large-scale human cooperation.

Human social behavior is controlled by many interacting factors including material cost–benefit considerations, genetically informed social instincts, personality, and culturally transmitted norms, values, and institutions (1⇓⇓⇓–5). A social norm is a behavior that one is expected to follow and expects others to follow in a given social situation (6, 7). Humans learn norms from parents, through educational and religious practices, and from friends and acquaintances, books, and media. The adherence to norms is socially reinforced by the approval of, and rewards to, individuals who follow them and punishment of norm violators. Certain norms are internalized, that is, acting according to a norm becomes an end in itself rather than merely a tool in achieving certain goals or avoiding social sanctions (1, 2, 8⇓⇓–11). For individuals who have strongly internalized a norm, violating it is psychologically painful even if the direct material benefits for the violation are positive. Many individuals and groups are willing to pay extremely high costs to enact, defend, or promulgate norms that they consider important (12). At the same time, virtually all norms can be violated by individuals under some conditions (e.g., if the costs of compliance are too high). Norms thus can be viewed as one of the arguments in the utility function that each individual maximizes (9).

Internalizing a norm has two significant effects upon human behavior: People who have internalized a norm follow it even when doing so is personally costly, and they will tend to criticize or punish norm violators (13). Norm internalization allows individuals to reduce the costs associated with information gathering, processing, and decision making (11) and the costs of monitoring, punishments, or conditional rewards that would otherwise be necessary to ensure cooperation (9, 14). Internalization of norms allows individuals and groups to adjust their utility functions in situations with a rapidly changing environment when genetic mechanisms would be too slow to react (9). A society’s values are transmitted through the internalization of norms (15), with some societies being more successful than others due to their norms and institutions (16). The presence of both costs and benefits of norm internalization suggests that the human ability to internalize norms has been subject to natural, sexual, and social selection for as long as human culture has been in existence.

Norm internalization is an elaboration of imitation and imprinting found in various species of birds and mammals (17). Plausibly, then, a propensity to follow norms is at least partly an innate feature of our social psychology, whereas the substantive content of the norms of a given society are largely cultural (18). Our models below are designed to explore these features of norms. There is a rapidly growing body of theoretical work on gene–culture coevolution and its effects on human behavior (3, 19). Some approaches take the capacity for internalized norms as a given and study the cultural evolution of specific norms that are internalized (e.g., ref. 20). However, except for an attempt in ref. 9 that equated internalization of norms with blind copying of a behavior from others (21, 22) while largely ignoring their associated material payoff or normative value to individuals, the question of how humans evolved to internalize norms has apparently not received much attention from theoreticians. We aim to fill this important gap in our knowledge. The key question we ask is, “How could … norm-using types of players have emerged and survived in a world of rational egoists?” (ref. 23, p.143).

Models and Results

We consider two general kinds of collective action problems our ancestors might have evolved to solve. The first includes group activities such as defense from predators, cooperative hunting, cooperative breeding, and so on. The success of a particular group in solving these problems does not depend very much on the actions of neighboring groups. We refer to such collective actions as “us-vs.-nature” games. The second kind of collective actions, which we refer to as “us-vs.-them” games, include direct conflicts and/or other costly competition with other groups over territory, mating opportunities, access to trade routes, and so on. The success of one group in an us-vs.-them game means failure or reduced success for other groups albeit at a cost to the winner as well. In both of these types of models, group success is an important component of fitness. Much previous modeling of group competition has not modeled this distinction but has leaned on us-vs.-them games when interpreting empirical examples just because of the costly self-sacrifice often displayed in violent conflict (3, 19). Both us-vs.-nature and us-vs.-them models will generate cooperation in the right circumstances (19, 24, 25), but modeling them in a comparative framework is instructive because the fitness payoffs to solving these two kinds of cooperative dilemmas are very different. In particular, escalation of efforts due to an intergroup arms race is common in the latter but absent in the former (SI Appendix). We will consider separately and contrast these two games.

We consider a population of individuals living in a large number of groups of constant size . Generations are discrete and nonoverlapping. During their lifetime, group members have an opportunity to participate in a number of collective actions. Individual participation in collective actions is costly although any benefit is shared equally among all group members; this creates an incentive to free ride (26). An effective mechanism to reduce free riding is punishment (27⇓–29). Therefore, we assume that individuals can punish their free-riding groupmates at a cost to themselves. Identifying free riders requires the individual to pay additional costs of monitoring the group. The costs of monitoring and punishing others, and being punished by them, increase linearly with group size (which will vary between different simulations). Individual efforts in a collective action and in punishing free riders will be described by variables and , respectively, each equal to or . As a result of participating in collective actions and punishment, individuals accumulate material payoff . At the end of each generation, groups survive and duplicate with probabilities dependent on their success in collective actions; in surviving groups, individuals reproduce with probabilities proportional to their accumulated material payoffs. Some offspring disperse randomly to different groups.

We extend the standard approach outlined above for the case of norm internalization. We assume that the society has a prosocial (injunctive) norm in the sense that individuals learn (e.g., from parents, elders, or peers) that they are expected to contribute to collective actions and punish free riders (i.e., choose ). However, individuals’ decisions are controlled by both the ability to internalize the norm and material payoff considerations. We treat as a continuous trait controlled genetically (). We postulate that any individual updating its behavior attempts to maximize the utility function[1]The two terms in Eq. 1 capture the effects of nature and culture, respectively. Individuals with are “undersocialized” (i.e., they do not care about the norm and only want to maximize their material payoff ) (1, 2). If , following the norm is a part of the individual’s preference. Individuals with are “oversocialized” [i.e., they do not care about the material payoff and always follow the norm (1, 2) by choosing ]. Nonnegative parameters and measure the maximum value of following the norms of contributing (i.e., choosing ) and punishing free riders (i.e., choosing ) for oversocialized individuals. These parameters increase with the strength of “social pressure” to follow the norm; we assume that they are exogenously specified. We wish to understand the evolution of , starting with very low values, and its effects on individual and group behavior.

