Introduction
Over the past decade, degradative cyber operations coincided with ongoing or emergent interstate disputes. From the Estonian and Russian debacle over a Second World War memorial, to Stuxnet and the Iranian nuclear programme, and more recently the disruption of the Ukrainian power grid, its strategic utility rests on its potential to threaten an adversary's cyber-enabled objectives.Footnote 1 Advancements in cyber capabilities, however, do not translate into strategic gains. Less than 5 per cent of observed cyber operations influenced interstate relations in favour of the aggressor.Footnote 2 Furthermore, rather than a surge in debilitating cyber operations, previous incidents occurred below the threshold for war and appear to demonstrate restraintFootnote 3 on the part of the aggressors.Footnote 4 One is then left to question the confidence in the domain's revolutionary potential relative to other instruments of power.Footnote 5
Despite the above, there is no shortage of alarm following the disclosure of new cyber operations. Those framed as existential threats are often attributed to state or state-associated actors and involve salient disputes between rivals.Footnote 6 While these interactions have yet to escalate into the physical domain, states are quick to implicate rivals despite the absence of definitive evidence.Footnote 7 This behaviour is puzzling given the rarity of cyber operations and the material, technological, and organisational resources required.Footnote 8 Consequently, we are prompted to ask why do states believe that cyber operations are more common than they are? Furthermore, why is there a belief that rival states are more likely to resort to cyber operations as means to meet strategic ends?
In response, this study argues that this apparent deviation from rationality stems from the decision-maker's failure to update existing beliefs with available information concerning the culpability of a suspected aggressor. It posits that this non-Bayesian approach to judgement is a function of pre-existing beliefs shaped by established rivalry interactions, the responsibilities assigned to decision-makers, and the cognitive load required to process available information. To support these propositions the study employs two between-subject Internet-based vignette experimentsFootnote 9 to surface the occurrence of suboptimal judgements involving the attribution of cyber operations. These emphasise micro-level factors that are often overlooked with respect to interstate interactions in cyberspace.Footnote 10
With these in mind the remainder of the article is organised as follows. The succeeding section presents the underlying theoretical framework that informs the study. This is followed by the design and methodology, which highlights the structure and limitations of the experiments. Following this, the reader is presented with the results and is provided with the general discussion of the findings. Finally, the implications of the study are surfaced with the direction that future research may take.
Before proceeding further, it is important to note that the study does not attempt to argue the novelty of micro-level factors. It, instead, demonstrates the applicability of existing frameworks in attempts to study events within cyberspace. In doing so it challenges the pervasive narrative of cyber exceptionalism to move the current discourse forward towards an empirically and theoretically grounded direction.
Theoretical framework
Non-Bayesian judgements
Alarm over cyber operations are inflated by the possible consequences resulting from the malicious disruption of critical infrastructure. While the term encapsulates a host of actions that vary in scope and severity, only a subset of these are viewed as legitimate national security concerns. While a universally accepted taxonomy of cyber operations remains elusive, incidents are classifiedFootnote 11 as either espionage, disruptive, or degradative cyber operations.Footnote 12
Whereas espionage in the conventional sense is well understood and needs no further elaboration; cyber espionage takes advantage of the increased utilisation of the domain in support of strategic goals. Cyber espionage involves the exploitation of flaws in the underlying infrastructure to exfiltrate privileged information. Moreover, these enable later, and more aggressive, operations.Footnote 13 Disruptive cyber operations, in contrast, typically manifest themselves through defacement and (distributed) denial-of-service (DoS) against an adversary. These require fewer resources than the former and are perceived as a nuisance given its transient and limited effects. Most interstate exchanges in cyberspace are categorised are either espionage or disruptive cyber operations.Footnote 14
Degradative cyber operationsFootnote 15 are characterised by technical sophistication, the organisational maturity of aggressors, and latent strategic potential. These exploit vulnerabilities that are typically out of reach for most actors (state and non-state) without the necessary competencies. While early advocates argued that cyberspace reduces material imbalances between actors, previous incidents prove otherwise. Operational outcomes depend not only on the availability of technical resources, but on the support offered by other policy instruments and enabled by mature organisational processes.Footnote 16 Ultimately, the ability to threaten infrastructure crucial to an adversary's political, economic, and/or military objectives facilitates its perceived utility.
With respect to real-world policy outcomes, the significance assigned to degradative cyber operations is reflected in several national and regional strategies and initiatives. With respect to the former, the recently published United States National Cyber Strategy associates the protection of critical infrastructure with ‘Protecting the American people, the American way of life…’,Footnote 17 effectively echoing previous rhetoric surrounding the potential impact of these types of cyber operations.Footnote 18 Similarly, the Singaporean Cyber Strategy also acknowledges that operations that affect critical infrastructure may result in ‘disruptions which could cripple economies, and lead to loss of life’.Footnote 19 At a regional level, organisations such as ASEAN recognise the economic and societal implications of these operations.Footnote 20 And while it is important to note that these concerns are not necessarily shared uniformly by all political decision-makers given the unequal dependence on these technologies, continued integration of ICT in day-to-day life at a global level suggests that global perceptions could gravitate towards this shared perspective.Footnote 21
The consequences of these operations relative to the growth of cyberspace is made starker by the domain's interconnected and interdependent nature.Footnote 22 As with similarly complex constructs, the deeper and more entangled these become, the greater the difficulty to identify and rectify flaws.Footnote 23 This sense of unknowability and vulnerability is made pressing by the possible discovery and exploitation by malicious actors.Footnote 24 Furthermore, the interconnection between components implies the potential for cascading damage across the domain.Footnote 25 Yet despite such, the above scenario remains unsubstantiated by past interstate cyber operations.
