Due to a phenomenon called crowd wisdom, the quality of a collective's decisions can be higher than those of the individuals composing it (Galton, 1907). Galton observed that aggregating evaluations across individuals gave a more accurate estimate than that of the median evaluation across participants. Under what conditions is the crowd wiser than the individuals who compose it? There is a large body of literature about improving crowd wisdom (Hertwig, 2012; Becker et al., 2017; Prelec et al., 2017). In general, the understood model of crowd wisdom suggests that it depends upon a large, unbiased group of independent judges. In this manner, averaging over the responses reduces the noise of an individual's response, therefore making the crowd decision more accurate. However, recent experiments reveal that this is not always the case (Lorenz et al., 2011): Exposing subjects to evaluations given by their peers can often improve collective estimates, by allowing people who are less confident in their estimates to change their mind (Jayles et al., 2017).

In more practical situations that require deliberation and collective decisions, social communication can either improve or undermine crowd wisdom depending on the structure of the social network (Centola, 2010; Becker et al., 2017). Consider, for example, two online communities. In one of them the social communication network is highly centralized, with some popular people serving as “communication hubs” in a small world network (Watts, 1999). In the other community, communications are more distributed and decentralized. Counterintuitively, Becker et al. (2017) found that social influences improved the wisdom of the crowd in the decentralized social network. In other words, pooling biased estimates may be superior to unbiased estimates if the biases themselves are broadly distributed. In contrast, they found that in social networks where there was a high degree of centrality, social influence has the opposite effect: it decreases crowd wisdom.

Note that the structure of a social network may have two distinct influences on crowd wisdom (Figure 2B): First, the structure of the social network may influence signaling quality (the accuracy of individual evaluations) via direct social influences (Jayles et al., 2017). Second, the structure of the social network may also determine how information is “filtered” while propagating through it, potentially affecting crowd wisdom via an implicit (and often obscured) process of data aggregation (Becker et al., 2017). For example, depending on network topology, minority opinions might be filtered out or amplified from a debate in different rates (Li et al., 2013).

In sum, crowd wisdom studies indicate that accuracy of collative estimates should improve with sample size and (depending on network topology) with a balanced social influence. Although crowd wisdom and cooperation are typically studied separately, in governance systems that are based on voluntary feedback, the two may interact: both sample size and topologies of social influences are outcomes of cooperation (i.e., of the motivation to participate). If improving crowd wisdom can affect the quality of collective decisions, then crowd wisdom might also, in turn, affect cooperation (Figure 1A).

Many social media platforms aim at maximizing communication efficiency, e.g., by nudging users to add connections and “friends.” However, it appears that efficiency and innovations of collective action depend on network structure in a complex manner (Mason et al., 2008). These researchers showed that expansive (wide-ranging) networks increase exploration, which is important in finding optimal solutions for complex problems, whereas highly connected (small world) networks may allow faster convergence but not necessarily on the optimal solution. Other studies further suggested that a focus on efficiency may engender a hidden cost with respect to crowd wisdom: One might imagine that informationally efficient collaboration networks should increase the ability of a group (say a task force) to find an innovative solution to a complex problem. However, networks that are efficient for gathering information are not necessarily efficient for performance (Kearns et al., 2006). For example, in an experiment, Brackbill and Centola (2020) gave groups of data scientists a task to find better solutions for complex statistical modeling problems. Participants were randomly assigned to either an efficient or an inefficient communication network. In both groups, subjects were exposed to the same “load” of information, the only difference was in the network efficiency of propagating information. Interestingly, groups in the efficient networks underperformed, while those assigned to inefficient communication networks reached highly efficient solutions. This result led Brackbill and Centola (2020) to suggest that there exists “a tradeoff between the network structures that promote a solution's rapid diffusion throughout a group and the network structures that promote the discovery of innovative solutions.” These results call for translational experiments with governance logic—e.g., regarding the efficiency of communication networks between committees and boards.