To study our models we used agent-based simulations. Figs. 1 and 2 illustrate observed evolutionary dynamics. In both cases shown, norm internalization trait evolves after some time; its evolution results in increasing within-group cooperation and punishment. In the us-vs.-nature game (Fig. 1), there is an increase in material payoffs and fitness. In contrast, in the us-vs.-them game (Fig. 2), material payoffs substantially decline as group members put increasingly more effort in between-group competition. In the example of the us-vs.-nature game shown, the population becomes dimorphic in with approximately two-thirds of the population having large values of and the rest with very small values of . These dynamics are analogous to those coming under the rubric of “evolutionary branching” identified by adaptive dynamics methods (30, 31).

Fig. 1.

Examples of evolutionary dynamics. Us-vs.-nature game with . (A) Frequencies of individuals using different combinations of strategies . (Inset) The average fitness. (B) The dynamics of the distribution of the internalization trait . The intensity of the black color is proportional to the number of individuals with the corresponding trait values present at a given time. The red line shows the mean value of . See Methods and SI Appendix for exact definitions of parameters.

Fig. 2.

Examples of evolutionary dynamics: us-vs.-them game with . (A) Frequencies of individuals using different combinations of strategies . (Inset) The average fitness. (B) The dynamics of the distribution of the internalization trait . The intensity of the black color is proportional to the number of individuals with the corresponding trait values present at a given time. The red line shows the mean value of . See Methods and the SI Appendix for exact definitions of parameters.

Overall, simulations show that if the norm internalization trait remains small, individuals make no effort (). If norm internalization trait instead evolves to a large value, individuals both contribute to collective action and monitor and punish occasional free riders (). With intermediate values of , the population represents a mix of free riders () and contributors who either punish () or do not (). The frequency of “selfish punishers” () typically remains very low. If norm internalization evolves, it often emerges quite rapidly after some waiting time, suggesting a transition between alternative quasi-stable states. Occasionally, rapid reverse transitions to low-internalization states are also observed.

The effects of parameters on evolutionary dynamics are illustrated in Figs. 3 and 4 (see also SI Appendix). Consider first what happens if no norm internalization is allowed to evolve [or production and peer punishment are not promoted (i.e., ; left part of the first column in each set of graphs). For parameters used here, in us-vs.-nature games individuals make no effort (Fig. 3). In us-vs.-them games (Fig. 4), some individuals volunteer for production; the average group effort is small and decreases with group size. In either type of games there is no punishment.

Fig. 3.

Summary graphs for us-vs.-nature games: efforts , punishment , internalization , fitness , and SD in internalization trait for different normative values of production and punishment , and group size . Other parameters: . Shown are averages based on 10 runs for each parameter combination.

Fig. 4.

Summary graphs for us-vs.-them games: efforts , punishment , internalization , fitness , and SD in internalization trait for different normative values of production and punishment , and group size . Other parameters: . Shown are averages based on 10 runs for each parameter combination.

If groups encourage their members to contribute to production but peer punishment is not promoted (i.e., ; left columns of graphs in Figs. 3 and 4), there is no punishment but some norm internalization evolves. In us-vs.-nature games, its effects on production and material payoffs are insignificant. In us-vs.-them games, there is a small increase in production accompanied by decreasing payoffs.

If group members are encouraged to punish free riders but production itself is not promoted (i.e., ; left part of the second and third columns of graphs in Figs. 3 and 4), then strong norm internalization can evolve simultaneously with a dramatic increase in production and punishment. In us-vs.-them games, increasing group size has a strong negative effect on internalization, production, and punishment. The effects of costs of punishing and of being punished on are relatively small (SI Appendix). There is a decline in fitness in us-vs.-them games but an increase in us-vs.-nature games.

If both production and punishment of free riders are encouraged (i.e., ), all efforts generally increase in both games. There are nonadditive interactions of parameters but the general patterns are hard to see. Increasing the costs of being punished increases production efforts. In us-vs.-them games, smaller groups typically have higher and production and punishment efforts than larger groups; larger groups can have higher fitness if they evolve no internalization and make fewer or no efforts in production and punishment. In us-vs.-nature games, the situation is more complex: If the cost of being punished is small, larger groups can evolve higher (because the total effect of punishment is higher). If this cost is moderate or large, smaller groups evolve higher . Larger groups can also have relatively high production effort. Interestingly, increasing the difficulty of us-vs.-nature games results in more norm internalization (SI Appendix, Table S1 and Figs. S14–S16


Collective online aggression directed towards actors of public interest is an increasing phenomenon. While various types of social media have been involved in such online firestorms (e.g. content communities such as YouTube), blogs and social networking sites such as Facebook are outstanding triggers [1]. In 2011, Christian Wulff, the former federal president of Germany, was accused of corruption–claims that afterwards were rejected as unfounded although they promptly led to his resignation. The Wulff-affair was massively amplified by the negative word-of-mouth dynamics in social media. In 2013, the company Amazon was accused of the ill treatment of temporary workers. The Amazon-affair led to floods of negative comments on Amazon’s Facebook profile. Firestorms also shake academia: In 2011, the former minister of defense of Germany, Karl-Theodor zu Guttenberg, was accused of plagiarism. These accusations triggered widespread online debates and ultimately led to the denial of his PhD and to his resignation.

