A Brief Overview of the Cascade Literature

The literature describes two kinds of foundational cascades: informational cascades and reputational cascade (Ellickson Reference Ellickson, Hechter and Opp2001, 51). Informational cascades, a concept developed by Siushil Bikhchandani, David Hirshleifer, and Ivo Welch (Reference Bikhchandani, Hirshleifer and Welch1992), occur when, faced with informational uncertainty, individuals mimic the behavior of others. In an informational cascade, individuals are motivated solely by informational needs. Other motivating factors such as reputational concerns are not incorporated into this model. Reputational cascades are similar; however, in the case of reputational cascades, individuals emulate the decisions of others not because of information uncertainty but because they seek social approval and wish to avoid disapproval (Sunstein Reference Sunstein1996a, 2033; Reference Sunstein1996b, 903; Kuran Reference Kuran1998, 623). The result is that people engage in preference falsification (expressing preferences that differ from what they genuinely believe) in order outwardly to conform to the views of others.

Timur Kuran developed the early literature on reputational cascades.Footnote ⁶ In “Sparks and Prairie Fires,” “Now Out of Never,” and more fully in Private Truths, Public Lies, Kuran discusses how people's willingness to speak is driven both by reputational concerns and by their informational needs (Kuran Reference Kuran1989, Reference Kuran1991, 7; Reference Kuran1997). He argues that both factors jointly determine individual “thresholds” and whether and when a cascade occurs and its subsequent impact (Kuran Reference Kuran1998, 623). Building on Kuran's work, Sunstein introduces the concept of norm cascades (Sunstein Reference Sunstein1996b). Norm cascades are similarly driven by a combination of motives: reputational concerns and information acquisition. Norm cascades model sudden collective swings in societal norms. Synthesizing much of this earlier work, Kuran and Sunstein (Reference Kuran and Sunstein1999) introduce the concept of availability cascades, which combines informational and reputational mechanisms.

The model we propose—cyberspeech cascades—weaves together these three cascade types: informational cascades, availability cascades, and norm cascades. These three theories are essential to our argument, and so they are explained in detail below.

A Closer Look: Informational Cascades

Informational cascades occur when an agent's actual beliefs change in response to the public conduct of others. The public information obtained from the behavior of previous actors overrides an agent's private signals, irrespective of accuracy. As each agent changes her behavior, this is observed by other actors, which in turn causes a change in the beliefs of other actors and their subsequent behavior. The process then repeats itself through multiple sequences of observation and belief modification, resulting in an informational cascade. The phenomenon is rooted in the volatility and fragility of mass behavior. Experts “call upon ideas like informational cascades to describe how peripheral knowledge can pervade a society, changing the perceptions and realities of political power” (Joyce Reference Joyce2010, 211). Bikhchandani, Hirshleifer, and Welch describe the concept as follows: “An informational cascade occurs when it is optimal for an individual, having observed the actions of those ahead of him, to follow the behaviour of the preceding individual without regard to his own information” (1992, 994). The agent alters her behavior on the assumption that the publicly available information is reliable.Footnote ⁷

Informational cascades are in fact quite common. For instance, in attempting to exit an airport after disembarking from an airplane, a passenger might join a small cluster of other passengers ahead of her moving in a particular direction, assuming that these individuals must be going toward the airport exit. That assumption, however, could be entirely erroneous—the fellow passengers may, for example, be simply heading to the restroom. As other disembarking passengers likewise act on this mistaken assumption and the crowd grows larger, an erroneous informational cascade forms.

Consider the following example of an informational cascade. A consumer must decide between shopping at two different stores—Store X and Store Y (Banerjee Reference Banerjee1992, 798–99).Footnote ⁸ The consumer has very little private information with which to make her decision. However, the consumer perceives that Store X is empty while Store Y has several shoppers milling about inside. Working on the assumption that the presence of shoppers is a reliable indication that Store Y is the superior choice, the consumer chooses Store Y. In doing so, however, the consumer adds to the informational appeal of Store Y for the following consumer faced with the same choice. Given similar thought processes, subsequent consumers faced with the same choice will replicate the actions of the first consumer, who through her initial choice unwittingly amplified the informational appeal of Store Y. In this fashion, Store Y attracts an increasing number of shoppers. This pattern is self-reinforcing: the impression that Store Y is the superior choice grows with each new shopper. Even a shopper who arrives late in the informational cascade and would prefer shopping at Store X may be persuaded to choose Store Y as a result of the large crowd of shoppers in Store Y.

However, Store Y—and this is an important point—may not necessarily be superior to Store X. It may in fact be inferior. The initial shoppers in Store Y may have simply preferred the color of the awning in front of Store Y to that of Store X. Or they might have perceived that Store X looked slightly better organized than Store Y in the way its products were stacked on the shelves. In either case, their choice is not a reliable indication of the superiority of Store Y. Erroneous informational cascades based on misplaced assumptions may be triggered “where initial misrepresentative signals start a chain of incorrect decisions that is not broken by more representative signals received later” (Anderson and Holt Reference Anderson and Holt1997, 859; Sterelny et al. Reference Sterelny, Joyce, Calcott and Fraser2013, 30–31). The information is inaccurate yet an informational cascade nevertheless takes hold.