Early studies outlined central government as a necessary tool for sustaining public goods in order to prevent a “tragedy of the commons” (Hardin, 1968). Three problematic behaviors, in particular, have been studied (Figure 2C): one is free ridership, where people take advantage of public goods but do not contribute enough resources to sustain them (Marwell and Ames, 1979). The second is the “prisoner's dilemma” defection strategy, where a greedy (and myopic) tendency of actors to maximize their immediate gains may erode cooperation over time. And the third is represented by “second-order social dilemmas,” in which monitoring, punishment, and other governance activities for preventing free-riders and defection themselves become vulnerable to free-riding or defection (Okada, 2008), because they are also costly and generate positive externalities. Later group and network studies, pioneered by Elinor Ostrom (see e.g., Janssen et al., 2010) and others (Ahn et al., 2009) suggested that subjects can and do creatively evade free-riding and defection in real-world situations that resemble second-order social dilemmas. Ostrom found that people can overcome these three types of behavior to organize locally and make sustainable arrangements for self-governing common pool resources, without any need for external government interventions. People are often willing to engage in altruistic (and costly) punishment of “free riders” (Ostrom, 1990). Consequently, over time, sanctioning institutions tend to be more competitive than institutions that do not punish free riders and defectors (Gürerk et al., 2006). In the same vein, several studies showed that, in a repeated game, reputation can play a major role in sustaining reciprocity (Axelrod and Hamilton, 1981).

Recently, however, massive online experiments showed that cooperation can persist via network level mechanisms even in the absence of punishment or reputational effects (Suri and Watts, 2011; Mao et al., 2017). Mao et al. (2017) performed an online experiment to study the long-term dynamics of cooperation. Participants played a prisoner's dilemma game repeatedly—that is, hundreds of times over several weeks. Although the Nash equilibrium corresponds to zero cooperation in the absence of mechanisms for punishing defectors, about 40% of participants were irrationally resilient cooperators: they would not be the first to defect, despite knowing that they were about to lose money. Interestingly, long-term social learning in the remaining (more rational) players slowly promoted an equilibrium of cooperation, with dynamics unfolding over long time-scales that could have never been revealed in a typical lab study. Further, willingness to directly reciprocate cooperative behavior has been shown to persist even in highly competitive settings, with NBA players continuing to reciprocate assists despite strong individual incentives not to Willer et al. (2012). Using a similar approach, Melamed et al. (2018) showed further that even in the complete absence of reputational memory, simply allowing network dynamics to evolve gives rise to clustering of participants that shield cooperators from defectors. Further, resilience of cooperation was also observed in a more complex online experiment: Melamed et al. (2020) showed that different modes of reciprocity are fairly independent from each other, such that the collapse of one form (such as direct reciprocity) does not necessarily affect the others (indirect and generalized reciprocity).

Although discoveries such as those presented above seem relevant for designing efficient governance systems, it may be challenging to translate them into the specifics of institutional logic. Based on the studies discussed above we propose two practical approaches for improving crowd wisdom and cooperation.

First is the introduction of social influences into crowdsourced feedback and petition systems: The statistical gold standard for designing crowdsourced information systems has been to obtain independent and unbiased evaluations, in order to minimize sampling endogeneity (Heckman, 1979). But as noted above, in governance systems that rely on voluntary cooperation (Figure 1), outcomes may depend on dynamic interactions between cooperation and crowd wisdom. In other words, minimizing social influences may bear the cost of compromising crowd wisdom. More importantly, minimizing social influences may interfere with efforts to perpetuate cooperation (Figure 2A). Sacrificing the independence of evaluations in order to promote a virtuous feedback loop between crowd wisdom and cooperation can therefore make sense. For example, in a field study some of us reported a positive outcome of increasing social influences in a rating system (Tchernichovski et al., 2017). We found that exposing service clients to trends in rating scores—just prior to rating—was associated with a persistent improvement in satisfaction with services over time. It was also associated with the community sustaining a high feedback rate over years. In addition, introducing social influences may promote self-organization of social information. For example, in a petition system, presenting users with similar petitions while filling out the petition form, may prompt users to amend and endorse existing petitions instead of creating redundant new petitions. Finally, as suggested by the Becker et al. (2017) study, setting a decentralized process for deliberating petitions could be advantageous.

Note, however, that the injecting of social influences into early stages of crowdsourced information systems may open gateways for groups of activists with bad intentions to manipulate and distort information. There are well-established mechanisms for online communities to deal with individual bad actors, but much more challenging are disruptions by collective actions of online “mobs” (Trice and Potts, 2018).