None of the degradative cyber operations observed to date are thought to have been ‘strategically successful’. While there is no doubt that some may be classified as tactically promising, none have managed to influence the behaviour of their respective targets in the manner intended by aggressors.Footnote 26 And while Brandon Valeriano, Benjamin Jensen, and Ryan ManessFootnote 27 identify a handful of successful operations, it is difficult to ascertain whether these were the result of independent use or joint employment with other foreign policy instruments.Footnote 28
Few states possess the technical capabilities required for these operations.Footnote 29 While cyberspace facilitates the diffusion of power, it does not follow that material imbalances are made irrelevant.Footnote 30 A degree of organisational maturity is required to effectively utilise latent cyber power.
Although the proliferation of cyber capabilities enables actors to operate within this domain, their ability to marshal their resources is just as important. Rebecca SlaytonFootnote 31 notes that the strategic potential of cyberspace is only realised through the careful use of skills, intelligence, and technology. The outcome of cyber operations is therefore determined by the ability to access and integrate these resources effectively.Footnote 32
Beyond technological and organisational considerations, the use of degradative cyber operations are limited by the need for restraint. Valeriano and ManessFootnote 33 observe that even with these requirements in their possession, states willingly engage in less effective operations. Citing the need to avoid crossing ‘red lines’, the authors argue that restraint serves to minimise the risk of escalation. And while there has yet to be a case wherein violence in cyberspace has carried over to the physical domain,Footnote 34 Ben BuchananFootnote 35 notes that the uncertainty surrounding the intent of these operations encourages such concerns. In a hypothetical attack against a state's nuclear command, control, and communications (NC3), Erik Gartzke and Jon LindsayFootnote 36 foresee an escalation spiral. Complementing this finding, Jacquelyn Schneider,Footnote 37 in a series of war games, demonstrates that the possibility of escalation is indeed recognised by elite decision-makers when faced with the option of utilising cyber operations.
In lieu of the technological, organisational, and strategic constraints that limit the exercise of degradative cyber operation, why are these perceived to be a frequent occurrence by policymakers and popular media? Authors such as Richard Clarke and Robert KnakeFootnote 38 regard it as a real and imminent threat to national security. Relatedly, a study of news articles from 2008 to 2013 highlights the elevated levels of concern in response to increasingly complex cyber operations.Footnote 39 Similarly, a recent survey in the United States indicates that 73 per cent of Americans perceive ‘cyber terrorism’ as a ‘critical threat’ to the United States.Footnote 40 Finally, the number of states instituting national cyber strategies has more than doubled since 2007Footnote 41 with interest in the politico-strategic implications of cyber conflict emerging in response to notable incidents in cyberspace.Footnote 42
Much like terrorism and other High-Impact Low-Probability Events (HILP), normative strategies in assessing risk fail and lead to an increase in dread risk.Footnote 43 While a host of factors aggravate this deviation from rationality, its causal process is depicted simply as: Fear + Uncertainty + Media Exposure – Experience = inflated assessments.Footnote 44 When these factors are present, decision-makers are prone to resorting to cognitive processes that result in suboptimal judgements.Footnote 45 Consequently, when faced with degradative cyber operations that exhibit the above characteristics, decision-makers are prone to disregarding probabilities through a combination of motivated reasoning and cognitive limitations; resulting in suboptimal judgements. As such:
Hypothesis 1. The misuse of probabilities results in misattribution when responding to degradative cyber operations
Cyber operations often involve states mired in rivalry over salient issues. On the one hand, these relationships frame these interactions as extensions of pre-existing rivalry behaviour and are, thus, stable.Footnote 46 On the other hand, the sudden appearance of novel offence-oriented capabilities may trigger a security dilemma that alters the rivalry dynamics between states.Footnote 47 This possible shift is worrisome as most states with the capacity to invest in cyber operations are both economically and militarily capable. At present, only the United States, China, Iran, Russia, and North Korea appear to demonstrate proficiency in these types of operations.Footnote 48 But given that these are notable conventional (and nuclear) powers, an unforeseen escalation is worrisome. Misattribution stemming from the incorrect use (or dismissal) of probabilities may, at best, result in the development of inappropriate policies while, at worst, lead to an escalatory spiral. The latter is potentially escalatory in cases where targets are assets crucial to a state's strategic or ideological interests.Footnote 49
Since attribution is ultimately a cognitive process where both technical and contextual evidence informs judgement, its outcome is best understood at this level of analysis. Broadly speaking, the study posits that an interaction between hot and cold cognitive processes results in the misuse of probabilities. Hot cognition is characterised by judgements that are affect-driven and manifests as the need to maintain existing beliefs. In contrast, cold cognition occurs independent of affect and is associated with individual cognitive limitations.Footnote 50
The motivated thinker
The notion of a rational actor requires the identification of a solution only once all possible alternatives and outcomes have been evaluated. In this idealised scenario, supporting and contradictory information are used to update existing beliefs. Since Jervis's seminal work, scholars have argued that the availability of new information serves to either reaffirm pre-existing beliefs if congruent or are ignored if contradictory.Footnote 51
Beliefs originate from the need to understand the environment and enables certain cognitive processes.Footnote 52 One of which is to set or influence expectations. For instance, if you routinely observe your neighbour's dog covered in mud and leaving muddy paw prints on the side walk, you associate the presence of these paw prints with the dog. If it occurs frequently, the idea that your neighbour's dog enjoys walking around in mud reinforces this belief. In the future, when the dog is not seen but the paw prints are present, this belief allows one to accept the plausibility of certain propositions (for example, the muddy paw prints came from the same muddy dog) such that judgements that emerge reinforce this existing belief (that is, the same dog made those prints) at the cost of alternative explanations (another dog may have made those prints). Adapting this argument to interstate relations, events such as the Yom Kippur War are indicative of pre-existing beliefs setting expectations.Footnote 53
Researchers posit that motivated reasoning is rooted in past affect-laden experiences. When a decision-maker processes new information, they depend on affect-laden information stored in long-term memory. Once accessed, a specific affect is triggered along with the associated information. A heuristic mechanism evaluates these feelings with respect to the information triggered and reinforces the existing affect irrespective of dis-confirmatory information.Footnote 54 Jonathan MercerFootnote 55 extends this argument and posits that emotions function as an assimilative mechanism for beliefs. Whereas rationalists assume that new information serves to update existing beliefs and eventually converges towards reality, Mercer argues that emotions assimilate data into beliefs; thus reinforcing it. These echo Jervis's rationale for maintaining beliefs and provides the necessary psychological basis for this argument. These beliefs are, in turn, reflected through images held by decision-makers.