The examples illustrate how online aggression has emerged from the private niche of limited email bullying to a publicly visible and relevant phenomenon. Dependent on the focus of the underlying research, the phenomenon of aggressive, offensive and emotional commenting in social media has been labeled flaming, cyberbullying, online harassment, cyber aggression, electronic aggression, toxic online disinhibition, trolling or, if the aggression resembles crowd-based outrage, online firestorms [1–5]. In online firestorms, large amounts of critique, insulting comments, and swearwords against a person, organization, or group may be formed by, and propagated via, thousands or millions of people within hours [1]. Social media enable these unleashed phenomena [2, 3, 6]. They allow attacking everywhere at anytime with the potential for an unlimited audience. They raise the likelihood for hostile misinterpretations due to limited discursive action and social media’s absence of nonverbal cues. They reduce the risk for feedback reactions because users can “sneak off” after the aggressive act.

The phenomenon of online aggression is not well understood despite the great deal of attention on hostile behavior in social media in both the mainstream media and the empirical literature [2, 7–16]. Most contributions are descriptive and are conducted largely in the absence of theories [2, 15]. If contributions refer to theories they are mainly guided by traditional bullying research theory, more precisely by the massive amout of existing research concerned with cyberbullying among adolescents. Within this view, online aggression is understood as an irrational and illegitimate behavior that is caused by underlying personality characteristics, such as a lack of empathy and social skills, narcissism, impulsivity, sensation seeking, emotional regulation problems or psychological symptoms such as loneliness, depression, and anxiety [15, 17]. Traditional bullying research theory, however, misses the point that in online firestorms, aggression happens in public, and not in private, social networks.

It therefore seems questionable whether bullying research theory is transferable to online firestorms. For example, a strong and commonly shared assumption within bullying research theory is that anonymity, understood as the degree to which a communicator perceives the message source as unknown and unspecified, promotes aggression through decreased inhibitions [3, 18–21]. For online firestorms it suggests that negative, and particularly aggressive, word-of-mouth propagation in social media will weaken if real-name policies are introduced. In this article we show that this assumption is not necessarily true because the reverse effect can be obtained: Individuals have a strong motivation for being non-anonymous when being aggressive in social media. We explain this behavior pattern by social norm theory. Social norm theory may be a more appropriate theory to understand communication behavior in social media and to draw conclusions, for example, that real-name policies will not weaken online firestorms.

The remainder of this paper is structured as follows: the next section introduces social norm theory to understand aggressive behavior in a social-political online setting, and develops hypotheses. The subsequent sections explain the dataset, the measurements and the method, and present the empirical findings. We conclude with a discussion of the findings, research limitations and suggestions for further research.

A social norm theory on online firestorms

Social norms are fundamental to human behavior [22, 23]. Former literature defines norms as statements “that something ought or ought not to be the case” ([24] page 132), as institutionalized role expectations [25], or as becoming apparent if behavior attracts punishments [26]. In general, norms are mental representations of appropriate behavior in society and smaller groups and, consequently, guide the behavior of individuals. Norms that are characterized as social “must be shared by other people and partly sustained by their approval and disapproval” ([23] page 99). Social norms are created intentionally because they promote the provision of a public good that benefits a collective, for example less pollution in a neighborhood due to less burning of leaves [27], less harm to health through cessation of smoking [28], or more fairness through income differentials [29, 30]. The public good view does not automatically imply that social norms are always beneficial for all persons concerned. In fact, many social norms exclude certain groups from public goods because they promote the interest of one subgroup, i.e., they serve “functions of inclusion and exclusion” ([23] page 108). For example, peer-group norms aim to strengthen cohesion within the group by offering group privileges [23, 31].

To be sustainable, social norms need to be enforced, otherwise Olson’s [32] zero contribution holds: “if all rational and self-interested individuals in a large group would gain as a group if they acted to achieve their common interest or objective, they will still not voluntarily act to achieve that common or group interest” ([32] page 2). Social norms are enforced by simple sanctions which trigger feelings of guilt and shame in the case of internalized social norms. Consequently, the mere expectation of sanctions, in turn, supports the enforcement [23]. Enforcement also happens through actual bilateral and multilateral costly sanctions where those who cause negative externalities are confronted with punishments and normative demands [28, 33]. Linked to Olson’s [32] zero contribution, norm enforcement itself is a second-order public good: self-interested and utility-maximizing individuals do not naturally contribute to norm enforcement and may prefer free riding [28, 33]. Ostrom [34] however stresses how, in practice, contextual variables and the engagement of certain types of individuals determine whether collective action and cooperation is enhanced or discouraged. Similarly, Ellickson [35] emphasizes how norms may emerge or shift dependent on cost-benefit conditions or group composition. Also the presence, salience, or strength of social ties can explain individual variation in social-political engagement [36, 37]. For example, diffuse networks of weak bridging ties encourage mobilization through communicative advantage [38]. Specifically, research shows that Olson’s [32] second-order public good dilemma can be overcome if (1) norm enforcement is cheap, i.e., it occurs in low cost situations [27, 39], (2) additional benefits are provided to the norm enforcers that disproportionately motivate them compared to non-enforcers, i.e., selective incentives are present [28, 32] and/or (3) if some individuals are present that are intrinsically motivated to enforce norms, i.e., some amount of altruistic punishment occurs [40–42]. In the following we elaborate these three conditions for social media to explain the phenomenon of online firestorms.

Online firestorms within a social norm theory

Aggressive word-of-mouth propagation in social media is the response to (perceived) violating behaviors of public actors. Public actors include, for example, politicians who disregard political correctness norms, corporations that violate human rights, or academics who violate scientific norms by engaging in plagiarism. In this view, online firestorms enforce social norms by expressing public disapproval with the aim of securing public goods, for example, honesty of politicians, companies or academics. The stunning waves of aggression typical for online firestorms can be explained by the characteristic features of social media that ideally contribute to the solution of the second-order public good dilemma of norm enforcement. Digital norm enforcement in social media is cheap, and selective incentives and intrinsically motivated individuals are present.