Under conditions of sequential observation and limited private information, an informational cascade is difficult to avoid: it will form either with the third or fourth agent. Because the model is sequential, after one consumer makes a choice, the next consumer follows such that each agent is able to see the choices of all the other agents who decided before her. Each consumer has a degree of private information concerning the decision (i.e., which store is better), but, operating in a state of informational uncertainty, assigns equal value to her private and public knowledge. The first consumer, utilizing strictly her private information (as this is the only information she has at her disposal at this juncture), decides to shop in Store Y (as her personal knowledge mandates). Assuming the next consumer's private information similarly suggests that Store Y is the superior option, an informational cascade will initiate immediately. This is because the third agent will choose Store Y even if her private information indicates that Store X is the better option. The third agent gives equal weight to her private and public information and her public information now exceeds her private information. If the second consumer's private information indicates that Store X is the better choice, the decisions of the first and second consumers will negate each other in terms of the available public information. As such, the third consumer will be placed in an informational void and, as a result, she will be directed by her private information. However, regardless of her choice, her decision will initiate an informational cascade because the fourth agent's public information will now exceed her private information (i.e., a store with two customers as opposed to another with only one). Thus, provided there is sequentiality of choice and sufficient limits on private information, an informational cascade will form either with the third or the fourth adopter.

A Closer Look: Availability Cascades

Availability cascades extend informational cascade theory by incorporating the idea of availability heuristics and adding a second component: reputational concerns (Kuran and Sunstein Reference Kuran and Sunstein1999). Kuran and Sunstein assume actors utilize heuristics to assess the current state of affairs in which they find themselves. This “involves estimating the probability of an event on the basis of how easily instances of it can be brought to mind” (1999, 706). Thus, they argue, actors imitate each other not only because they believe that others have more reliable information (as is the case with informational cascades), but also “because they take the simple availability of information as an indication of its reliability” (Lemieux Reference Lemieux2004, 20). This in turn stimulates further adoption of the belief, triggering “a chain reaction that gives the perception increasing plausibility through its rising availability in public discourse” (Kuran and Sunstein Reference Kuran and Sunstein1999, 683). This self-reinforcing process of collective belief is, critically, strengthened by reputational concerns, as individuals change “their public responses in the interest of maintaining social acceptance” (Kuran and Sunstein Reference Kuran and Sunstein1999, 683). Actors adjust their behavior, and in doing so, they reinforce the cascade. Sunstein and Kuran posit the idea of availability entrepreneurs, social agents who “attempt to trigger availability cascades likely to advance their own agendas” (Kuran and Sunstein Reference Kuran and Sunstein1999, 687).

An availability cascade can alter people's assessments of risk (the model has found extensive application in the field of regulatory research and finance theory). Availability “shapes, on the one hand, judgments about the magnitudes of various risks and, on the other, the acceptability of these risks” (Kuran and Sunstein Reference Kuran and Sunstein1999, 712). For instance, the availability cascade model informs us that because of its elevated profile in the public conversation, people are likely to overestimate the risk of being wounded in a terrorist attack compared with being injured in the bathroom (in the United States, approximately 235,000 people are injured in bathrooms every year; see Bakalar Reference Bakalar2011). This shift in risk perception caused by availability cascades, as we will see, also contributes to breakdowns in online censorial control.

A Closer Look: Norm Cascades

The third cascade type, norm cascades, explains abrupt transformations in societal norms (Sunstein Reference Sunstein1996b). Sunstein argues that, for reputational reasons, a certain number of actors conceal their privately held views but may grow more emboldened to express them as others do likewise. Sunstein argues that norms may be far more fragile than is generally believed. He distinguishes between norm bandwagons and norm cascades. According to Sunstein: “Norm bandwagons occur when small shifts lead to large ones, as people join the ‘bandwagon’; norm cascades occur when there are rapid shifts in norms” (Sunstein Reference Sunstein1996b, 909). A norm bandwagon will emerge when “the lowered cost of expressing new norms encourages an ever-increasing number of people to reject previously popular norms, to a ‘tipping point’ where it is adherence to the old norms that produces social disapproval” (Sunstein Reference Sunstein1996b, 912).Footnote ⁹ At that point, a norm cascade will erupt, bringing about a sudden collapse of the norm. Norm cascade theory is rather broad in scope, focusing on abrupt wholesale cultural or political reconfigurations, such as the collapse of communism in Eastern Europe and the dismantling of apartheid in South Africa (Sunstein Reference Sunstein1996b, 912). Robert Ellickson describes the norm cascade process as follows: “[I]n the paradigm case an upstart norm starts slowly, gains momentum, and culminates in a triumphant rush. Various authors refer to a tipping point being passed, to an equilibrium changing not gradually but in punctuated fashion, or to a cascade being triggered” (Ellickson Reference Ellickson, Hechter and Opp2001, 51).

