1 Introduction

The most well-known groupthink model (Janis 1972) has been criticized for its linear structure and the exclusive relationship between the antecedents and quality of decision-making. Since the assumption that groupthink leads to decision-making failures is based on its linear relationships, it has been argued that curbing its antecedents is necessary to prevent organizational fiascos. Previous studies on groupthink have not only criticized its power of explanation but also suggested considering the phenomenon’s different aspects, such as the psychological or political. In these arguments, some solutions have been presented, such as particularistic interpretation of the mode (Turner et al. 1992), dynamic factors (Hart 1991), and empirical analysis (Turner and Pratkanis 1998a). Despite these alternatives, however, their findings have not supported Janis’ groupthink model consistently. Consequently, collective intelligence, instead of groupthink, has received more attention in the group decision-making field—since Levy et al. (1995) proposed this concept in the social system, groupthink has been considered as a failed case of collective intelligence. In particular, information communication technology (ICT) has stimulated collective behaviors of both online and offline groups. Hence, groupthink discussions are moving toward collective intelligence with the development of information technology (IT).

A commonality between collective intelligence and groupthink is the tendency of concurrence seeking. Collective intelligence has been studied as a concept related to the positive rather than negative aspects of the effects (e.g. Leimeister 2010; Maleewong et al. 2011) and emphasizes the dynamic aspects (Kim et al. 2008; Mayo-Wilson et al. 2013; Schut 2010). Most studies have focused on investigating the characteristics, mechanisms, and effects rather than the relationship between collective intelligence and groupthink. The finding that the collective may be produced by interpersonal communication made collective intelligence an emergent phenomenon (Bates and Gupta 2017; Bigham et al. 2015; Woolley et al. 2010). Contrary to groupthink, the framework of collective intelligence is based on a dynamic perspective rather than a set of linear relationships; therefore, it is hard to directly link the two.

The separation of groupthink and collective intelligence is a result of a lack of clear understanding. Since it is impossible to exclude entirely the possibility of a correlation between them, this study attempts to discover the key factors that connect the two and their effects. It, therefore, defines “switching factor,” which plays an important role in the transition between groupthink and collective intelligence. Furthermore, we have examined three switching factors: knowledge conflict, reconsideration, and storage. The goal of this study is to understand the effect of these three and the optimal strategy based on it.

Collective intelligence is an efficient method of solution creation for an organization (Luo et al. 2009; Reia et al. 2019). However, developing it is fundamentally different from both organizational interactions and the decision-making process. Although either interactive behavior or individual capability is closely related to the group’s performance, this is a field of collaboration rather than collective intelligence (Driskell and Salas 1992; Hansen and Vaagen 2016). Thus, the existing strategies may not encourage collective intelligence. This is where this study contributes—it provides a strategy for developing collective intelligence when groupthink occurs or is suppressed. Ultimately, the strategies arrived at in this research will be helpful in solving complex problems that require collective ability, such as enhancing a firm’s knowledge creation, formulating an effective policy, and evolving social knowledge.

In Sect. 2, a comparison between groupthink and collective intelligence and their relationship are presented, as per previous studies. We suggest that knowledge conflict is a potential “switching factor” of collective intelligence. In Sect. 3, the effect of knowledge conflict on collective intelligence is investigated through an agent-based model (ABM). Section 4 proposes the optimal combinations of the three switching factors for transforming groupthink into collective intelligence. For this, meta-frontier analysis (MFA) is carried out, which compares the efficiency frontiers of heterogeneous groups. In Sect. 5, the summary and interpretation of the results are provided. Finally, in Sect. 6, the practical implications and limitations of this study are given.

2 Theoretical background

2.1 Groupthink and collective intelligence

Development of technology can be a critical factor for collective intelligence. In particular, since information flow is a key input in developmental activities (Benkler 2006; Mchombu 2003), ICT has radically extended the boundary of human capabilities through strengthened social networks, accessibility, and capacity (Smith et al. 2011). Therefore, individual intelligence has become insignificant when an organization builds internal collective intelligence (Täuscher 2017). Lykourentzou et al. (2011), and Maciuliene and Skarzauskiene (2016) made a similar argument—that ICT platforms such as the internet, social network service, and online portals can contribute toward evolving the collective capability of an organization.

Interaction-based characteristics such as diversity, social sensitivity, connectivity, and accessibility have played a dominant role in developing organizational capability, because collective intelligence began from a dynamic perspective. Skaržauskienė et al. (2013) presented that there are three major dimensions and that collective intelligence can be induced through their interaction and ICT; this means it is possible to induce it through systematic methods. Therefore, collective intelligence may be one of the organizational capabilities rather than a temporal phenomenon (Table 1).