In social media, sanctioning norm violations occurs in low-cost situations. The basic idea of the low-cost hypothesis is that attitudes or preferences are more likely to guide individual behaviors when norm enforcement behavior is relatively cheap [27, 39, 43]. Evidence in various research fields supports this basic tenet (for an overview see [43]). For example, the voting paradox [32], i.e., the fact that citizens participate in elections even though they are aware of the marginal influence of their vote, is often explained by referring to the low-cost hypothesis [44]. In social media, a number of factors contribute to such low-cost situations. First, social media mobilize former free riders because online criticism is monetarily inexpensive, hardly time-consuming and can be performed anywhere and anytime, compared, for example, to elaborate street protests [1, 2]. One example is the limited message length in the social media platform Twitter, which obliges communication to be short and quick. It is less astonishing that Twitter has been involved in most of the recent cases of online firestorms [1]. Second, in social media, people who are geographically completely removed from each other can assault each other verbally without fear of bodily harm. Nonverbal cues such as facial expression and physical size are lacking, thus reducing the empathy of the aggressor and the impact of authority of the victims typically expressed by dress, body langugage, and social setting [2, 3, 45]. Third, social media give ordinary people the power to communicate (perceived) norm violations to a very large audience [46, 47]. The internet re-creates village-like interconnectedness within a global, pluralistic society by crossing local, or even national, boundaries due to unrestrained information flow [48]. To compare, while aggressive norm enforcement is a rare behavior in the non-digital context (Brauer and Chekroun [49] found that max. 4% of bystanders aggressively sanction daily deviant behavior by insulting or aggressive shouting), we should observe it more frequently in the digital social media context for the reasons given above.

Hypothesis 1.

Provided that a social-political issue finds its way into social media platforms, online aggression takes place more frequently than in the non-digital context because sanctioning of (perceived) norm violations occurs in low-cost situations.

In social media, selective incentives that benefit a latent group of norm enforcers are disproportionally present [28, 32]. Individuals only bear the costs of norm enforcement if the potential benefits of their actions exceed the costs [50]. Selective incentives translate resentment for norm breaching into action in situations where it is unclear whether a necessary critical mass of other norm enforcers will join the action. In such situations, cost sharing cannot be expected, nor can clear benefits from norm enforcement, such as an actual behavioral change by the accused person or organization, be predicted. In the case of selective incentives, individuals participate in collective action in response to salient private benefits [51]. Whether individuals are able to reap selective incentives is dependent on the issue at stake and on certain individual or group characteristics. Social media contribute to the presence of selective incentives by enhancing the salience of private benefits. In social media, for example, highly controversial topics are debated. Social media are, in addition, highly influenced by the multiplication of cross-media dynamics, for example by public scandals taken up or created by news media leading to comments in social media. Broad public discussions and connections to public scandals give credible signals that a norm infringement at the expense of a latent interest group–be it the group an individual belongs to or identifies with–has occurred [52].

Hypothesis 2.

Online aggression in social media is encouraged by salient selective incentives, for example, in highly controversial topics or in topics connected with public scandals.

Social media ensure that a high amount of intrinsically motivated actors are present. Individuals engage in costly norm enforcement if they have an intrinsic desire to “make the world a better place” [53–55]. This type of norm enforcement has been intensively discussed as “altruistic punishment”, i.e., individuals punish, although the punishment is costly for them and yields no material gain [42]. Altruistic punishment is driven by strong negative emotions towards the norm defector [40, 41, 56] and by people’s perception of a state of affairs as illegitimate [57–61]. Strong intrinsic motivation, however, is only likely to encourage participation if it is reinforced by organizational or individual ties [37]. This requirement is given in the infrastructural setting surrounding online firestorms. The technical mechanisms of social media such as newsletters, newsgroups, followers, or social media sharing ensure that intrinsically motivated individuals are optimally informed about cases that, in their view, represent offenses against existing social norms. Beyond this, they provide opportunities to tackle these norm violations by commenting on them.

Hypothesis 3.

Intrinsically motivated actors encourage online aggression in social media.

The non-anonymity of negative word-of-mouth dynamics in social media

In social media, people can hide or alter their identity. They may either comment by providing no name or at least not their real name, i.e., a (random or stable) pseudonym. Existing literature on online behavior hypothesizes that such online anonymity is one of the principle factors that decreases inhibitions, increases self-disclosures and therefore promotes online aggression [3, 18–21]. This causal mechanism is also assumed by social media consultants who attempt to explain online firestorms [62].

In general, anonymity produces the “stranger on a train” phenomenon, wherein people share intimate self-disclosures with strangers as they do not expect a reunion and hence do not fear any risks and constraints [63]. To that effect, “when people have the opportunity to separate their actions online from their in-person lifestyle and identity, they feel less vulnerable about self-disclosing and acting out” ([3] page 322). With regard to heightened aggression and inappropriate behavior, psychosocial motives exist for being anonymous [19]. Anonymity first detaches from normative and social behavioral constraints [64]. Second, it allows to bypass moral responsibility for deviant actions [3]. Third, it reduces the probability of social punishments through law and other authorities [20]. Fourth, it triggers an imbalance of power which limits the ability of the victim to apply ordinary techniques for punishing aggressive behavior [65]. Fifth, it gives people the courage to ignore social desirability issues [3] and finally, it encourages the presentation of minority viewpoints or viewpoints subjectively perceived as such [66–70].