Similar to availability entrepreneurs, Sunstein argues that norm cascades are jumpstarted by norm entrepreneurs who attempt to rally support for their views by exploiting “private dissatisfaction with existing norms in order to bring about large-scale social change” (Sunstein Reference Sunstein1996b, 929). Sunstein cites political actors such as Martin Luther King, Jr., Ronald Reagan, and Jerry Falwell as examples of norm entrepreneurs (Sunstein Reference Sunstein1996b, 929). Certainly, figures such as Adolf Hitler and Mahatma Gandhi would also fall into this camp. For our model, norm cascades are highly relevant because they may play a key role in producing the sometimes sudden and punctuated character of a cyberspeech cascade.

A cyberspeech cascade, which occurs entirely in the context of online behavior, comprises informational cascades, availability cascades, and norm cascades. The two sections that follow explain how these three cascade types combine to produce cyberspeech cascades.

Informational Cascades

In the previous section, we described an informational cascade using the example of choosing between two stores. Incorrect beliefs about the acceptable boundaries of open online expression can, we argue, cascade in a similar fashion. To apply the model of informational cascades to online speech, we need to substitute certain variables (Bikhchandani, Hirshleifer, and Welch Reference Bikhchandani, Hirshleifer and Welch1992).Footnote ¹⁰ Rather than a choice between two stores, the competing choices in our model are self-censorship versus non-self-censorship. For expository convenience, we define these two choices as “closed speech” (i.e., self-censored speech) and “open speech” (i.e., non-self-censored speech).Footnote ¹¹ Agents possess both private and public information. A speaker's private information consists of the understanding that certain speech is impermissible; her public information is gleaned from the online conduct of other speakers. For the purposes of our model, a distinction must also be made between “cyberspeech bandwagons” and “cyberspeech cascades.”Footnote ¹² The former involves a modest uptick in speakers communicating openly. The latter occurs when multiple bandwagons of open speech converge in a synergistic manner. This takes the form of a large-scale eruption of open speech that ripples across social media and other online speech platforms, causing vast numbers of Internet users to suddenly doubt their previously held perceptions of the permissibility of certain forms of speech.

The essential component for an informational cascade (uncertainty) is built into the strategy of using legal vagueness to induce self-censorship. Because the dividing line between permissible and impermissible speech is left unclear in authoritarian societies (Ford Reference Ford2015, 24), Internet users are stuck in a state of informational uncertainty. Although legal ambiguity as to permissive speech is very effective in encouraging self-censorship, this lack of clarity can cut both ways. This state of uncertainty may keep a lid on dissent in general, but public perceptions of what constitutes permissible speech are vulnerable to misinterpretation at the perimeters of collective understanding.

Lacking clear demarcations of what constitutes impermissible speech, the average Internet user, on the (unwarranted) assumption that numbers convey reliability, will look to her peers to determine what exactly acceptable speech is. Like the shopper in our above example who, while inclined to shop at Store X, is nevertheless persuaded by the behavior of other consumers to shop at Store Y, an Internet user may be persuaded by a cyberspeech bandwagon's growing numbers to likewise engage in open speech. Bikhchandani, Hirshleifer, and Welch have shown that in a “general setting with sequential choices … at some stage a decision maker will ignore his private information and act only on the information obtained from previous decisions.… In the absence of external disturbances, so do all later individuals” (1992, 994). This same dynamic may occur with respect to open online speech. Although extreme forms of open speech are clearly understood as impermissible, a great deal of online discourse cannot be easily classified by Internet users, and so on the periphery of what is understood as acceptable expression, the Internet user's private information is extremely weak. It is here, along the margins of public discourse, that the envelope may be incrementally and even inadvertently pushed by the behavior of cyberspeech entrepreneurs.

The model does not require cyberspeech entrepreneurs to engage in extreme forms of political speech, nor does it even have to be performed intentionally. In many instances, cyberspeech entrepreneurs may not even be conscious that they are crossing the acceptable line of expression; their conduct may simply flow from a natural lack of caution and a penchant for audacious speech. This is a unique aspect to our model that distinguishes it from the existing cascade literature.Footnote ¹³ Unlike norm entrepreneurs and availability entrepreneurs, cyberspeech entrepreneurs are not envisioned here as having to push so vigorously or in such an organized manner against prevailing norms and beliefs. Cyberspeech entrepreneurs may act in a highly decentralized, spontaneous, and, most crucially, unintentional fashion. This is due to the unique structural character of online communicationFootnote ¹⁴ —the highly interconnected character of online communication can rapidly amplify small shifts in speech. The interesting component of our model is that the entire process may unfold as an unintended consequence of very small acts. No one may wish for a cyberspeech cascade to emerge, yet one may emerge nevertheless.

In any case, intentional or not, the model assumes that cyberspeech entrepreneurs will only push very gently against the boundaries of permissible speech, and will not do so if it represents a significant departure from their understanding of what is deemed acceptable by the censors. They will not engage in political speech that may threaten their own physical safety and liberty. Thus, each cyberspeech bandwagon will shift perceptions of acceptability only to a very minor, almost negligible degree. The loosening of closed speech is initially slight and proceeds at first incrementally. After building on itself in an almost imperceptible manner, however, the proliferation of cyberspeech bandwagons may eventually reach a point such that a cascade emerges, triggering a large-scale shift in the character of online speech.