Table 1 Comparison of groupthink and collective intelligence
Full size table

Despite of gap between the concept of groupthink and collective intelligence (see Table 1), Previous studies have tried to ascertain the links between groupthink and collective intelligence in various ways. Solomon (2006) argued that groupthink can be transformed into wisdom of the crowd when two conditions are satisfied. He suggested organizational diversity and decentralization to prevent the “tipping point” of groupthink. However, his was a conceptual study, so it lacked a detailed investigation to understand the mechanism of interconversion between groupthink and collective intelligence. In this context, several studies have proposed plausible ideas for such interconversion. Erdem (2003) introduced the concept of trust to explain why distrust groups can have better organizational performance than trust ones. In that study, it was shown that excessive mutual trust increases groupthink and decreases teamthink. Contrary to that, an appropriate level of trust can be better. Reia et al. (2019) proposed that the way to collaboration can determine whether an organization falls into groupthink or develops collective intelligence. According to their study, sharing limited knowledge, rather than all of it, is likely to induce collective intelligence. In addition, Jafari et al. (2015) conducted a social experiment on the virtual space. Their study presented two aspects as determinants of whether an organization will face groupthink or have collective intelligence: diversity and creativity.

If groupthink can be changed to collective intelligence, the main problem is how to realize this change. As mentioned above, groupthink and collective intelligence have similarities in their perspectives on mechanism. Thus, we hypothesize that their intersections could be a source of interconnection. To that end, switching factor is defined as an essential determinant that transforms an organization’s groupthink situations into collective intelligence ones. For that purpose, Previous studies on groupthink and collective intelligence have presented three common factors for preventing groupthink and promoting collective intelligence, as shown in Table 2; in this study, switching factors are selected based on the intersections between the groupthink solutions and determinants of collective intelligence proposed in those studies (Table 2).

Table 2 Switching factor as an intersection of groupthink and collective intelligence
Full size table

2.2 Switching factor

The first factor is “organizational conflict.” When multiple knowledge exists in an organization, people evaluate each based on their own background (Heit and Bott 2000) or prior experience in that organization (Allee 2012). Existing knowledge can be replaced by new alternatives through social collaboration and individual cognition (Nonaka 1994; Pentland 1995). This denotes that not only individual differences but also interactions are necessary to create organizational knowledge. In other words, an organization requires a compound of cooperation and competition called “coopetition” (Gast et al. 2019). Organizational conflict is a superficial output manifested through “coopetition.”

Organizational conflict emerges through heterogeneous perspectives during the decision-making process (Jehn and Mannix 2001). Previous studies have divided it into “task conflict” and “relational conflict.” (Jehn and Mannix 2001; Amason 1997; Hon 2007; Jehn 1995). Both of these can hamper organizational performance in the absence of appropriate management (Amason 1996; Hambrick et al. 1996). Organizational conflict is not just one of the major factors enhancing organizational performance (e.g., Deutsch 2000; Greenhalgh 1987; Pondy 1967; Robey et al. 1989); it also leads to the cognitive growth of group members (Ames and Murray 1982). That is why previous studies have suggested organizational conflict as an effective solution to groupthink (Janis 1972; Ellis et al. 2003; Fernandez 2005; Solomon 2006; Janis 1982; Solomon 2006). Furthermore, it has been revealed that productive conflict can promote (Solomon 2006; Sunstein 2005; Surowiecki 2004) and reinforce collective intelligence (Page 2007; Woolley et al. 2010). In the early days of the study into organizational conflict, the main line of argumentation emphasized the negative effects (Brett 1990; De Dreu and Weingart 2003; Jehn and Bendersky 2003).

Considering the possibility of a positive effect became part of the discussion later (Cronin and Weingart 2007). These studies have pointed out that the context of conflict is more important than the conflict itself, dividing organizational conflict into the afore-mentioned two types to explain the different effects. Task conflict can have positive effects on organizational performance (Kanter 1988)—it can provide the chance to reconsider contradictory alternatives as well as share new information (De Dreu 2008; De Dreu and Van de Vliert 1997; Xie and John 1995). Conversely, relational conflict induced through individual preference, favor, and value can raise emotional discord, which affects organizational or individual capabilities (Amason 1997; Simons and Peterson 2000). In most cases, relational conflict is known to have a detrimental effect on organizational performance. Therefore, in this study, the conceptual boundary is limited to task conflict.