Former research has concluded that the possibility for anonymity in the internet fosters aggressive comments. It is assumed that online aggression is driven by lower-order moral ideals and principles and, consequently, people feel ashamed to aggress under their real names. However, the empirical evidence for such a link is scarce and no definitive cause-effect relationship has evolved. Studies suggest that anonymity only increases online aggression in competitive situations [71], that anonymity does not increase online aggression but does increase critical comments [72], or that the effect of forced non-anonymity on the amount of online aggression is a function of certain characteristics of user groups, e.g. their general frequency of commenting behavior [73].

The former conceptualization of online aggression is rather narrow, in particular for aggression in social media. According to social norm theory, in social media, individuals mostly use aggressive word-of-mouth propagation to criticize the behavior of public actors. As people enforce social norms and promote public goods, it is most likely that they perceive the behavior of the accused public actors as driven by lower-order moral ideals and principles while that they perceive their own behavior as driven by higher-order moral ideals and principles. From this point of view there is no need to hide their identity.

Furthermore, aggressive word-of-mouth propagation in a social-political online setting is much more effective if criticism is brought forward non-anonymously. This is due to the fact that non-anonymity inceases the trustworthiness of the masses of weak social ties to which we are linked, but not necessarily familiar with, in our digital social networks. Trustworthiness of former firestorm commenters encourage us to contribute ourselves. First, non-anonymity is more effective as the credibility of sanctions increases if individuals use their real name [70, 74]. Anonymity makes “information more suspect because it [is] difficult to verify the source’s credibility” ([70] page 450). This removes accountability cues and lets one assume that individuals present socially undesirable arguments [74, 75]. Second, the views of non-anonymous individuals are given more weight: “Just as people are unattached to their own statements when they communicate anonymously, they are analogously unaffected by the anonymous statements of others” ([69] page 197). Anonymous comments have less impact on the formation of personal opinions [69, 76], on the formation of group opinions [74], and on final decision making [77]. Third, anonymity lowers the identification with, support of, and recognition by, kindred spirit [78]. In anonymous settings, individuals cannot determine who made a particular argument, how many different people expressed similar arguments, whether a series of arguments are all coming from the same person, or the degree to which other commenting individuals are similar to oneself [74, 79–81]. Anonymity filters out cues that communicate social identity, cues that are necessary to characterize comments by others [74, 82], to identify with individuals in social comparison processes [74] and to coordinate group interactions [80]. Finally, anonymity reduces the benefit to be positively evaluated by others [83, 84]. Studies show that exclusively anonymous conditions induce little mobilization because anonymity excludes the benefit of recognition by others [85].

From a social norm point of view, the arguments suggest that aggressive word-of-mouth propagation in a social-political online setting takes place non-anonymously. People have a strong feeling to stand up for higher-order moral ideals and principles. Commenting anonymously is a costly, wasteful behavior, as sanctions are less credible, create less awareness, less support and offer few benefits. These considerations make particular sense in the usual setting of firestorms, namely social media where usually, weak social ties are clustered around ideologically like-minded networks. Such networks likely support non-anonymous aggressive sanctions that confirm their worldview.

Hypothesis 4.

In a social-political online setting, non-anonymous individuals, compared to anonymous individuals, show more online aggression.

As stated earlier, norm enforcement is fostered if selective incentives and intrinsically motivated actors are present. Consequently if social norm theory is an appropriate theory for online aggression in a social-political online setting, these groups in particular should give more weight to the benefits of non-anonymous aggressive word-of-mouth propagation. Simultaneously, they give less weight to potential risky consequences such as being subject to deletion, banned from websites, formally convicted by the accused actor for defamation of character and/or damage to reputation, or informally sanctioned by social disapproval from online or offline individuals [86].

Hypothesis 5.

In a social-political online setting, in situations that offer selective incentives, compared to situations without selective incentives, more online aggression by non-anonymous individuals is observed.

Hypothesis 6.

In a social-political online setting, intrinsically motivated aggressors (i.e. aggressive commenters), compared to aggressors without intrinsic motivation, show more online non-anonymous aggression.

Materials and Methods


We test the hypotheses with a census of a major social media platform concerned with public affairs. We analyze all comments on online petitions published at the German social media platform between May 2010, the launching of the online portal, and July 2013. Online petitions exemplarily include protests against pay-scale reform of the German society for musical performing and mechanical reproduction rights called GEMA (305,118 signers), against the enforcement to finance public service media (136,010 signers), against the closing of the medical faculty at the University Halle (58,577), or for the resignation of an Austrian politician (9,196 signers) or the Bavarian minister of justice (6,810 signers). Online petition platforms seem very suitable to investigate the phenomenon of negative word-of-mouth in a social-political online media setting. First, online petitions are concerned with public actors and public affairs, for example, internet security, misbehavior of firms, politicians, or academics, public spending, tax issues, animal protection, etc., and thus provide a central location where public norms are negotiated. Second, online petition platforms are prototypical social media platforms: everybody is allowed to participate and create content for any kind of topic, and the debates and comments are publicly visible. Third, qualitative evidence suggests that many popular firestorms have been triggered or have been surrounded by online petition platforms, for example the Deutsche Telekom firestorm in 2013, or the firestorm leading to the displacement of the German Federal President Christian Wulff in 2011. Fourth, online petition platforms are concerned with real-life cases. Many former studies are based on artificial laboratory experiments to study negative word-of-mouth behavior on the internet. Finally, online petition platforms cover a wide range of public issues and affairs, implying lower selection biases as compared to case studies about online firestorms (such as in [1]).

The final dataset includes 532,197 comments on 1,612 online petitions. There were a total of 3,858,131 signatures over the 1,612 petitions between 2010 and 2013, with detailed information about the wording of the comment, the commenters, the signers and the petition. The dataset was provided to the authors in an anonymous form by the platform owner. For each signer and commenter, however, the dataset indicated whether he/she had originally contributed anonymously (= 1) or non-anonymously (= 0). For this study, no approval of any ethics committee was sought because all data are publicly accessible on and no names of signers or commenters can be tracked and identified in the dataset. In order to prepare the dataset in accordance with our theory, we rely on a mixed-method big-data approach. For many variables we use a qualitative approach to arrive at meaningful quantitative measurements.