A lone cyberspeech bandwagon will produce little more than a slight uptick in open speech. Most of these bandwagons will remain isolated and ultimately peter out. However, because the interlinked structure of the Internet allows for exponential growth, a cyberspeech bandwagon has the potential to spawn additional bandwagons. If this occurs, the process may begin to snowball, and as is the nature of exponential growth, the perceived boundaries of censorship may begin to move more quickly. This process of exponential increase is illustrated in Figure 1.

Figure 1.

Depiction of a Cyberspeech Cascade

Figure 1 visually depicts a cyberspeech cascade, describing it on four levels. The first level shows a cyberspeech bandwagon yielding new ones that, in turn, trigger others. This may exceed the rate of one to two, as illustrated in the diagram's fourth level, which depicts a rate of one to three. The important point is that the cascade can explode rapidly into multiple layers of growth, and from only an initial trickle, it can eventually capture within its sweep millions of Internet users. At each successive level, there are lesser degrees of self-censorship.

Of course, Figure 1 does not capture the complex reality of cyberspeech cascades. We have simplified it for purposes of illustration. Cyberspeech cascades are rarely so linear and precise in how they unfold—in fact, cyberspeech cascades may result in a matrix of complexity. Such may be the dizzying intricacy of the chain reaction that charting its exact pathways of interconnection would be an exercise in futility.

Regardless of the precise circuitry of growth, however, a cyberspeech cascade will conform to a general pattern. Each successive bandwagon further widens the borders of permissible speech ever so slightly but repeatedly so that the boundaries of what is perceived as acceptable quickly recede. Each bandwagon builds on the one before it, further pushing back the tacitly understood boundaries of permissible speech as people misinterpret the previous bandwagon as reliably signaling that a wider ambit of public expression is now officially tolerated. The scale of each bandwagon may range in number—a mere handful to millions of Internet users could be implicated at various stages.

The process can be explained more technically as follows: Bandwagon 1 yields Speech X, where X signifies a slight increase of open speech. Subsequently, cyberspeech entrepreneurs in Bandwagon 2, misinterpreting that Speech X reliably indicates a greater tolerance of uncensored speech, produce Speech X+1, where 1 denotes an even higher degree of open speech. Bandwagon 3 then builds on Bandwagon 2, with cyberspeech entrepreneurs delivering Speech X+2, which bears even fewer indicia of censorship. In sequential fashion, Bandwagon 4 employs speech X+3. This cycle repeats itself in an iterated fashion until a cyberspeech cascade occurs. The effect of the cyberspeech cascade will be self-reinforcing as it reorders public perceptions regarding what constitutes permissible speech. Previously stable norms of self-censorship collapse as online discourse is overwhelmed by the large-scale eruption of open speech.

The tightly bound and hyper-connected nature of the Internet allows even tiny acts of dissent (even unintentional ones) to reverberate and rapidly gain momentum as they ripple out across the Internet. Like a light touch on a spider's web, this impact is amplified and may be felt throughout the entire system. What began as small shifts in speech patterns eventually turns into a flood of open discourse—eddies become ripples and ripples crest into tidal waves. Ironically, the informational uncertainty that is so effective in inducing self-censorship sets the very conditions for cyberspeech cascades.

Online Anonymity Between Users

Anonymity between users varies considerably across Internet platforms.Footnote ¹⁵ For instance, speech that occurs on social networking services, such as Facebook, is far from anonymous and undoubtedly gives rise to powerful reputational concerns. These reputational concerns are especially robust because these connections usually comprise a speaker's close social ties with users who engage in repeated interactions. Online speech in group e-mail trails and chat groups on instant messaging clients—such as Whatsapp, WeChat, and Tencent QQ—that comprise personal (or semi-personal) ties may likewise produce strong reputational concerns. Although the connections are less intimate, professional networking services, such as LinkedIn and ResearchGate, will also generate powerful reputational motives. Users on micro-blogging platforms such as Twitter and Sina Weibo (China's Twitter) also lack anonymity. Compared with other forms of social media, micro blogging is generally more open to the public and typically consists less of personal connections (in the form of followers). The lack of anonymity on these online speech platforms may give rise to powerful reputational concerns.

However, we must be cautious here. Although social networking platforms like Facebook are anything but anonymous, a great deal of online communication in fact occurs in what is the practical equivalent of anonymity. We give the term functional anonymity to the situation in which an Internet user employs a pseudonym (online handles or screen names) that reveals sparse to no information regarding their personal identity. It would be difficult to argue that the use of online aliases generates strong reputational motives. This is especially true when online communication involves predominately one-shot interactions (as opposed to repeated interactions). Moreover, many of these screen names are site specific, with Internet users employing multiple online identities in which they have very little invested.

Indeed, the lack of reputational concerns is clear from the level of vitriol, undisguised racism, sexism, and verbal abuse that saturates online discourse. For instance, the comment threads on YouTube, news sites, blog sites, Internet forums, online chats, multiplayer online games, and other online fora are replete with speech clearly liberated from the constraints of basic civility. It is often the case that with “commenters able to hide behind a cloak of anonymity, the blog and chatroom have become forums for hatred and bile” (Adams Reference Adams2011). Social norms and reputational concerns that would otherwise keep such speech in check offline are routinely ignored by people when they interact online. This fact speaks to the lack of strong reputational motives in these fora—the result of functional anonymity. It is telling that the caustic discourse so commonplace in the comment threads of YouTube videos is rarely seen on personal Facebook pages. The use of screen names is insufficient to trigger reputational concerns. Reputational concerns will not arise, or if they do, they will be comparatively weak. Although star bloggers or YouTube celebrities with millions of followers may be heavily invested in their online identities, this is not true for their millions of followers operating under random screen names posting the odd comment here and there. Even where stable online identities are used (e.g., Facebook profiles) on speech platforms such as the comment threads beneath news stories, a lack of repeated interaction and the one-shot nature of these interactions inhibit reputational concerns of any significant nature.