The second factor is “reconsidering decision-making.” This provides a chance to reconsider existing solutions or create new ones, known as alternatives. Additional discussion on alternatives is perceived as an effective solution to attenuate groupthink (Chapman 2006; Janis 1982; Longley and Pruitt 1980; Park 2000). Park (2000) pointed out that a lack of alternatives makes groupthink the source of organizational failure. Turner and Pratkanis (1998a) suggested reconsideration of decisions to develop alternative knowledge, although the decisions have been considered as defective ones (Esser 1998).

In terms of collective intelligence, reconsideration of alternatives is an effective system to generate better solutions, because it can extend an organization’s information pool (Miranda and Saunders 1995). Filtering of knowledge is also a critical issue in the collective intelligence system. Reconsidering alternatives can provide quality knowledge to the organization through repeated verification and sharing of results among people (Reia et al. 2019). Another positive aspect is that reconsideration can improve individual capabilities. Accumulated reconsiderations of alternatives extend the individual cognitive boundary, which is called background knowledge (Massari et al. 2019), and abundant background knowledge contributes to organizational flexibility, creativity (Hällgren 2010), and constructive discussion (Ellis et al. 2003). Clearly, expansion of the individual knowledge domain can attenuate groupthink (Baron 2005; Flippen 1999). This study describes reconsideration as the tendency to keep existing knowledge and interact with more agents.

The last factor is organizational memory, including both storage and the reuse process. This factor is defined as the way organizations store knowledge for future use (Cyert and March 1963; Levitt and March 1988; Stein and Zwass 1995). In previous studies, grasping prior knowledge has been considered crucial for collective intelligence (Denning et al. 2005) because the crowd can make a good decision only if accumulated knowledge exists in its domain (Surowiecki 2004). In a decentralized organization without any integrated system, individual knowledge is stored by each member (Atlee 2003; Lee and Chang 2010). Despite it not costing much, over time, the location and content of knowledge become unclear or are even completely discarded within the organization. This loss can disrupt cooperation among heterogeneous members (Faraj et al. 2011; Kane et al. 2009). Consequently, the lack of organizational memory can block essential channels for generating new organizational knowledge (Maciuliene and Skarzauskiene 2016).

From the organizational viewpoint, this factor performs two roles in the knowledge creation process (Moorman and Miner 1997)—first, it functions as an interpreter by filtering the knowledge (Cohen and Levinthal 1990; Day 1994; Sinkula 1994; Walsh and Ungson 1991), and second, it creates a guideline by influencing the organization’s behavior (Cyert and March 1963; March and Simon 1958; Moorman and Miner 1997). It also plays an important role in the organizational learning process (Huber 1991). Therefore, an organization’s performance and outcomes can be determined by organizational memory and its learning-related behaviors (Antunes and Pinheiro 2019).

In summary, organizational memory not only supports the performance of collective intelligence but also enhances organizational performance and outcomes. That is why it is rational to consider the switching factor of groupthink. This study defines organizational memory as a process that includes acquisition, storage, and retrieval (Stein and Zwass 1995).

3 Effect of switching factor on quality of organizational knowledge

ABM is a powerful method to analyze large-scale and complex systems comprising autonomous agents (Bobashev et al. 2007; Miller and Page 2007; Cornforth et al. 2005; Xiao and Han 2016). The agents of our ABM were “stochastic” because they did not guarantee the recursion of events (North and Macal 2007). The model’s consistency was tested through a sensitivity analysis (Fig. 1).

Fig. 1
figure 1

Description of the agent-based model

Full size image

In this study, the ABM was designed as two layers—the agent layer (the first) and the environment layer as shown in Fig. 1. The agents in the first layer had four common characteristics: autonomy, interdependence, rule compliance, and adaptation (Macy and Willer 2002). They were capable of creating patterns through local and global interactions, which is a self-organization system (Kaufman 1996). This study developed the agent who meets these characteristics.

In the agent layer, arbitrary agent i attempts to make decisions to enhance the utility determined by the utility function (\({U}_{i}\)). We adopted a logarithmic utility function that is more effective than a quadratic form when expressing the behaviors of agents (Kraus and Litzenberger 1975). The utility function is composed of two variables: individual performance and organizational performance. Through the relative utility, agents recognize their own (Arentze et al. 2013). Thus, individual performance (\({p}_{ind}\)) is defined as a relative rank of performance among neighborhoods.