The present dataset allows us to exclude two biases which, in other studies, frequently affect findings on relations between anonymity and aggression. First, there was no active intervention in the ratio of anonymous and non-anonymous aggressive comments in the dataset. In the period of data collection, the platform owner did not moderate the comments on his own initiative. However, he reacted by deleting selected inappropriate comments when the user community reported them. According to the platform owner, a deletion was independent of whether the inappropriate comment was provided anonymously or not, as he explicitly considered this difference as irrelevant to liability issues. Second, we may also exclude any bias stemming from differing legal jurisdictions: Potential legal implications for identified aggressors are the same across the entire study. In Germany, the jurisdiction on defamation and insult is part of the federal law [87], i.e., as the entire study pertains to the same legal jurisdiction, all defamatory or aggressive commenters across all German states face the same potential costs for their actions.

Measurement of Variables

We measure online aggression in the following manner. In general, inconsistency in the operationalization of online aggression dominates research [88]. Operationalization includes impolite statements, swearing, flirting, exclamations, expressions of personal feelings, use of superlatives [89] to profanity, typographic energy (e.g. exclamation marks), name calling, swearing, and general negative effect [72, 88]. We rely on the definition of online aggression in firestorms, i.e., large amounts of critique, insulting comments, and swearwords against a person, organization, or group formed by, and propagated via, social media platforms [1]. Accordingly, we measure online aggression by direct offenses within the comments on online petitions (e.g. “I hate GEMA, complete morons and exploiters”, ID469090), swearwords (e.g. “Fuck that Shit!”, ID477368), and expressions of disgust or contempt (e.g. “The deportation policies of German authorities is commonly a disgusting, repulsive and inhuman mess!”, ID418089). Expressions of disgust and contempt are typical responses to morally offensive behavior [90]. Importantly, even from the outside perspective, we confidently evaluate these expressions to be intended as aggression. This is because we do not expect close relationships or shared, subcultural interactional norms between the commenter and the targeted actor in petitions, in contrast to profane language between friends representing covert closeness and not aggression [91].

To systematically collect online aggression, we compile a list of frequently used swearwords from synonym reference books and online databases of swearword collections (e.g. This approach corresponds to previous studies that count aggressive postings by using a pre-defined set of aggressive words (such as in [73]). Then, we disaggregate the 532,197 comments into single words and count them. Frequently occurring words are manually checked and classified as online aggression if applicable. Subsequently, we exclude all words that can be used for different meanings, for example, as swearwords or as terms for animals. These steps led to a final list of 1,481 words that express offenses, swearwords, and disgust. Using this final list of aggressive expressions, we count the amount of online aggression in each comment. Subsequently we qualitatively check the appropriateness of our approach by comparing subsamples of comments with our quantitative measurement. We take the logarithm added by 1 to create an approximate normal distribution of the variable.

Independent variables.

Anonymity is measured in the following way: Before online users sign a petition and subsequently formulate a voluntary comment, they are requested to provide their real names and addresses. In regard to public visibility, they are given the choice to allow their real name to be published or to remain anonymous, i.e., only the postal code is visible to other users (0 = non-anonymous, 1 = anonymous). Although the theoretical possibility of using pseudonyms does exist, we expect that commenters’ incentive for pseudonyms is low. This is because anonymity complies with the hidden name option and petition organizers may classify the signature of pseudonyms as invalid.

Controversy that accompanies a petition is measured by the level of debate. Each petition provides the opportunity to start a debate on the petition homepage, a tool used in most petitions by supporters and opponents. A debate is structured by denoted pro- and contra-arguments, i.e., by arguments that underpin or oppose the petition’s concerns. Only arguments that differ in their content from formerly mentioned arguments are additionally incorporated. Within the pro- and contra-sections, commenters are allowed to oppose arguments by adding sub-replies (pro-reply-/contra-reply-arguments). More controversial topics lead to a higher diversity of pro-, contra-, pro-reply- and contra-reply-arguments. Thus, to measure controversy, we construct a Herfindahl index by taking the percentage of arguments within each category, i.e., pro-/contra-/pro-reply-/contra-reply-arguments, squaring it, adding them together and subtracting the final result from 1. The index measures the controversy that surrounds the topics of petitions from no controversy (= 0) to a maximum of controversy (= 1).

To identify scandals, we measure whether the accusation against an actor forwarded by a petition, for example corruption of a politician, is covered and framed as scandal by traditional news media (1 = yes / 0 = no). We define keywords that describe the content and concerns of the petition. In the database LexisNexis we search for whether these keywords are associated with the term “scandal” in the German-speaking media within a time period of one year before the starting date of each petition.

To measure actors’ intrinsic motivation, we operationalize fairness perceptions of commenters. We compile a list of 579 expressions frequently used in ideological discourses that indicate fairness issues, for example, expressions such as “injustice” or “unfair”. In addition, we use synonym reference books and databases, manually check frequently occurring words within comments and exclude ambiguous words. For each commenter we count the amount of intrinsic motivation by taking the sum of fairness words in the comment. We take the logarithm, added by 1, to create an approximate normal distribution of the variable.

Control variables.

We control for factors that influence the amount of online aggression.

The length of comment is measured by the total number of words in a comment. Longer comments are more likely to entail more aggression.

The time period between opening a petition and submitting a comment is included because the time point of comment submission may influence commenters’ level of aggression. Aggression may either take place in the very beginning, because most signatures and comment activity in petitions are submitted within the first days [92], or alternatively, in advanced stages, in the case where a petition experiences a boost due to revived public debate. We measure how many minutes after petition opens that a comment has been submitted.