Second is the methods of data aggregation and presentation: Typical governance challenges are very different than those presented in most crowd wisdom experiments. In such experiments people are typically asked to make a quantitative estimate where there is a known ground truth, which is rarely the case in governance. However, crowd wisdom might be more relevant to governance at the low level of aggregating feedback from the community. Online community members often share ratings and “likes” to posts and services, generating a constant stream of quantitative data. It is rarely possible to optimize crowd wisdom in aggregates of such subjective evaluations with respect to a ground truth (which is often intractable). However, it is often possible and practical to improve crowd wisdom in order to promote the early detection of trends (Tchernichovski et al., 2017). That is, crowd wisdom can be defined as the speed and accuracy of detecting a change (Tchernichovski et al., 2019). The early detection of trends can be particularly useful in online communities, where social feedback is continuous. Say, for example, that an online community changed a policy regarding publishing posts in their archive. Detecting negative user feedback early could allow a speedy correction before too many users have “voted” by exit (Schneider, 2019). In the next section we will introduce experimental methods of optimizing crowd wisdom in feedback systems, including challenges in sustaining high quality of signaling (Figure 2B), which often suffers from sampling endogeneity and from poor information quality (Stocker, 2006; Moe and Schweidel, 2011; Ho et al., 2014).

In sum, improving crowd wisdom and cooperation should be considered while designing online governance. We highlighted two aspects where this might be particularly relevant: one is in the design of social influences during submission of petitions or evaluations, and the second is in the design of data aggregation in evaluations and ratings. We will return to these issues below while presenting frameworks for virtual worlds experiments with governance.

Observational studies on the governance systems of online communities revealed important dynamic relations between evolving governance structure and outcomes. For instance, Tan and Zargham (2021) recently published the GovBase database: a crowdsourced summary of tools used for online governance. The database allows for exploration of how different governance structures that have been adopted by online communities may correlate with (and perhaps predict) the survival of those communities over time. Several studies explored the evolution of governance in online communities. For example, Frey and Sumner (2019) studied governance rules in 5,000 online communities of video game players and described how they evolve over time. They found that the structure (number and scope) of governance rules given population size can, to a certain extent, predict growth of the group of core members. Because users can be enculturated to prefer a new form of governance, and conversely, can exert selective pressure on expressed governance forms by opting out of communities they don't like, it is possible for social feedback loops to drive the evolution of both governance forms and preferences. To this end, a follow-up study by Zhong and Frey (2020) finds evidence that, although influence is evident in both potential directions, the selective pressure mechanism is much stronger than the cultural evolution of preferences. Observational studies have illuminated several questions about the formation of self-directed governance systems. For example, how strong a force is emergent centralization? Shaw and Hill (2014) documented the unexpected emergence of oligarchy in the radically egalitarian domain of “wiki” knowledge bases. They observe that a small administrator class becomes increasingly distinguishable as wikis grow, and that the goals and standards of that class diverge from those of regular volunteers. How do different modalities of positive and negative interactions between agents aggregate to produce social outcomes? Szell et al. (2010) analyze the multidimensional relationships between hundreds of thousands of players in an online game to illuminate the game's emergent system of power dynamics. With a multiplex network approach, they demonstrate alliance and conflict dynamics playing out in a coordinated manner over several types of game action, including trade, conflict, and friendship relations. There are, however, strong limitations to such observational studies: causality is difficult to assert, and the space of scenarios that can be explored is constrained by the current state of the world. With an experimental approach to the governance of online communities, one can test cause and effect directly and explore governance outcomes in social scenarios that cannot be easily observed otherwise.