Images are the ‘total cognitive, affective, and evaluative structure of the behavior unit or its internal view of itself and its universe’.Footnote 56 This links the construct of images with both affect and belief. With respect to state behaviour, images serve three different tasks: (1) the evaluation of the relative capability of another actor; (2) the perceived threat and/or opportunity given to another actor; and (3) the perceived culture of another actor. The study focuses on the second function and the decision-maker's need to maintain an ‘enemy image’ of a threatening other.Footnote 57
Enemy images are constructs through which other actors are perceived to behave in ‘bad faith’.Footnote 58 While several factors result in the emergence of enemy images, the study focuses on conflict accumulation as a reinforcement mechanism. David DreyerFootnote 59 notes that constant exposure to multiple issues in a rivalry environment reinforces these images and increases the likelihood conflict. Crucial to this notion is the idea that an enemy image is strengthened irrespective of subsequent issues (that is, the dispute need not be over related actions or issues). This lack of differentiation is unsurprising as individuals only recall basic concepts of past events known as gist.Footnote 60 This absence of context allows repeated exposure to negative interactions to strengthen pre-existing images.Footnote 61 Consequently, if the notion that state behaviour in cyberspace is a reflection of rivalry dynamics, it is likely that an accumulative process that strengthens these images is at work.Footnote 62 More worrying, the need to maintain an existing enemy image may further be aggravated by existing processes that may result in attribution in support of a political interest among a small group of decision-makers. As noted by Thomas Rid and Ben Buchannan,Footnote 63 the attribution of cyber operations includes a strategic component during which decision-makers take into consideration their understanding of an adversary's motivation and past behaviour. Given the need to maintain pre-existing beliefs, it follows that:
Hypothesis 2.1. Decision-makers are likely to ignore probabilities to support information aligned with pre-existing enemy images
The need to maintain beliefs cannot exclusively account for suboptimal judgements. The need for cognitive closure driven by associated costs among elites is equally important.Footnote 64 For elites, associated costs are the function of a decision-maker's role and accountability.Footnote 65 This is particularly true in democratic regimes where elites are held accountable for policy failures. Self-interest motivates elites to maintain existing beliefs at the expense of possible alternatives.Footnote 66 Consequently, closure must occur as quickly as possible (urgency) and must last for as long as possibleFootnote 67 (permanence).
Urgency encourages decision-makers to ‘seize’ upon cues that provide immediate closure. For instance, events that appear as a rival's attempt to obtain an advantage are believed to be so. Once seized, decision-makers ‘freeze’Footnote 68 on their judgements when presented with contradictory information. Building on the preceding example, even if a rival's actions are later deemed to be benign, this information is disregarded in favour of maintaining an existing belief. As such:
Hypothesis 2.2. Decision-makers in positions of authority are likely to utilise information that results in immediate closure
The miserly processor
Other non-affective processes are equally informative in determining the tendency of decision-makers to misuse probabilities. Although the underlying political environment suggests a greater role for motivated cognition, the inherent characteristics of cyberspace impose substantial cognitive constraints.Footnote 69 Lene Hansen and Helen NissenbaumFootnote 70 assert that the knowledge-gap between technology experts and policymakers resulted in threat inflation among the latter. Additionally, SlaytonFootnote 71 notes that a lack of technical understanding regarding the difficulties associated with developing offensive campaigns in cyberspace fuels the myth of an offensive advantage within the domain. These biased judgements are less a function of motivated thinking and more a result of inherent cognitive limitations. For this study, suboptimal judgements are believed to occur in part due to the task-difficulty and limited cognitive capacity that constrains processing motivation.Footnote 72 Processing motivation does not refer to the maintenance of specific beliefs nor does it denote efforts to reach a predefined conclusion. It pertains instead to the amount of cognitive effort required to accomplish a task.
The processes and structures of cyberspace are highly abstracted for the benefit of non-experts. On the one hand, this permits the utilisation and expansion of the domain without the need to comprehend its inner workings. On the other, an objective analysis of security incidents requires a nuanced understanding of cyberspace. Together with the politico-strategic milieu of interstate cyber interactions, these burden the cognitive capacities of decision-makers. In response, cognitive shortcuts are invoked to formulate judgements using the least amount of (cognitive) resources. And while cognitive heuristics have been proven to assist experts in formulating judgements, these may promote biases and suboptimal judgements.Footnote 73 While several factorsFootnote 74 enable these biases, the study focuses on information order.