$${p}_{ind,i}=1-\frac{rank\;of\;{U}_{i}}{Number\;of\;neighborhood}$$

Self-centeredness (\({\alpha }_{i}\)) and sensitivity (K) were adopted to reflect the heterogeneous personality of agents. The utility function of each agent is defined as:

$${U}_{i}={K}_{i} {p}_{ind,i}^{\alpha }{ p}_{org}^{1-\alpha }$$

Since individual knowledge is created through fragmented information and its interpretation and combination (McHugh et al. 2016), knowledge can be defective and uncertain (Davis 1986). This study describes individual knowledge \(({knw}_{i}\)) as a distribution of unit information. Individual knowledge can be changed by two factors: first, interactions among the agents can change the shape of knowledge distribution, and second, time increases the uncertainty of knowledge. Here, the change in knowledge distribution is explained by a stochastic differential equation called the Langevin equation. It provides an insight for describing the change of probability distribution with respect to time. If we assume that the location of the agent is x, according to the equation, the changes in the location can be expressed by the sum of the deterministic and stochastic terms. Thus, the differential of x can be derived from a deterministic location (\(a\left(x,t\right)\)) and the drift term (\(b(x,t)\)) with noise (\(\xi\)).

$$\frac{dx}{{dt}} = a\,\left( {x,t} \right) + b\,\left( {x,t} \right)\xi \,\left( t \right),\,\xi \,\left( t \right)\sim {\text{random walk}}$$

Assuming the current location (x) is a random variable, the above equation is represented thus:

$$\left.{knw}_{i}^{t}=A\left({knw}_{j(i\ne j)}^{t-1}\right)+\xi (t)B({knw}_{i}^{t-1}\right)$$

In this equation, the drift term, \(A\left({knw}_{j(i\ne j)}^{t-1}\right)\), means that the expectation of location at t-1 and \(B({knw}_{i}^{t-1})\) refers to a stochastic turbulence term at t. These two terms change location and shape of knowledge distribution over time, as shown in Fig. 2. This equation means that knowledge distribution at t is calculated by deterministic information and stochastic noise (Fig. 2).

Fig. 2
figure 2

Visual description of the drift term and stochastic turbulence of knowledge distribution

Full size image

The second is the environment layer. Despite it not having the authority to intervene in the behavior of agents, it can affect the agent indirectly (Cha et al. 2019) through two essential pieces of information. First, organizational performance is an important signal. This information is calculated by comparing organizational knowledge with the predefined solution set called “fitness” in evolutionary computation (Levitt and March 1988; March 1991). Second, this layer defines the environmental conditions the agent belongs in. The initial condition is invariant if there is no external impact.

The learning process has been considered as one of the basic interactions to improve organizational performance (e.g., Levitt and March 1988; Posen et al. 2013). Despite imitating different knowledge, it is the most well-known method (Gilbert and Terna 2000). However, it is too simple to describe interactions between the distributions of individual knowledge. Thus, we represented the learning of knowledge distribution as a stochastic drift.

To reflect this drift, we assumed that the agent wants to get closer to the knowledge of another agent having higher utility. As mentioned above, the accumulated information of shift is in \(A\left({knw}_{j}^{t-1}\right)\). Thus, we can denote the stochastic drift as the expected location term (Fig. 3).

Fig. 3
figure 3

The learning process of individual knowledge distribution

Full size image

The last role of the environment layer provides the outcomes of the system. Since groupthink and collective intelligence are patterns rather than events (Turner and Pratkanis 1998c), it is hard to observe them on the intra-organizational level (Park 1990; Turner and Pratkanis 1998a, b, c). Therefore, the transformation of groupthink can be observed on the environment layer. To discover large-scale patterns, we observed the shape, bias, quality of organizational knowledge, and average utility.

Organizational knowledge (\({knw}_{org}^{t})\) is described as merged individual knowledge distribution with a total area of 1.0. The quality of knowledge (\({Q}_{t}\)) is calculated by comparing the organizational knowledge (\({knw}_{org}^{t})\) with the optimal knowledge (\({knw}_{opt})\) randomly generated from the distribution following N (0,10).

$${knw}_{org}^{t}=\frac{1}{N}\sum_{i=1}^{N}{knw}_{i}^{t}, {\text{N is the number of agent}}$$
$${knw}_{opt}\sim N(0, 10)$$
$${Q}_{t}=1/\int |{knw}_{opt}-{knw}_{org}^{t}|.$$

The ABM simulation in this study comprised five steps—“initialization,” “updating environment layer,” “interaction and decision-making of agents,” “updating agent layer,” and “termination.” Initialization was a stage to allocate parameters and construct an agent network. In the second step, we derived the organizational phenomena observed in the environment layer and updated the information accordingly. The third step was the main part of the ABM simulation; here, the agents demonstrated behaviors and complex interactions. The information of the agent layer was updated in the fourth step. Steps from the second to fourth were iterated until the system satisfied the termination conditions. When we entered the termination step, information from both the agent and environment layers was merged and sorted by time. All ABM simulations follow the same process explained above. Figure 4 shows the simulation carried out in this study.