It should be noted that the point here is not that Internet users communicating in functional anonymity do so with respect to government monitors—potential surveillance does not change with the use of screen names. The point here is merely that there is no, or very weak, reputational concerns with respect to other users. This is important because it means that fear of social disapproval or a desire for social approval will not play a role under such conditions.

It is important that we really appreciate the varied character of online communication. In some cases, speakers possess well-established online social identities and repeatedly interact; in other cases, speakers operate in functional anonymity. The first will give rise to reputational concerns; the second will not. As a result, online communication produces a mix of cascade types. Because online speech operates across a broad range of speech platforms—some that provide anonymity between users and some that do not—a full-blown cyberspeech cascade will involve informational and reputation-related cascades as it sweeps across the Internet. Where reputational motives are actuated, cascades that involve reputational motives, such as availability and norm cascades, will emerge. Let us now examine, in turn, the impact of each of these.

Availability Cascades

To the extent that they involve a reputational aspect and not simply informational issues, cyberspeech cascades are better captured by the concept of availability cascades. Where the speaker possesses a robust online identity, availability cascades may emerge because reputational factors will come into play. Unlike simple informational cascades, availability cascades are also driven by powerful reputational motives. Reputational motives such as earning social approval and avoiding disapproval come to the fore (Kuran Reference Kuran1998, 686). Reluctant to deviate from the script, many Internet users join the growing bandwagon because they wish to be seen as adopting the correct position (i.e., engaging in online speech critical of the government). An Internet user asks: “How will my community view me if I fail to endorse prevailing opinion?” This provides incentive to parrot these views in public (Kuran Reference Kuran1998, 729). Reputational concerns can be very powerful when conditions allow. If an actor believes that the acceptable limits of public expression have really changed,Footnote ¹⁶ otherwise politically indifferent Internet users may engage in online speech that is critical of the government driven solely by a desire to gain social approval and avoid social disapproval. The degree of concern about reputation will vary between people: “people differ from one another along these dimensions; some care enormously about their reputations, while others care very little” (Kuran and Sunstein Reference Kuran and Sunstein1999, 729). However, on the macro level of vast numbers of agents interacting, reputational motives will exert a discernable impact. One need only look at the tidal waves of ideological conformity that wash across Facebook feeds to begin to see how powerful reputational concerns can be in patterning collective online opinion, genuine or not.

Availability cascades also have an informational component. Like informational cascades, Internet users’ uncertainty regarding the precise boundaries of permissible speech facilitates the emergence of availability cascades. However, the availability cascade model invokes an informational mechanism that differs from informational cascades. This mechanism, the availability heuristic, was briefly discussed earlier.Footnote ¹⁷ The availability heuristic “involves estimating the probability of an event on the basis of how easily instances of it can be brought to mind. This heuristic can produce substantial distortions whenever certain alternatives are easier to imagine than others” (Kuran Reference Kuran1998, 706). Kuran and Sunstein give the example of a person overestimating the prevalence of AIDS “simply because many of his acquaintances have the disease and he can easily think of AIDS cases” (Kuran Reference Kuran1998, 706). Applied to online speech, the availability heuristic suggests that an Internet user's understanding of what constitutes impermissible online expression will be distorted by how easily she can recall encountering examples of open speech online (this tendency is compounded by the fact that instances of self-censorship are, by definition, not easily discerned). Thus, a rise in the availability of uncensored speech, particularly if this is concentrated in a localized network of Internet users (e.g., between a cluster of interlinked Facebook accounts), can produce the incorrect impression that the acceptable limits of permissible speech have changed.

Availability cascades have been traditionally used to explain misperceptions of increased risk conflated by media attention (e.g., disease threats, airplane crashes, the threat of terrorism). However, availability cascades may also in principle explain a mistaken belief in decreased risk—that is, the incorrect perception, bolstered by its simple availability, that engaging in open speech online has become less risky when in fact it has not. With their perception of risk now slightly altered, Internet users become more inclined to engage in open speech themselves, and in so doing, further increase the availability of open speech. This triggers a chain reaction of individual responses that make the perception that a wider range of open speech is now being tolerated by the censors appear increasingly plausible simply by virtue of its growing online availability, when in fact a wider range of open speech is not being officially tolerated by the censors (Kuran Reference Kuran1998, 685).

Reputational concerns will have an impact only if actors believe it is safe to engage in open speech. This needs to be emphasized. The average Internet user will not risk personal safety, life, and liberty merely for reputational gains. The perception that a wider ambit of liberal expression is actually being tolerated by the censors remains a necessary condition. When this condition is satisfied, however, reputational motives can powerfully reinforce shifts in collective perception, adding further fuel to the fire. As people join growing bandwagons of open speech motivated by reputational concerns, other users will misinterpret this increase in availability as an indication that the censors have grown more tolerant of open speech.