The number of protesters having signed is included because larger protests are likely to attract more online aggression. We measure how many individuals sign a particular petition and consequently match this data with the comments on a certain day. The median of protesters amounts to 76 signers per day with a maximum of 2,926 signers per day. We take the logarithm of the number of protesters to create an approximate normal distribution of the variable.

The status of the accused may also influence online aggression. Theoretically, public actors with a high social status may be either protected from sanctions as they have more resources to reply to punishments by even more painful punishments, or, to the contrary, they can attract sanctions because they are also more vulnerable than lower status actors [93]. In practice, high status celebrities or politicians may also refrain from suing laypersons as it is counterproductive to their reputation. To take these complex influences into account, we control for the status of the accused. As a proxy for social status of the accused public actors, we collect the number of Google hits for the accused’s name (1 = <1000; 2 = <10,000; 3 = <100,000; 4 = <500,000; 5 = <1,000,000; 6 = >1,000,000). Google hits tend to reflect social status. To decrease measurement errors, for example due to actors sharing the same name, we additionally check whether the accused is listed in the German online encyclopedia Wikipedia (0 = no entry, 1 = entry in article’s subtitle, 2 = entry as main article). Wikipedia exclusively lists actors with a minimum public status. We add both variables and take the logarithm of the mean value.

We measure also whether the accused is a natural person or a legal entity. Legal entities professionally monitor the internet for defamation and gather more resources to fight accusations than do natural persons. To avoid that commenters anticipate differing costs for their aggressive behavior dependent on whom the accused actor is, we control for this factor. Two independent coders manually check whether the target is a natural person such as a scientist or politician (= 1) or a legal entity such as a government or an organization (= 0). In 4% of the petitions, the target is a natural person and not a legal entity.

The anonymity of the social environment of commenters measures the anonymity of the environment in which commenters live. This may influence how much aggression is expressed [94]. Less anonymous villages with tight social control likely increase sanctioning costs. As a proxy for the anonymity of commenters’ social environment, we measure the size, i.e., the number of inhabitants, of the city or village in which commenters live. The postal codes of each signer are aggregated such that individuals living in the same city or village are merged. The dataset includes 23,977 cities and villages. We count the number of signers for each city or village, and by random checking, we find that the correlation of the number of signers within a postcode region, and the de facto size of this region, is 0.92, validating our proxy. We allocate the size of residence variable to all signers and commenters. Bigger values indicate that commenters originate from more anonymous environments.

The regional scope of a protest is measured because issues of broad public relevance may attract more aggression. We measure the regional diversity of a petition by constructing a Herfindahl index ranging from no regional diversity (= 0) to a maximum of regional diversity (= 1). Signers are assigned to different German federal states on the basis of residential postal codes. We take the percentage of signers within each federal state, square it, add them together, and subtract the final result from 1.

The success of a petition is measured because successful petitions potentially deal with more relevant topics, which may indirectly influence the amount of online aggression. A petition is considered successful if the petition initiator defines the petition goals to be achieved in full or at least in part (1 = yes; 0 = no).

The petition motive may influence the amount of online aggression. Using a petition’s title and leading text, two independent coders classify the petitions with regard to their underlying motives by using the classification by Reiss [95]. Five major concerns are identified, namely idealism/fairness (42%), income/costs (19%), security/social order (13%), autonomy/self-determination (14%), and quality of life/competences (52%). Multiple assignments of petitions are possible. Idealism/ fairness serves as the reference group in the regression models.

Similarly, the petition topic may influence anonymity considerations and the amount of aggression. Depending on the societal area, be it the economy, politics, or culture, accused actors may differ in their thresholds of wanting to sue aggressive online commenters. Commenters may anticipate these thresholds and the related differing costs of being aggressive. This in turn affects commenters’ actual behavior. Using a petition’s title and leading text, two independent coders classify the petitions with regard to their underlying topics using the functional systems of a society [96]. Six major topics are identified, namely society (41%), arts (20%), economics (13%), politics (8%), media (8%), and environment and animal protection (8%). Multiple assignments of petitions are avoided. Society, including topics such as sport or solidarity, is the most general category and serves as reference group in the regression models.

For the summary of the descriptive statistics and bivariate correlations of the former variables, see S1 Table.


We apply random-effects and fixed-effects models to predict online aggression in petitions (for access to data, syntax, and Permission for using data of, see the Data availability statement). In both models the comments are grouped on the petition level. The random-effects model keeps within- and between-petition variation in the analysis. We assume that petitions vary not only within, but also between, each other, for example because some petitions have many supporters while other petitions have only a few supporters, or because of differences in the underlying goals and motives. We analyze whether online aggression within and between petitions changes when other variables within and between the petitions change. The fixed-effects model keeps only within-petition variation in the analysis. We also analyze whether the aggression within petitions changes when other variables change, for example the anonymity of commenters, the amount of intrinsic motivation or the amount of selective incentives within the petitions. Many variables of our dataset are time-invariant, i.e., constant petition features that do not vary on the petition level. In the fixed-effects model these variables are omitted. Both models have advantages as well as disadvantages. The fixed-effects model excludes all random noise between the petitions and is therefore often preferred as the golden standard. However, differences between the petitions, for example the number of supporters, may also be important in explaining negative word-of-mouth behavior within petitions. This speaks in favor of the random-effects model. We therefore apply both models and compare the results. We additionally run alternative conceivable models for the data structure, for example, logistic regression, Poisson regression, or negative binomial regression for panel data, as our dependent variable is (if not transformed) a count variable, or can be transformed into a binary variable that indicates whether a person is an aggressor or not. The results are similar with the results that follow and will therefore not be presented here.