There is no means to guarantee the order in which relevant information is obtained in response to malicious activities in cyberspace. While frameworks exist to improve this process, analysts depend on crucial evidence wherever and whenever available. And while it would be ideal to treat each piece of evidence objectively, recently acquired information may have greater saliency due to the limited capacity of working memory. The demonstration of this recency effect confirms the tendency of individuals to prefer the path of least cognitive resistance.Footnote 75 With respect to probabilities, experiments illustrate the tendency to shy away from cognitively strenuous computation of actual probabilities in favour of using the most recent information provided. As such:
Hypothesis 3. Decision-makers are influenced by the order of information and are likely to privilege the most recent information received
While the importance of motivated reasoning and cognitive heuristics in International Relations and political science has been studied for quite some time, there is no reason to believe that the novelty of cyberspace nullifies lessons learned. While there is no doubt that the nature of cyberspace is distinct from that of the physical domain, there is no reason to believe that existing theories and frameworks are inapplicable. Given the centrality of human actors, it is more than likely that cognitive processes provide crucial insight into state behaviour in cyberspace.
Research design and methodology
Operationalisation
The study utilises two between-subject Internet-based vignette survey experiments to test the proposed framework. The vignette depicts an ongoing territorial dispute between two states (Country A and Country B) involving resource-rich islands in contested waters.Footnote 76 In the vignette, Country A's claim is recognised by the international community. At a certain point in time, Country A's offshore rigs are subjected to a cyber operation that disables these facilities for several hours. After which participants are exposed to treatments corresponding to the relevant variablesFootnote 77 in this study.Footnote 78
As noted earlier, motivated reasoning emerges as a function of enemy images (Image) and existing organisational role (Role). The treatment for images is applied by manipulating Country B's past behaviour. An enemy image is reflected in past non-cooperative and martial behaviour. Correspondingly, a non-enemy image is present when efforts to reach a compromise are observed. The latter serves as the control for this treatment. To test for the impact of organisational role, participants are assigned one of two roles. The first is that of the administrator of Country A's National Cyber Command (NCC) responsible for advising the head of state as to the appropriate policy response. The second is that of one of many analysts responsible for providing his or her superiors with the necessary information. The latter serves as the control for this treatment. Role is interacted with both Image and Order to test for the effect of seizing. Participants playing the role of the NCC administrator are more likely to accept confirmatory evidence.
In testing for non-motivated reasoning, the study manipulates the order in which the information is presented (Order). The original ordering presents the base rate information before individuating information. This serves as the control for the treatment. To test for the impact of information order the sequence is reversed. If decision-makers are indeed susceptible to the sequence associated with information acquisition, later information is prioritised compared to those presented earlier.
After participants are presented with these details, they provide an assessment as to whether Country B is responsible for the cyber operation. Their response is collected in the form of a probability (Probability) from 0–100 per cent.Footnote 79
External and internal validity issues
As with other experimental studies, issues of external and internal validity are often a point of contention. Chief among these are concerns over the study's external validity. To begin with, the study does not claim generalisability. These results are only applicable with the sample in hand and generalisation beyond this is discouraged. Moreover, with experiments involving elite decision-making, claims made must be taken with a degree of caution given the underlying differences between elites and non-elite samples.Footnote 80
To mitigate issues concerning external validity the study employs two separate samples: students and government employees.Footnote 81 This considers the possible variation between these two groups. In addition, the study utilises the Internet to recruit participants to provide a wider demographic reach that serves to mitigate any culture-specific effects.Footnote 82 Finally, the design is strengthened by providing a scenario that integrates aspects of real-world events in an attempt to increase the impact of this fictitious scenario.
Besides external validity, questions of internal validity need to be addressed as well. Two stand out in this regard: the pre-treatment effect and engagement with the study. The former refers to the possibility that participants may already possess knowledge related to the study that may skew the results towards their pre-existing beliefs. This may come about due to exposure (directly or indirectly) to similar cases or through formal education or first-hand experience. While limiting contact with similar events is unlikely given the media coverage surrounding cyber operations, it is possible to address this by not restricting participation to individuals from a specific educational background thus increasing variation within the samples.
The issue of engagement, in comparison, requires a more stringent approach. Participants of Internet-based experiments tend to be disengaged with the material (that is, less attentive). This limits the effectiveness of the intended treatment. To overcome this issue, four (4) attention check questions are used within the experiment. Failure to correctly answer all four questions results in the exclusion of a participant. Previous research demonstrates that Internet-based experiments can expect an exclusion rate of 30–50 per cent.Footnote 83
Participant recruitment
The study recruits its participants through the online platform Prolific. While the debate continues regarding the validity of results obtained from Internet-based experiments, research shows no significant difference between this and conventional lab-based environments.Footnote 84 However, issues pertaining to external and internal validity demands care during participant recruitment.
To begin with, the study is conducted using two separate groups. The first consisting of undergraduate students and the other of government employees.Footnote 85 The former represents the usual approach in these types of experiments.Footnote 86 The latter, however, is used to allay fears concerning the non-representativeness of the sample.Footnote 87 Given the difficulty of recruiting elites, government employees serve as a readily accessible proxy.
For both groups, additional steps are taken to ensure the rigour of the design. First, only those with English as their primary language are invited to participate. This addresses concerns regarding comprehension that may alter the accuracy of the response. Second, participants are not selected based on a specific educational background. This increases the variation between those with greater knowledge about cyberspace and those with less. Finally, only those with an approval rating of 90 per cent or higher are accepted. Prolific grades possible participants according to the number of times their responses have been accepted by other researchers. This aids in reducing the engagement problem discussed earlier.
Once the participants are selected, they are randomly assigned to one of eight (8) possible scenarios that correspond to the manipulation of Image, Role, and Order. Upon completing the experiment, participants are rewarded US $0.35 for their participation in the three to seven-minute-long activity.
Key limitations
Before proceeding further, the limitations of this design must be surfaced to better frame the context in which the results are interpreted. First among these is the limited generalisability of the findings. While experimental designs allow unprecedented control over the variables of interest, it comes at the cost of wider and more complex processes that readily occur in the social world.Footnote 88 Although the manipulation of Image, Role, and Order permits us to establish their function in the formation of judgements, this comes at the cost of other externalities. Moreover, the manipulation of these variables requires the (over) simplification of real-world processes that results in a much narrower analysis of this phenomenon.