Norm Cascades

Reputational motives are also very useful in accounting for the sudden, punctuated shifts in Internet self-censorship patterns that have been observed in many instances.Footnote ¹⁸ Sunstein's norm cascade model describes in a very detailed fashion how this may come about. Norm cascades are fascinating in how they play upon collective fear and distort mass perceptions. Again, uncertainty regarding the precise boundaries of permissible speech provides the initial space for norm cascades to arise; however, the reputational component operates in a distinctive manner. The norm cascade model holds that “social conditions are often more fragile than might be supposed, because they depend on social norms to which—and this is the key point—people may not have much allegiance” (Sunstein Reference Sunstein1996b, 909). Many individuals, driven by reputational fear, engage in preference falsification, and so wait in the wings for an opportunity to defy the prevailing norm. Like norm cascades, the present model assumes that a significant percentage of Internet users self-censor out of fear, engaging in preference falsification. An uptick in open speech, however, lowers the perceived danger in engaging in open speech. This shift in perception emboldens Internet users previously reluctant to engage in open speech to join the growing bandwagon.

This continues until a tipping point from predominately closed speech to predominately open speech is reached. At this stage, the cost structure flips so that it is reputationally more costly to appear out of step with the growing chorus of speech critical of the government than remain silent. Misperceiving that the official boundaries of permissible speech have now shifted, large numbers of Internet users who previously self-censored will cease doing so. Even more crucially, however, Internet users otherwise indifferent, motivated by the reputational consequences of not conforming, become more likely to engage in such speech. Alexis de Tocqueville described the dynamic well: “More frightened of isolation than of committing an error, they [join] the masses even though they [do] not agree with them” (de Tocqueville Reference deTocqueville1856, 259). When this stage is reached, a communal shift in self-censorship norms may occur (Roe Reference Roe1996, 663–65; Picker Reference Picker1997, 1250–51) in the punctuated, sudden “triumphant rush” described by Ellickson (Reference Ellickson, Hechter and Opp2001, 51).

This inversion in collective fear has close parallels to the spiral of silence theory in the political science literature and mass communication theory (Noelle-Neumann Reference Noelle-Neumann1974, 43–51). The theory holds that as a particular view begins to gain ascendency in a social environment, individuals adjust their behavior accordingly. The result is that those holding the view become more vocal and those who do not become less willing to speak out, triggering “a spiraling process which increasingly establishes one opinion as the prevailing one” (Noelle-Neumann Reference Noelle-Neumann1974, 44). Spiral of silence theory explains the same en masse shift in social behavior described by the norm cascade model. Fear of voicing a minority opinion cows people into silence, and those holding a majority view grow more emboldened to voice their view.

Mutually Reinforcing

Given the character of online communication, the potential for informational cascades is the greatest. This is because informational cascades are not driven by reputational motives, and as such, may emerge across any speech platform. Availability cascades and norm cascades, on the other hand, may only emerge through speech platforms where reputational motives come into play—that is, where speakers possess a strong online identity. Nevertheless, as cascades of open speech gain momentum and flood across a mixture of online speech platforms, all three cascade types will invariably be ignited. Once this occurs, informational cascades, availability cascades, and norm cascades will feed off one another.

As changes in speech patterns move into areas of the Internet that give rise to reputational concerns, informational cascades will help produce availability and norm cascades by providing the initial uptick in open speech critical to triggering these cascade types. In the cascade literature, availability cascade and norm cascade models require availability entrepreneurs and norm entrepreneurs to achieve this. The present model does not require such committed campaigners for social change. Shifts in speech patterns may occur from the unintended actions of Internet users and extremely trivial incidences of open speech that quickly build on one another in a sequential manner.

In the other direction, availability and norm cascades can in turn powerfully reinforce cyberspeech bandwagons that begin as simple informational cascades. As cyberspeech bandwagons flood into areas of Internet communication that generate reputational concerns, such as Facebook pages or other forms of social media, fear of social disapproval and a desire for social approval will further incentivize actors to engage in open speech. In this manner, cascades in one domain can trigger or reinforce cascades in another. Although informational cascades and cascades that involve reputational concerns may initially emerge in different sectors of the Internet, once arisen, these cascade models may reinforce each other like the three legs of a tripod.

This is an important point because simple informational cascades, the literature tells us, are extremely fragile—they can be reversed by new public information (Bikhchandani, Hirshleifer, and Welch Reference Bikhchandani, Hirshleifer and Welch1992, 994). Conformity “is brittle. The arrival of a little information … can shatter an informational cascade” (Bikhchandani, Hirshleifer, and Welch Reference Bikhchandani, Hirshleifer and Welch1992, 1004). It is thus not difficult to dam up an informational cascade. The new public information “needs only to offset the information conveyed by the action of the last individual before the start of the cascade, even if millions subsequently imitated” (Bikhchandani, Hirshleifer, and Welch Reference Bikhchandani, Hirshleifer and Welch1992, 1006). As such, cyberspeech cascades are not a perfectly stable equilibrium. Cascades of open speech are themselves just as vulnerable to changes in counterinformational cascades.