In accordance with Hypothesis 1, the data substantiate that online aggression in social media is a more frequent phenomenon than in the non-digital context. In the analyzed online petition platform we find 197,410 aggressions according to our definition. 20.62% of all comments entail a minimum of one aggressive expression (Fig 1). In 9% of all comments we find two, up to fifteen, aggressive expressions. On the petition level, only 11% of all petitions include no aggressions. 34% include a negligible amount of aggressions from 1, up to 10. 37% include 11 up to 100 aggressions. 16% include 101 up to 1,000 aggressions. 2% include 1,001, up to 25,360, aggressions. Even if the prevailing majority of commenters make no use of aggressive language in social media, the numbers demonstrate that online aggression occurs not only in a vanishing minority of comments or petitions (compared to the observed vanishing minority of max 4% of bystanders aggressively sanctioning in the non-digital context [49]). This supports the claim that in social media, aggressive sanctioning behavior is a relatively frequent phenomenon because it takes place in low-cost situations.

We now move to the presence of selective incentives and intrinsically motivated actors in social media. The descriptive findings show that 47% of all petitions are accompanied by a highly controversial debate, 6% of the petitions are associated with a scandal in news media, and 26% of the commenters are motivated by fairness concerns. Social media thus indeed seem to offer an environment in which the second-order public good dilemma of norm enforcement can be overcome. Whether these conditions indeed contribute to norm enforcement is tested in Tables 1 and 2.

The random-effects model in Table 1, Model 1, confirms that situations that offer selective incentives, i.e., a petition is accompanied by a highly controversial debate or is connected with a scandal in news media, significantly encourage online aggression in comments. This preliminarily supports Hypothesis 2 (for the size of the effects see Figs 2 and 3). The fixed-effect model in Table 2 entails no results for selective incentives because petition-invariant effects are dropped. Further, the random-effects as well as the fixed-effects models in Tables 1 and 2, Model 1, preliminarily support Hypothesis 3: online aggression is encouraged by intrinsically motivated actors as compared to individuals without fairness concerns (for the size of the effects see Figs 4 and 5).

Building on the view that social media today are a major channel for digital social norm enforcement, which until now is not rejected by the data, Hypothesis 4 assumes that online aggression takes place non-anonmously. Aggressive commentors have nothing to hide: they stand up for higher-order moral ideals and principles. The goal of norm enforcement can be reached most effectively if sanctions are forwarded non-anonymously because they are credible, create awareness, support, and offer benefits. The descriptive statistics show that only 29.2% of all commenters prefer to remain anonymous. Anonymity of commenters is thus a characteristic feature of social media; however, a vast majority still comments under their real names. The results in Tables 1 and 2, Model 1, show the impact of commenters’ anonymity to predict online aggression in comments. In line with Hypothesis 4, both the random-effects and fixed-effects models show that more online aggression is obtained by non-anonymous commenters and not by anonymous commenters.

Exemplarily, we present three of the most aggressive comments by non-anonymous commenters: “Silly, fake, inhuman and degrading, racist, defamatory and ugly theses like those of Sarrazin (author's note: a former German politician) have no place in this world, let alone in the SPD (author's note: Social democratic party). Sarrazin certainly has no business in the Social democratic party and should try his luck with the Nazis” (ID352216); “HC Strache (author's note: Austrian politician) has an evil, inhuman character, he lies and tries to persuade other people of wrong ideas.” (ID284846); “These authorities are mostly no people, but §§§- and regulatory machines! I detest authorities–with my 67 years’ life experience after all!” (ID418089).

Figs 6 and 7 illustrate the size of the effect as predicted in the random- and fixed-effects regressions. The average effect of anonymity on aggression becomes sharper in the fixed-effects model. The random-effects model additionally illustrates that many of the very aggressive commenters appear non-anonymously. Overall, the effect size is small. However, the data clearly show that social norm enforcement, and not as popularly assumed, the risks of detection, seems the major motivation for aggression in social media because persons often aggress under their real names.

If norm enforcement is indeed the major motivation for aggression in social media, the highest amount of non-anonymous negative word-of-mouth should be obtained in situations that offer selective incentives and for intrinsically motivated actors. Model 2, in Tables 1 and 2, tests this assumption by introducing interaction effects between the anonymity of commenters and the presence of selective incentives and their intrinsic motivation. The results give preliminary support for Hypotheses 5 and 6. The highest amount of non-anonymous aggression is observed if a petition is accompanied by a highly controversial debate, is connected with a scandal in news media, and if persons are motivated by fairness concerns. By introducing these interaction effects, the main effect of anonymity on online aggression becomes insignificant, and thus suggests that the underlying reasons for non-anonymous aggression can be indeed explained by social norm theory, namely by selective incentives and intrinsic motivation.

Figs 2 and 8 illustrate the effect for the level of controversy within a debate. In the case of highly controversial topics, individuals clearly prefer to aggress non-anonymously, indicating that selective incentives are present (we code debates as highly controversial if the Herfindahl index is higher than 0.3, and as less controversial if the Herfindahl index is 0.3 or smaller). Figs 3 and 9 illustrate the effect for the connection with a scandal in news media. Particularly for scandalized topics, the biggest gap arises between the aggression of non-anonymous and anonymous commenters, with the former showing more aggression. Again it supports that scandals offer selective incentives for norm enforcement. Finally, Figs 4 and 5 illustrate the effect for intrinsically motivated individuals. Intrinsically motivated individuals clearly prefer to aggress non-anonymously.

Fig 8. Online aggression dependent on controversy and anonymity (fixed-effects).

Predictions of

One thought on “Collective Action And The Evolution Of Social Norms Essay

Leave a Reply

Your email address will not be published. Required fields are marked *