Fig. 4
figure 4

Process of ABM simulation

Full size image

To conduct such simulations, the parameters should be defined. Table 3 shows the initial value of each parameter and the basis for value setting.

Table 3 Initial parameters of ABM simulation
Full size table

3.1 3.2. Reference model

To analyze the effect of switching factors, the reference model was tested; its output can provide the required criterion to compare the results of the other experiment models. In addition, since robustness of the ABM simulation is a critical issue, a sensitivity test was conducted (Pannell 1997). Saltelli et al. (2008) and Guus ten Broeke et al. (2016) suggested that the robustness of a model is verifiable by the variance of the output expressed as a scatter plot of output; this study followed that method. Figure 5 shows the results of the sensitivity analysis. The left figure provides the overall changes by time, and the right indicates the distribution of the final performance. These results suggest that the changes in assumed parameters, number of agents, and density of network do not change the overall result significantly. According to these results, our simulation models were consistent enough on the fluctuation of inputs.

Fig. 5
figure 5

Result of sensitivity test

Full size image

3.2 Knowledge optimization and knowledge bias

As mentioned above, this study employs three switching factors to resolve groupthink. The first is knowledge conflict—in this study, it is defined as combining knowledge of two agents, i and j, who have a high heterogeneous score \(({H}_{i,j})\). Two interacting agents are determined by the score of heterogeneity as follows:

$${H}_{i,j}=\int [{knw}_{i}\left(t\right)-{knw}_{j}\left(t\right)]$$

The second factor is reconsideration of alternatives, providing additional chances for exploring better solutions. To reflect reconsideration in the ABM simulation, each agent was assigned a probability that they would delay learning the other agents’ knowledge, allowing them to consider the existing knowledge before adopting another’s.

Not only existing but also obsolete knowledge can be exploited usefully during the organizational knowledge creation process. That is why organizations store their knowledge in an explicit form. The stored knowledge can be mutated in many ways when agents retrieve it based on their individual context, such as background knowledge, experience, and prejudice (Gammelgaard 2010; Nonaka 1994). Furthermore, the content of knowledge can be distorted if that knowledge has not been used for a long period of time. In this analysis, all created knowledge should accumulate in organizational memory, and the knowledge (\({k}_{i}^{t})\) created by agent i at time t was randomly mutated based on the temporal distance \((d=t-{t}_{0})\) from the time it was stored (\({t}_{0}\)).

Table 4 includes the results of knowledge optimization performance and its bias—knowledge optimization describes the differences between organizational knowledge and the optimal solution set in terms of the shape of knowledge. Although the best case is to choose the average of the set, we cannot know its exact shape. Therefore, organizational knowledge tries to stay near this average. Knowledge bias, on the other hand, refers to how individual knowledge is dispersed. The x-axis of the figure refers to the diversity of individual knowledge shape, and the y-axis to the overall diversity of the kind of individual knowledge. From these two aspects, we observed that all three switching factors may enhance the fitness to the optimal solution set because organizational knowledge converges closer to the set’s average.

Table 4 Effect of postponing decision making
Full size table

The first row of Table 4 shows the results of the reference model. Following results compared with this reference model. “Knowledge conflict” makes organizational knowledge dramatically converge with the optimal knowledge set. At the same time, it eliminates the variance of individual knowledge, and this means individuals have formed their knowledge conclusively.

“Reconsideration of alternatives” and “organizational memory” bring the better of the organizational knowledge than the reference model case. With respect to knowledge bias, “reconsideration of alternatives” has no major effect on the diversity of individual knowledge. An interesting result is that “organizational memory” forms an island-like area of individual knowledge over time. This result implies that individual knowledge may be separated into two groups.

Despite the status of knowledge optimization and bias briefly describing their influence, it was not certain if they could transform groupthink into collective intelligence. As mentioned previously, groupthink and collective intelligence cannot be identified until their final outcomes are produced. Thus, the quality of organizational knowledge and average utility of agents were calculated to compare the performance of the final outcomes.

3.3 Quality of knowledge and average utility

Through three simulation experiments, this study captures the effect of each switching factor on the quality of organizational knowledge and average utility of agents. Since the dominant difference between groupthink and collective intelligence manifests in the quality of their final outcomes (Hansen and Vaagen 2016; Täuscher 2017), better performance of both the organization and individual may guarantee collective intelligence rather than groupthink.