Second, despite attempts to recruit individuals more closely aligned with those expected to face similar judgmental tasks, there is no guarantee that participants will behave in a similar fashion. Key differences are noted between elites and non-elites with regards to the formation of judgements. Elites are believed to be: (1) less prone to loss aversion; (2) may be better at multi-stage thinking; (3) able to utilise heuristics more effectively; (4) prone to overconfidence; (5) and are thought to be more cooperative.Footnote 89 Having participants that better fit the notion of non-elites would result in findings that do not account for the behaviour of the population of interest.
Third, foreign policy judgements are rarely, if ever, made in isolation. These are subject to bureaucratic and organisational processes that either enforce or temper individual idiosyncrasies. This study, however, avoids this process in the interest of analytical parsimony. At most, the importance of the organisation is collapsed into the construct of Role and its influence on individual judgement. In doing so, the study is unable to elaborate on the persistence of biases once individual perspectives are subjected to organisational and bureaucratic processes.
Fourth, the study lacks the dynamic nature of real-world cases of cyber operations. Information is not static and the process of obtaining evidence may, itself, be subject to a separate set of biases that ultimately skews judgements concerning accountability. For instance, evidence may be collected by an organisation in a manner to support its own pre-held conceptions of reality. Consequently, this skews the judgement made by the information consumer. The design, as it stands, is unable to account for such.
Yet despite the above limitations, the findings surfaced by this study are not without merit. Instead, these constraints guide the reader as how best to interpret the results presented in the succeeding section. With respect to other scholars, these limitations invite further studies that could either support or refute conclusions that emerge from the present design.
Experimental results
Experiment 1. Undergraduate sample
For this experiment a total of 448 participants were recruited. Of these, 36.61 per cent (164) were removed after failing the attention check questions. To ensure a balanced analysis, random samples were drawn based on the size of the smallest treatment group resulting in 176 samples with 22 samples per treatment group. This sample had 55.11 per cent (97) male and 44.89 per cent (79) female participants with an average age of 25.46 years. Geographically, most participants were from either Europe (47.73 per cent) or the Americas (44.32 per cent). On average, participants took approximately 2 minutes and 57 seconds to complete the experiment. The average probability that Country B was responsible for the cyber operation is 21.65 per cent.
To determine the effect of Image, Role, and Order on Probability, a Type I Analysis of Variance (ANOVA) is performed. In the model, the interaction terms Role x Image and Role x Order are used to test for the presence of seizing identified in the above framework.
The results highlight a significant Average Treatment Effect (ATE) due to Order F(1,163) = 4.612 treatment on Probability at the p < 0.05 level with an effect size of f = 0.166. The small effect size is not exclusively associated with sample size. The extent to which extraneous factors were controlled for, the specific manipulations, and the medium used for this experiment may have reduced the magnitude, but not the significance, of the treatment's impact on participants.Footnote 90 None of the other variables and interaction terms were found to be statistically significant. A post-hoc comparison further reveals that the main effects of Order are significant at the p < 0.05 level. When individuating information is presented first, Probability increases by 7.58 percentage points.
Experiment 2. Government employee sample
For this experiment a total of 423 participants were recruited. Of these, 40.9 per cent (173) were removed after failing the attention check questions. To ensure a balanced analysis, random samples were drawn based on the size of the smallest treatment group resulting in 200 samples with 25 samples per treatment group. This sample had 37 per cent (74) male and 63 per cemt (126) female participants with an average age of 36.49 years. Geographically, most participants were from either Europe (59.5 per cent) or the Americas (37 per cent). On average, participants took approximately 3 minutes 16 seconds to complete the experiment. The average probability that Country B was responsible for the cyber operation is 21.82 per cent.Footnote 91
To determine the effect of Image, Role, and Order on Probability, a Type I Analysis of Variance (ANOVA) is performed. In the model, the interaction terms Role x Image and Role x Order are used to test for the presence of seizing identified in the above framework.
The results highlight a significant Average Treatment Effect (ATE) due to Image F(1,194) = 8.787 treatment on Probability at the p < 0.05 level with an effect size of f = 0.213. None of the other variables and interaction terms were found to be statistically significant. A post-hoc comparison further reveals that the main effects of Image are significant at the p < 0.05 level. When an enemy image is present, Probability increases by 9.32 percentage points.
General discussion
While the mean of Probability for both experiments were not statistically different from one another (x1 = 21.653, x2 = 21.82, p = 0.994), these were not consistent with the predicted value of 14.3 per cent of a perfectly Bayesian actor.Footnote 92 Despite this, neither experiment reflected the tendency of decision-makers to ignore base rates outright. Over half of the participants in both experiments identified Probability with that of the base rate (that is, 10 per cent). This suggests that effort was exerted to reach a sufficiently acceptable, but not necessarily optimal, judgement and highlights the fact that humans tend to act as satisfiers rather than maximisers.Footnote 93 Further analysis reveals that for both experiments Probability ranged from 10 per cent to 60 per cent suggesting that both individuating and base rate information were considered by participants in their final assessment. This allows the claims of Hypothesis 1 to be challenged but surfaces a separate line of inquiry. If base rates are not neglected outright, then what factors cause the limited use of both the base rate and individuating information resulting in suboptimal judgements?
Information order and bias
For Experiment 1, Order is the sole factor influencing the participants’ judgements. Hypothesis 3 postulates that given cognitive constraints, individuals use the most recent information made available. This recency effect is well documented in the literature but appears to be absent in this case. Instead, the results suggest that earlier information plays a larger role in the inflating Probability and that the primacy effect is what is being observed instead.