Reputational Concerns Make Informational Cascades More Robust and Harder to Reverse

This can be discerned in the government control strategies employed by state censors. For example, the use of “Internet commentators” by Chinese censors (some estimates put the number of these Internet commentators as high as 300,000; see D'Hooghe Reference D'Hooghe2014, 55) to guide and redirect online discourse by writing responses to online postings (So Reference So2012, 449) is designed precisely to achieve this end—reversing growing cyberspeech bandwagons by introducing counterbandwagons. Russian censors have similar programs, employing an army of trolls in an effort to shape online speech.Footnote ¹⁹ The Kremlin's Internet Research Agency systematically spreads false information, posting “pro-Kremlin propaganda online under fake identities … in order to create the illusion of a massive army of supporters,” smear political opposition, and neutralize the impact of online speech of an antigovernment flavor.

This administrative arm of the Russian government—informally referred to as the Troll Farm (Chen Reference Chen2015)—strategically shapes public opinion by creating content across various social networks, such as Twitter, Facebook, VKontakte (Russia's Facebook), and the comment threads of Russian news outlets (Chen Reference Chen2015). While it is unlikely that the Chinese or Russian censors conceptualize their efforts in terms of cyberspeech bandwagons and cascades, the goal of such soft control methods is to redirect the flow of online discourse so as to thwart the emergence of cyberspeech cascades (Rawnsley and Rawnsley Reference Rawnsley and Rawnsley2015). These methods of curating online discourse constitute the external disturbances identified by Bikhchandani, Hirshleifer, and Welch (Reference Bikhchandani, Hirshleifer and Welch1992) that may disrupt an informational cascade.

Cascades that possess a reputational component (i.e., availability and norm cascades), however, are far more resilient than lone informational cascades (Lemieux Reference Lemieux2004, 20).Footnote ²⁰ This is because the addition of “reputational factors to informational motivations implies that individuals will ignore their private signals even more often, and that the information made available by everyone's choices will be even more biased toward following the crowd” (Lemieux Reference Lemieux2004, 20). As such, when informational cascades combine with availability and norm cascades, an eruption of open speech on the Internet may be quite difficult to stop once begun.Footnote ²¹

Sequentiality and a Clearer Awareness of Others' Behavior

Informational cascades are open to the criticism that they assume perfect sequentiality with each agent able to observe the behaviors of all the agents before her (Shiller Reference Shiller1995, 183). Actors make decisions sequentially, “with later people watching the actions of earlier people, and from these actions inferring something about what the earlier people know” (Easley and Kleinberg Reference Easley and Kleinberg2010, 484). Yet, situations that exhibit perfect sequentiality are in reality more the exception than the norm (Shiller Reference Shiller1995, 182–83). The Internet, however, better simulates this with users behaving in a highly structured, usually sequential fashion. Indeed, it is often perfectly sequential.

One need only glance at the standard layout of chat rooms, message boards, micro-blog posts, social media such as Facebook, e-mail threads, and the comments to online news articles to appreciate their sequential structure: commenters respond in a successive fashion, able to view previous comments (Holt Reference Holt2004, 15–16; Adams Reference Adams2007).Footnote ²³ The number of likes on a Facebook page or post proceeds in a sequential fashion and, as such, is susceptible to a bandwagon effect as the number of likes reaches a tipping point. This sequential nature is true for Internet chatter as a whole: each user has access, either in its entirety or partially, to the past online behavior of other users—literally at their fingertips.Footnote ²⁴

In the real world, agents do not always act in such a well-formatted sequential pattern and, in general, are not perfectly able to view the behavior of preceding agents. Thus, the informational cascade model's assumption that the behavior of all the previous agents is known is better approximated in an online setting and more readily maps onto online behavior than it does onto other informational environments. This heightened availability of information also increases the likelihood of availability and norm cascades. Belief patterns can proliferate and cluster in semi-insulated corners of the Internet, producing an increased availability of other actors’ behavior. The structure of Internet communication allows for clear and immediate access to these availability swells.

Connectivity and Speed

Behavioral cascades apply more robustly to Internet-based communities also in terms of the networked character and speed of online communication. During the 2009 antigovernment protests in Iran (referenced again later), Iranians began taking to city rooftops in a nightly ritual to shout “Allahu Akbar!” (“God is Great!”) as a form of protest (Mackey Reference Mackey2009). With the practice spreading and in danger of cascading, authorities promptly cracked down on the practice, successfully extinguishing its growth. The speed and hyper-connectivity of online communication, however, renders similar methods of cascade control more challenging to pull off in the cyber world.