Figure 6 shows that knowledge conflict (experiment 1), as the switching factor, significantly increases both organizational performance and individual utility; it encourages repeated organizational conflict, but before there are sufficient interactions, it can lead to an inefficiency of collective intelligence. However, as the results of organizational conflicts accumulate, the interaction between disparate knowledge becomes more likely to produce better knowledge than an existing one. This is called “constructive conflict” of an organization (Ellis et al. 2003; Hall and Williams 1970; Maier and Hoffman 1964).

Fig. 6
figure 6

Organization performance and average utility from the experiments

Full size image

The performance and utility of the organizational memory model (experiment 3) were rapidly improved in the early stages. Since then, they have been lower than those of the reference model. Insufficient organizational memory makes it difficult for organizational decisions to fully benefit from valuable individual knowledge (Park et al. 2013). However, organizational memory can have advantages in a decoupled organization where knowledge sharing is rare (Wieck 1976) in other words, constant interactions among the members are likely to constrain its benefits (Tufool and Gerge 2013). That is why the effect of organizational memory on the performance and utility of organizations turns negative as the knowledge interaction repeats.

Reconsideration of existing knowledge has not shown a significant difference with the reference model. Evaluating alternatives has been easy for the individual agent if the number of alternatives is small—knowledge performance and average utility increased in the very early stages because the volume of alternatives was adequate for an individual agent to handle; however, repeated knowledge interactions rapidly increased the number of alternatives, ultimately causing confusion. With respect to this perspective, previous studies have argued that a strong leadership (Courtright 1978; Montanari 1986) or group cohesiveness (McCauley 1989) is required to evaluate a large number of alternatives (Breitsohl et al. 2015). In addition, even though considerable alternatives are available, individual feedback that is too far from the organizational goals leads to the wrong belief that all alternatives were fully evaluated (Flippen 1999).

From the three ABM simulation experiments, this section provides some evidence on the influence of one switching factor on the performance and average utility of an organization. However, the total effect of multiple factors is not the same as their sum in a complex adaptive system because interactions intervene in a system’s evolution (Kauffman 1996). Thus, combinations of switching factors were tested next through the MFA, which can compare the efficiency of heterogeneous strategies.

4 Finding the optimal strategy

4.1 Meta-frontier analysis

When measuring the maximum efficiency of industrial productivity, two kinds can be considered in terms of industry segmentation. Stochastic frontier analysis (SFA) is an important methodology to compare the intra-efficiency of a homogeneous group—it has only one productivity frontier estimated from the entities in the group; however, if an industry has multiple heterogeneous groups, comparison of the productivity frontiers of all groups is required. MFA is for comparing the productivity efficiency among the sub-groups—it compares the theoretical efficiency based on the production function of the industry to reflect the unique characteristics of heterogeneous sub-groups (Battese et al. 2004; Coelli et al. 2005). Through the MFA, the between-group efficiency is calculated by the distance between meta-frontier and group-frontier (O’Donnell et al. 2008).

This study utilized a virtual dataset generated by the ABM simulation to identify the efficiency of each combination of the switching factors. The ABM groups were divided into eight, each with a unique strategy. Details of the classification are given in Table 5.

Table 5 Description of sub-models
Full size table

Generally, the efficiency of each group is defined as a ratio between the input and output. This study assumed learning capability and diversity as inputs of organizations and the organizational knowledge performance as the output. Appropriate levels of learning capability (Levitt and March 1988; March 1991) and organizational diversity (Aggarwal and Woolley 2013; Kozhevnikov et al. 2014) are considered as important factors in determining the quality of organizational knowledge; excessive amounts of the same can hamper the quality (Levitt B; March 1988; Woolley et al. 2015). Thus, this study assumed a polynomial function to estimate the relationship between inputs and output.

$$perf_{i\left( j \right)} = \alpha_{i} + \beta_{1} div_{i} + \beta_{2} learn_{i} + \beta_{3} div_{i}^{2} + \beta_{4} learn_{i}^{2} + \beta_{5} div_{i} learn_{j} + \left( {V_{i} - U_{i} } \right),\,{\text{where }}V_{i} \sim N\left( {0,\sigma_{V}^{2} } \right),{ }U_{i} \sim \left| {N\left( {0,\sigma_{U}^{2} } \right)} \right|$$

In every ABM simulation experiment, the organization’s learning capability and diversity were randomly assigned, and its performance was measured by the quality of knowledge explained in the previous section. The stochastic frontiers of each group were estimated by FRONTIER 4.1 software and the meta-frontier for each by MATLAB R2017a.