The primacy and the recency effects illustrate the tendency of individuals to recall the first and last items in a series best. These are the consequences of limited cognitive resources. The primacy effect is believed to occur since initial information is more effectively stored and accessed in long-term memory. In contrast, the recency effect occurs due to the limited availability of short-term or working memory, which requires less effort to access. However, the recency effect is mitigated when additional tasks involve the use working memory.Footnote 94
Individuals can store approximately seven (7) items (plus or minus 2) in short-term memory at any given time.Footnote 95 With respect to the vignette, participants would have to take note of at least seven (7) details prior to the presentation of base rate and individuating information. This increases the likelihood of using earlier information (either of the above depending on Order) in determining probability due to the cognitive limits imposed.
Empirically, the fact that the value of Probability is higher when Order is reversed supports this claim. If the recency effect is present, then treatment groups where individuating information is presented last should have a higher Probability compared to when it is provided earlier. However, the results illustrate that Probability is, on average, 7.58 percentage points lower in these cases. Inversely, when base rates are presented last, Probability increases. This implies that participants access information provided earlier more readily than those received later. This may be due to the exhaustion of working memory and the need to retain as much information as possible.
Alternatively, another reason for this observation may be due in part to the structure of both the individuating and base rate information. Other experiments demonstrate that both format and length account for the decision to use or ignore specific pieces of information.Footnote 96 Longer formats or mathematically complex ones encourage subjects to choose simpler and easier to comprehend alternatives. This, however, cannot account for preceding observation as the vignette was structured in such a way that both individuating and base rate information are reasonably similar in terms of length and format. As such, it is unlikely that tendency to favour the initial information is a function of its appearance.
Similarly, the effects of a limited attention span may also be dismissed. The experiment employs attention checks meant to ensure that participants are focused on the task at hand. Those that fail this are removed from the study. Similarly, the experiment only involved participants that have been favourably rated by other researchers. While this does not necessarily guarantee their attention, it does suggest that a certain level of engagement with the material.
Given the above observations and the unlikely occurrence of alternative explanations, this weakens the proposition of Hypothesis 3 regarding the recency effect but does not discount the importance of the order in which information is presented. This finding raises three important implications regarding the attribution of these events.
Procedurally, evidence in the wake of a cyber operation is not received or processed in a fixed order. Although Rid and BuchananFootnote 97 have suggested the Q-Model in which tactical, operational, and strategic information are best processed in that sequence, they acknowledge that the process can start in any of these stages depending on the availability of information.
Organisationally, the continued shortage of cyber security expertise in government institutions has stimulated programmes that recruit from the private sector. These individuals enter these roles without the experience and, more importantly, without motivated biases that their counterparts from government may have. Methodologically, this aligns them more closely with the profile of students participating in this experiment and may manifest the same behaviours. As such, processes and technologies aimed at managing the volume of information are a worthwhile investment to avoid bias such as that encountered in Experiment 1.
Politically, and perhaps more importantly, if more recent information is treated as salient, actors involved in persistent disputes in either the physical or virtual domain are likely to ignore existing probabilities. As cyber operations are increasingly employed as adjunctive tools, these are likely to be framed in the context of the existing relationship between parties involved. Furthermore, a sudden escalation in the virtual domain may be perceived as an escalation of the conflict itself.Footnote 98 For instance, the disruption of the Ukrainian power grid was quickly attributed to the Russian regime as an escalation of the conflict when in fact subsequent analysis showed that these systems had been compromised prior to the conflict and may not have been linked to the desire to escalate.Footnote 99
And while the above case and subsequent research has yet to demonstrate the occurrence of cross-domain escalation, its absence does not discount its possibility in the future. Consequently, the tendency to rely on earlier information increases the risk of misattribution, threat inflation, and possibly escalation in response to malicious events in cyberspace.
Enemy images and bias
Although it may be comforting to suggest that motivated bias may be done away with by recruiting outside government or military institutions, most national cyber security organisations are staffed by these individuals. As such, their pre-existing beliefs become a crucial component in judgement. The results of Experiment 2 demonstrate that the effect of Images in the context of cyber operations mimics that of other conflicts such that judgements are formed in accordance with pre-existing beliefs. Image, however, only appears to be of consequence with participants in the second experiment. Accounting for the difference between samples may be found in the fundamental characteristics of participants and the overarching theme presented in the vignette.
The second experiment involved noticeably older individuals who were, on average, at least nine years older than those in the first. This age gap suggests that the former may have a greater interest in politics than the latter as they may operate outside an academic environment and would have a greater stake in day-to-day affairs.Footnote 100 Complementing this difference in priorities, the theme presented in the vignette (that is, territorial disputes) may resonate more given its political, economic, and social implications. The attempt to seize territory could have increased the salience of the treatment with participants who are more attentive with respect to current events and are politically engaged. This observation does not necessarily raise questions regarding the validity of this experiment but instead highlights the possibility that exogenous factors can have a direct influence on experimental outcomes. In this case, it serves to reinforce the impact of the treatment used (that is, Images) and results in Probability increasing by 9.32 percentage points when participants are treated to an aggressive and uncooperative neighbour. As with the student sample, participants in this experiment appear to integrate both base rate and individuating information into their evaluation. Order, however, is not statistically significant.
This does not discredit Hypothesis 2.1 but requires its assumed outcome to be relaxed from outright neglect to that of an inflated assessment. The importance of this finding cannot be disregarded given ongoing efforts to securitise the domain. Since 2007, the number of states with national cyber strategies has more than doubled. Furthermore, with states investing in offensive capabilities; the effect of casting cyberspace as a domain of conflict between rivals risks the possibility of a security dilemma and unintended escalation.