The hyper-degree of connectivity the Internet affords is nowhere approximated outside of the cyber realm (Friedman Reference Friedman2007, 516–17). The Internet connects vast numbers of people, transcending geographical, class, and social boundaries. This connectivity is currently experiencing explosive growth with the increased usage of smartphones and other Internet-enabled electronic devices (OECD 2012, 21). The Internet is now in its second stage of development, “evolving from a data network connecting PCs with wires to a much broader network reaching a wide range of new portable devices such as mobile phones and tablet computers” (OECD 2012, 21). The highly interconnected nature of the Internet makes cyberspeech cascades even more likely to occur in that it accelerates the transmission of information, and where communication is not anonymous, will produce powerful reputational concerns. What was once referred to as a global village has now arguably become a global parlor: the collective chatter and discussion of large masses of people is occurring on a scale hitherto unimagined in human history (Ren and Zhai Reference Ren and Zhai2013, 112–13). The astonishing speed at which information can now be transmitted complements this hyper-connectivity. This speed of communication amplifies the effects of informational availability, as well as norm cascades, making sudden perceptual shifts regarding what constitutes acceptable online speech far more likely than in the offline world.

Bounded Rationality

Another reason suggesting that the informational cascade model and the availability cascade model better track online behavior is the limited knowledge that participants have of their fellow netizens. Human beings are not engines of perfect rationality. This is captured by the concept of bounded rationality. Bounded rationality is simply the idea that actors’ decisions are limited by the information they have, the cognitive limitations of their minds, and the time available to make decisions. The informational cascade model assumes that actors are boundedly rational—that is, they make their decisions within a field of incomplete information (Bikhchandani, Hirshleifer, and Welch Reference Bikhchandani, Hirshleifer and Welch1992). A key component of the model is that persons are unable to observe others’ private information directly, but they may make inferences about this private information based on what these other persons do.

Each Internet user is operating from a position of imperfect information regarding other agents. Users only see the online behavior of other agents, but they have no way of knowing if the actions of these other actors are predicated on authoritative information. As a result, users often assume (wrongly) that other actors are working with more perfect information and are therefore inclined to interpret their behavior as conveying reliable information. In our example of the shopper choosing between two stores, if the latecomer shopper does not directly inquire about the decision process of all the other shoppers, she has no way of knowing that their decision to shop at that store was merely the result of an informational cascade. This is precisely the case with online behavior. The behavior of others is highly de-contextualized. There is seldom any way of discovering the motivation of other Internet users and, as result, a general impression that other agents are acting on bona fide information will tend to arise, particularly if we see the same behavior in very large numbers.

Bounded rationality likewise enhances the strength of availability cascades. The cognitive biases that give rise to an availability heuristic are the result of bounded rationality. Essential to Kuran and Sunstein's model is that a society is “composed of boundedly rational individuals who benefit immensely from using cognitive rules of thumb” (Kuran and Sunstein Reference Kuran and Sunstein1999, 690). This cognitive limitation forces actors to rely on heuristics to assess “the magnitude of a risk or the seriousness of an alleged social problem” (Kuran and Sunstein Reference Kuran and Sunstein1999, 689). Because Internet users are operating under conditions of limited information, they are frequently more boundedly rational than offline actors. Hence, we argue, both informational and availability cascades may emerge more robustly through online communication, as the conditions are ripe for their emergence.

Better “Rational” Actors

One final point should be made. Informational, availability, and norm cascades assume a rational-choice model: a cascade occurs because each actor rationally makes her decision based on the behavior of preceding actors or seizes the opportunity to enhance her reputational standing (Sunstein Reference Sunstein1996b, 935; Kuran and Sunstein Reference Kuran and Sunstein1999, 689). Given the cognitive limitations under which actors operate, it is rational to look to the informational signals of others (Kuran and Sunstein Reference Kuran and Sunstein1999, 689). Likewise, it is rational to be concerned about one's reputation (Kuran and Sunstein Reference Kuran and Sunstein1999, 689). Internet users are arguably better rational actors. In the non-cyber world, a host of variables may reframe the information upon which a player is acting. For instance, an actor may attribute the fact that Store Y is full of shoppers not to the fact that Store Y's products or prices are better, but to the fact that its awning is a more attractive color. This additional information may blunt the impact of an informational cascade. On the reputational side of the ledger, knowledge that other actors are merely jumping on a bandwagon in an attempt to merely gain social approval may deflate the reputational appeal of also joining the bandwagon.

However, with respect to online speech that occurs on online bulletin boards or is conveyed through brief tweets via social media, there is a relative paucity of information (particularly where functional anonymity predominates). A vast amount of context is stripped away. The behavior of others is often reduced to a string of utterances, a like on a Facebook page, or a reposting of a message or image. Actors thus form their impressions based on the very narrow slice of information provided them. Such a myopically focused set of information is, arguably, more fertile soil for informational and availability cascades given that in cyberspace a great deal of context—other actor's motivations, external considerations, even tone—is absent. Behavior reinforced by the credibility of large numbers will exert a more powerful impact because there is less information that may correct the mistaken inferences drawn. Under such conditions, agents are better “rational” actors, even if the conclusions they draw are in error.

For these reasons, we argue that online speech is especially susceptible to sudden breakdowns in self-censorship norms—more so than traditional conduits of communication. Many authoritarian governments have not adjusted to this new reality.Footnote ²⁵ Sudden upticks in open speech online, especially in the initial stages, if unchecked, with the passage of time can become increasingly difficult to dam. By dissipating the sense of collective paranoia that undergirds self-censorship, an unexpected explosion of open speech online in the form of a cyberspeech cascade can seriously destabilize, and ultimately bring down, an authoritarian government's control over the Internet.