4.2 Strategy comparison result

The efficiencies of the strategies used by each group are shown in Table 6. According to the estimation result, all in-group efficiencies (TE), that the average efficiency of each simulation result based on the same condition, were high, meaning the agents of each model were fully utilizing the input resources to enhance their organizational knowledge performance. This also indicates that it is impossible to increase the organization’s efficiency with individual efforts alone.

Table 6 Estimation results for the SFA and MFA
Full size table

In contrast, the between-group efficiency (TGR), which means the relative efficiency of a certain condition of simulation, toward the meta-frontier of each model showed a wider gap—group 2, which had knowledge conflict and reconsideration, showed the highest, while groups 3 and 5 showed higher compared to the others; the latter two also ranked first in total efficiency (TE*) that is the relative efficiency of each simultion based on the optimal conditions. Therefore, we can say that the strategy of group 2 was the best and those of groups 3 and 5 good enough to be considered as alternatives. In contrast, groups 4 and 7 showed remarkably low efficiency.

In sum, the strategy of group 2 may be the optimal combination of switching factors—combining knowledge conflict and reconsideration guarantees high efficiency in organizational knowledge creation. According to the results of group 7, reconsideration of alternatives is not an effective strategy when adopted alone. Similarly, knowledge conflict becomes a defective strategy when considered simultaneously with organizational memory.

The primary goal of this analysis was to identify the optimal strategy by comparing combinations of switching factors for efficiency. The MFA results also provided some evidence on the strategies organizations should avoid. Initiating a new strategy is hard and risky in real-world situations. Stopping the implemented strategy, on the other hand, is relatively easier from the organization’s perspective. Thus, knowing what should not be done is sometimes more beneficial than knowing what to do.

The MFA produced two groups with strategies that should not be adopted. First, reconsideration of alternatives can be effective only when a sufficient number of alternatives are available (Flippen 1999); its use on its own creates inefficiency in the organizational knowledge creation process. Second, the combination of knowledge conflict and organizational memory shows a much lower efficiency than the others. Previous studies have argued that verifying the quality of knowledge is an important issue in the collective intelligence system (Choi 2009) and that the validity of knowledge is strongly influenced by the evaluations of other people (Flanagin and Metzger 2000; Flanagin and Metzger. 2008). Thus, it can be inferred that the inefficiency of group 4 is due to the lack of an evaluation process or filtering of the accumulated knowledge through knowledge conflict and memory.

5 Discussion

In this study, ABM simulation was conducted to suggest that “switching factors” stimulate collective intelligence, and the MFA to arrive at the optimal strategy. To understand the creation of organizational knowledge and decision-making, our analyses present valuable lessons on how to use “switching factors” to transform groupthink into collective intelligence.

Despite contemporary organizations’ complexity and dynamic behaviors (Coff et al. 2006; Milosevic et al. 2018; Smith and Lewis 2011), resolving groupthink has only remained in Janis’ framework or its modified theories (Rajakumar 2019). To create quality organizational knowledge, previous studies have focused on how to nullify groupthink’s effects based on the linear causalities (Esser 1998; Rajakumar 2019; Turner and Pratkanis 1998c). However, such nullification alone in an organization is not the best solution because it is an effective way to handle simple or routinized problems at low cost (Janis 1982). In addition, knowledge bias, which is recognized as a source of defective decision-making, also can be a natural product of the organizational consensus process (Solomon 2006). Therefore, this study is interesting in that it demonstrates how to transform groupthink into collective intelligence.

Our findings indicate that an organization having groupthink can be moved closer to collective intelligence by the strategic use of switching factors. The ABM simulation and MFA illustrate two facets of the switching factors.

In the ABM simulation, the influence of each switching factor has been investigated—knowledge conflict clearly increases knowledge optimization performance but considerably biases the domain of individual knowledge at the same time. This finding supports the earlier idea that knowledge conflict can be constructive when the organization’s task is complex. Conflicts between heterogeneous knowledge incur substantial costs when the organization has problems such as inconsistent and uncertain goals or defective communication (Chiocchio et al. 2011); in such a situation, knowledge conflict is likely to work negatively. On the other hand, an organization facing a complex task requires sufficient knowledge conflicts (Jehn and Mannix 2001) to not only expand the domain of knowledge but also acquire new knowledge (Miranda and Saunders 1995). More specifically, as Jehn and Mannix (2001) explained, the need for knowledge conflict increases when the organization has multiple perspectives.