The nature of cyberspace finds parallels with the enduring challenges of ascertaining another state's capabilities.Footnote 101 The limited lifespans of cyber weapons, to complicate matters, renders public demonstrations infeasible. Motivated by the fear of a ‘bad actor's’ intentions, states may engage in pre-emptive cyber operations to obtain intelligence regarding a rival's cyber capabilities. These operations require the compromise of a rival's cyber infrastructure to serve as either: (1) a channel through which sensitive information is exfiltrated; or (2) as an initial stage of a larger and (potentially) damaging operation. While the former is accepted as routine for interstate relations, the latter signals malicious intent that ought to be contained. Unfortunately, cyberspace is not conducive for signalling that allows for these to be distinguished from one another and may result in an unintended security dilemma.Footnote 102
The tools employed to exfiltrate information can be modified to disrupt or degrade crucial systems. Short of an initiator publicly declaring their intent, the discovery of these operations can, at worst, result in an escalatory spiral between states.Footnote 103 Should the victim of these pre-emptive operations harbour pre-existing beliefs regarding the malicious nature of the initiator, it is likely that the former would respond in a manner that reinforces this belief.
Given the absence of norms regulating action in cyberspace, the influence of pre-existing beliefs suggests that a state's response is informed by its experience with a suspected aggressor. To minimise losses, states faced with a consistently belligerent adversary may opt to develop offensive capabilities that serve to deter further attacks or to respond in kind. Consequently, the suspected aggressor may increase its capabilities in cyberspace resulting in an arms race between these states. While most scholars are sceptical as to the extent of damaged possible through cyber means alone, this escalation does not serve to ease tensions in an already unstable and hostile cyberspace.
The influence of role
A curious outcome of this study is the seeming insignificance of Role for both experiments. The influence of organisational role and accountability is well documented in the literature. As a factor that aggravates or mitigates bias, its muted effects are surprising to say the least. For both experiments, the lack of significance appears to be confirmed by its small effect size and limited statistical power (f1 = 0.046, 1 – β1 = 0.091 and f2 = 0.006, 1 – β2 = 0.051). These observations lead to three possible explanations to account for this outcome.
One is the inability of the treatment to elicit the desired affect from the participants. Given that both samples are unlikely to have ever found themselves in a position of authority, it is possible that simply stating their role in this vignette as the administrator of the NCC did not form a substantive affective association with the possible consequences of tarrying or rushing into a decision. A possible work-around for future experiments is to engage in power priming participants. Galinsky et al.Footnote 104 demonstrate that an individual not normally in a position of authority could be primed to act in such a way as if they were through specific recall tasks. Inversely, individuals could also be primed to behave as if they were in not in a position of authority.
A second possibility that complements the first is the absence of consequences associated with the failure to adhere to the expectations of a participant's assigned role. As was stated in the preceding theoretical framework, costs are often associated with positions of greater responsibility. Failure to act accordingly may result in punishments stemming from organisational/bureaucratic practices or from the electorate (for example, not getting re-elected). Experimentally, these consequences may be simulated through graduated rewards based on performance. For this study, the use of a uniform reward system may have omitted this consequence-based reasoning from the participants resulting in Role being statistically insignificant.
A third possibility is that cultural factors may contribute directly to participant behaviour. Despite efforts to increase variation using Internet-based experiments, participants were predominantly from the United States and Europe. Although variations in behaviour as a function of culture is by no means novel, the results of the previous experiments do not allow for this assumption to be tested.
While the first two possibilities may be addressed through a redesign of the experiment to use priming and a graduated reward system, administrative and logistical considerationsFootnote 105 rendered this infeasible at the time. Instead, a third experiment was conducted using Amazon's Mechanical Turk (MTurk) to account for cultural differences. The decision to perform this on MTurk was due to its demographic differences with Prolific; specifically its larger number of Asian participants.
A total of 120 participants with 15 participants per treatment group were recruited for this experiment. Of these, 25.83 per cent (31) were from Asia. When modelled to include the dummy variable Region, only Role (f = 0.310, p = 0.001) and Region (f = 0.274, p = 0.017) were found to be statistically significant at the p < 0.05 level.Footnote 106 Probability increased by 14.33 percentage points when participants were from Asia compared to those from the Americas. An increase of 15.04 percentage points was also observed when assigned an administrative role. There is, however, no significant interaction between Region and Role. But this is most likely due to a lack of statistical power for this experiment (f = 0.144, 1 – β = 0.255).
This follow-up experiment suggests that cultural factors may moderate the impact of Role in the formation of judgement. Given the relatively small sample size, however, confidence in this outcome is in question. Moreover, the importance of having an affective association with a given role and the threat of consequences cannot, as of yet, be ruled out. Consequently, these findings do not allow us to confirm or refute Hypothesis 2.2.
Theoretical, methodological, and policy implications
The result of the experiments reflect the apprehension associated with degradative cyber operations. While base rates were not found to have been dismissed outright, judgements tend to inflate the likelihood of state involvement resulting in suboptimal judgements such as misattribution. The complexity that characterises cyberspace encourages decision-makers to resort to heuristics to reduce the cognitive load required. Although this permits faster processing, it increases the likelihood that only a handful of facts are considered, resulting in biased judgements. In the experiment, the appearance of higher probabilities earlier in the vignette prompted participants to base their assessments on these values resulting in a stronger belief that Country B is indeed the aggressor. Similarly, the context that surrounds these events encourages the maintenance of pre-existing beliefs. When informed of State B's past aggression, decision-makers perceived a greater probability of their involvement. This serves to reinforce the idea that Country B is indeed a bad actor willing to further their own interests.
Theoretically, these findings are neither novel nor surprising. The role of both heuristics and motivated reasoning in political decision-making has been extensively studied. These concepts, however, remain understudied vis-à-vis cyber operations. Consequently, the results contained herein are a decisive development within the field for two reasons. First, they serve to dispel the myth of the domain's exceptionalism. While there is no denying that the manufactured nature of cyberspace enables actors to behave in ways different from the conventional domains of air, land, sea, and space; judgements are ultimately the product of human cognition. As such, these outcomes are subject to cognitive limitations – motivated or otherwise – that manifest themselves across operational domains.
Second, this behavioural shift in the study of interstate interactions in cyberspace complements the ongoing efforts at the state and systemic levels. Despite its emergent status, the cyber security literature is rich with studies that leverage existing state-level theories to explain the use of cyber operations. Recently, this has served to temper persistent belief in the revolutionary potential of these operations but fails to penetrate the black box surrounding decision-making processes. Experiments such as this extend the findings of earlier efforts to better evaluate the behaviour of individuals that make up the state and are ultimately tasked to respond to threats and opportunities within the domain.
Methodologically, the study imparts important lessons for future researchers. Since its earliest days, the field of cyber security has relied on observational studies to both propose and develop appropriate theoretical frameworks. The opaque nature of this phenomenon, unfortunately, results in limited, and possibly, biased data. Consequently, this study demonstrates the feasibility of an experimental design in appraising this phenomenon. However, two caveats exist in this respect. First is the importance of selecting appropriate samples. While undergraduate students are easily accessible, their knowledge and experience may prove limited as argued by other scholars. Although this has not negatively impacted this study, this is not always the case. Therefore, careful participant selection is needed prior to the start of any experiment. Second is the recognition and study of organisational and bureaucratic processes associated with foreign policy. The study did not tackle this aspect and it is prudent to point out the policies are not the sole creation of an individual. Even the most despotic of regimes depend upon a small circle of advisers. While the observed biases are unlikely to completely disappear when decisions emerge from group interactions, these may either be tempered or aggravated in kind. As such, future studies would do well to develop more complex experimental designs that better simulate the distributed and shared nature of decision-making.
The results of the experiments also highlight the continued need for technologies and processes to assist in the analysis of security incidents. Even the best-trained analysts are constrained by their cognitive abilities. It is then appropriate to invest in tools that reduce this workload that, in turn, minimises dependence on fast – but possibly flawed – cognitive short-cuts. Similarly, scholars such as Rid and BuchananFootnote 107 have called for processes that divide analysis between groups and individuals best suited to analyse information at different points in the analytical process. Building on this argument, national cybersecurity organisations may also benefit from hiring individuals from a diversified pool. An increase in variation may serve to temper biases prevalent in one group compared to another.
While the outcome of this study serves only as an initial foray into a larger research programme, its significance should not be judge as trivial. Its theoretical, methodological, and policy implications validate the need to broaden our study of cyber security to include individual behaviour as a crucial aspect of this domain.
Author ORCIDs
Miguel Alberto N. Gomez, 0000-0001-8575-2188
Acknowledgements
I would like to extend my gratitude to my colleagues at the Center for Security Studies at the ETH for their initial comments, which helped to set the tone of this article. I would also like to acknowledge my reviewers, whose comments serve to strengthen the arguments presented herein.
Miguel Alberto Gomez is a Senior Researcher at the Center for Security Studies at the ETH in Zurich. His research explores the role of heuristics and cognitive bias in the formation of cybersecurity policy and interstate interactions in cyberspace.
Appendix
Vignette structure
For the past ten years, a territorial dispute has existed over a series of islands in the sea that separates your country from your southern neighbour. Although unpopulated, maritime surveys have discovered rich oil deposits in the area. Your country's claim is based on the location of these islands within your territorial waters. In contrast, your southern neighbour insists that it has an underlying historical claim represented in a series of antiquated documents. Two years ago, both countries presented their respective cases to the International Court of Arbitration in The Hague where the court confirmed your country's legal right to the territory. In response, your country has begun to deploy offshore rigs to harvest these oil deposits.
Image
Non-Enemy
In previous territorial disputes, both you and your southern neighbour have cooperated with one another. Despite your respective self-interests, both are motivated by the need to maintain peace.
Enemy
In previous territorial disputes, your southern neighbour deployed its military to occupy the contested territory. Despite efforts to ensure peace, your southern neighbour appears to be motivated primarily by self-interest.
Yesterday, a cyber attack targeted your country's off-shore rigs in the disputed territory. The attack appears to have employed advanced tools and techniques and disrupted operations on these off-shore rigs for several hours.
Role
Non-Administrative
As one of many analysts with your country's National Cyber Command, you are tasked with evaluating the available evidence to provide your superiors with information regarding cyber attacks against your country.
Administrative
As the head of your country's National Cyber Command, you are tasked with providing the head of state with information regarding cyber attacks against your country that will be used to formulate the ensuing response.
Order
Original
Trend data indicate that 10 per cent of cyber attacks are initiated by foreign governments while the remaining 90 per cent originate from non-governmental sources (for example, criminal organisations, private individuals, etc.). Although the investigation is still underway, those involved have begun the process of identifying the source of the attack. Early reports concerning the source of cyber attacks are correct 60 per cent of the time but are incorrect for the remaining 40 per cent.
Reversed
Although the investigation is still underway, those involved have begun the process of identifying the source of the attack. Early reports concerning the source of cyber attacks are correct 60 per cent of the time but are incorrect for the remaining 40 per cent. Trend data indicate that 10 per cent of cyber attacks are initiated by foreign governments while the remaining 90 per cent originate from non-governmental sources (for example, criminal organisations, private individuals, etc.).
As information continues to be collected, you have been asked to provide an initial assessment of the situation.