Reconsideration of alternative knowledge is a factor that has been emphasized, especially in research on the prevention of groupthink (Janis 1982; Rajakumar 2019; Riccobono et al. 2016; Turner and Pratkanis 1998c). The results of ABM simulation suggest that while it is not effective with respect to knowledge quality and individual utility, it does help preserve the diversity of individual knowledge. Previous studies have pointed out that maintaining organizational diversity contributes to either a reduction in groupthink (Fernandez 2005; Solomon 2006) or development in collective intelligence (Hwang et al. 2009; Loasby 2002; Surowiecki 2004).

Organizational memory has been highlighted in studies on collective intelligence as well as knowledge management. Our findings present a model wherein organizational memory creates an isolated knowledge area, which refers to organizational knowledge memory. For creating collective intelligence, each organizational knowledge should be stored as a specialized form (Malone and Bernstein 2015). In terms of knowledge quality and average utility, there is skepticism over the effectiveness of organizational memory. Over time, stored organizational knowledge loses efficacy and becomes an obstacle to change because individuals depend on retrieving knowledge rather than creating new (Starbuck and Hedberg 1977; Walsh and Ungson 1991). In the model with organizational memory, despite knowledge quality and utility being raised till midway, they decreased rapidly in the second-half of the experiment; in the end, this model presented the lowest level in both knowledge quality and average utility. This finding provides evidence on the negative side of the organizational memory system.

Each switching factor can be combined with the others. In this study, we define eight strategic groups based on such combinations. According to our findings, the strategy with knowledge conflict and reconsideration (group 2) has maximum efficiency. Knowledge conflict is based on heterogeneity, and thus, the agent who interacts with other agents has two options. The first is adhering to the existing knowledge. If all agents choose this option, knowledge conflict will be meaningless behavior. When the agents choose the second option and adapt their knowledge based on the heterogeneous ones, there remains a problem. They should determine how much they will learn from heterogeneous knowledge. Reconsideration of alternative knowledge provides some clues. The agents engaging in it can determine the level of learning from heterogeneous knowledge by comparing it with the alternatives they have.

The combination of knowledge conflict and memory has the lowest efficiency with the exception of the groupthink model (Group 1). As mentioned above, people can choose whether or not they interact with heterogeneous knowledge. Organizational memory provides a chance to interact with knowledge that no one currently has. As a result, these organizational behaviors amplify the knowledge of mutual learning with a lack of adequate consideration or evaluation.

According to the MFA results in Fig. 7, those of the best and worst strategic groups highlight the importance of “reconsideration.” However, a point to remember is that reconsideration is not an effective strategy when utilized alone; it is when adopted with other switching factors. Knowledge conflict focuses on knowledge diversity (Miranda and Saunders 1995), and organizational memory on the volume of organizational knowledge (Kruse 2003). Our findings show that the volume and diversity of knowledge cannot guarantee the quality of organizational knowledge created. To filter accumulated knowledge, reconsideration of existing alternatives is imperative. Hence, this study provides support to the idea that accumulated knowledge and its appropriate evaluation result in organizational diversity and decentralization, which are essential factors for collective intelligence (Solomon 2006).

Fig. 7
figure 7

Comparison of the strategy of each group

Full size image

6 Conclusion and limitations

When an organization aims to effectively solve enormously complex problems, internal organizational interactions are emphasized over individual capacity (Chiocchio et al. 2011). Organizational knowledge is created through various organizational behaviors (Alavi and Leidner 2001; Nonaka and Konno 1998), and knowledge bias occurs during this process. In the absence of appropriate management, such bias can lead to groupthink in organizations. Previous studies have suggested diverse ideas to prevent organizational failure and promote collective intelligence (Rajakumar 2019). In particular, the development of IT has made collective intelligence a prominent capacity of organizations (Alag 2008; Lykourentzou et al. 2011; Musser and O’reilly 2007). However, despite the technological foundation, there have been few discussions on strategic solutions for groupthink and collective intelligence.

This study provides several lessons in the knowledge management area. Reconsideration of alternatives is an essential process to fully exploit the existing organizational knowledge. In particular, the combination of knowledge conflict and reconsideration may be the best way to stimulate collective intelligence in a groupthink situation; conversely, simultaneous use of knowledge conflict and organizational memory should be avoided, for the same purpose. Thus, these results have a fundamental implication in terms of collective intelligence. Indeed, the switching factors in this study do not involve complex or risky behaviors, rather, routine and low-cost ones. Based on this, we can conclude that collective intelligence of an organization may be the product of several simple routines.

If ABM simulation is designed based on empirical data, it has an advantage in its accuracy. However, this study does not calibrate the ABM due to lack of real data relevant to the research topic. This is one of its limitations, as calibration based on real data is expected to enhance the reality of the simulation model. Also, the extension of parameters and functions is needed to make the ABM more elaborate.