The challenge we would like to focus on hereafter is how to better connect the basic research of optimizing collective actions with observations of online governance (section Correlational studies in online communities) in order to test practical solutions. This would require translational experiments. Experimenting directly with social media governance may raise ethical concerns, but when done properly it can be highly valuable (Lazer et al., 2020). Concerns about early experiments and ethically questionable data-mining have strongly constrained subsequent experiments (Kramer et al., 2014; Lazer et al., 2020). That said, among the most promising recent approaches is the Matias CivilServant System (Matias and Mou, 2018) for community-led experiments in platform governance. An example implementation of CivilServant is a recent experiment that was done in a Reddit community with 13 million subscribers. An announcement about a new community rule was randomly assigned to some of the new members in an online science discussion forum (Matias, 2019). Presenting the announcement informed new participants about the community's rules: no jokes, no abusive content, and so on. The question was if this message would influence newcomers' choices to contribute content, and if it would affect harassment levels. Results showed that the intervention reduced cases of harassment without decreasing participation in the forum. We hope that such an experimental approach would become increasingly useful in optimizing cooperation, hence promoting a virtuous loop between cooperation and high-quality collective decisions, as discussed earlier (Figure 2). One way of achieving this is by complementing community-led experiments with virtual world experiments, where risks and attendant ethical concerns can be minimized. In addition, virtual world experiments provide the possibility to perform much more substantial control experiments and test more radical modification to the services including the possibility of “collapse” of the community, something that should be avoided in a community-led experiments.

As discussed above, online experiments in crowd wisdom and cooperation have contributed to a deeper understanding of collective behavior in humans. The key to this progress is in technological innovations that now allow for the testing of thousands of participants over hundreds of iterations while controlling the topologies of social communication networks. But given the highly stylized nature of experimental exchange systems, even such large-scale experiments have limited bearing on complex, real-life governance. We now turn to the challenge of conducting online experiments in virtual worlds, with the aim of studying governance in simulated, complex environments. There are two motivations to conduct such experiments: First, the governance systems of the virtual world can permit adventurous experiments in the varieties of social life with low enough stakes that they do not risk any meaningful form of social collapse. Second, placing people in situations that allow them to tweak or elaborate their setting makes it possible to study the evolution of complex institutions from simple ones. This means that we need not only to conduct long-term experiments, but also to place people in situations that mirror the real-world complexities of pooling and sharing multiple resources. Former studies utilize virtual worlds for conducting simple experimental interventions. For example, introducing bots into popular areas in the multiplayer game Second Life allowed researchers to examine how the “age” and “gender” of bots' avatars affected the avatar's chances of obtaining help from other players (Zhang et al., 2020b). They can also provide and validate governance tools for virtual worlds, such as research providing governance tools on the integrated game chat platform Discord (Zhang et al., 2020a), or community monitoring tools on the video game Minecraft (Müller et al., 2015). However, to our knowledge, this article is the first to present a framework for virtual world experiments that are designed specifically for studying governance.

We developed a method for experimenting with governance in a virtual world, which we call Ferry Services Game (FSG). FSG is an in-browser WebGL game (see Material and equipment). Participants (e.g., MTurk workers) manipulate their avatar in a 3D world to collect coins, which are redeemed for money at the end of the game (as a compensation bonus). The 3D world is composed of a long chain of islands (Figure 3A). Participants must use simulated ferry services to visit each island, but some ferries are fast and others are slow, and players are motivated to have faster ferries since time spent on ferries competes with time spent earning revenue (Figures 3B,D). Through several mechanisms, players may gain and contribute information about ferry characteristics. After a ferry ride the player may be prompted to rate the service. Ratings may be pooled and shared with other players via a dashboard. The experimenter sets the distributions of service speeds and delays, and then evaluates the ferry rating scores against the ground truth of that distribution. Therefore, as in many real-world situations, rating information can help players select (or collectively own) better services.

Several game parameters are adjustable: Rating the ferry may be either voluntary or obligatory. Players may be given no incentive to rate the ferry (which is often the case in real world rating systems). Alternatively, rating the ferry can be made a social process. For example, in Figure 3C we present a design where players can choose between two ferry services in each island. Here, sharing rating scores with other players would benefit all players by allowing them to quickly pool their knowledge about which ferry service is better. However, this public goods game incentivizes free ridership (Marwell and Ames, 1979). In a different design, only members of a “club” may share rating information, as in a common pool resource game. Here, club governance rules may be either randomly assigned, or set by the players. In this manner, layers of governance can be superimposed on the game, as we will show in the next section. In this section, we focus on low-level experimentation with optimizing social feedback.

Many online communities use feedback systems that provide quantitative evaluations such as “likes,” and rating scores. Feedback systems have evolved rapidly over the last two decades. They turned from an industry standard mechanism designed for obtaining information from—and retaining—clients, into a rich ecosystem of independent crowdsourced rating platforms, which now guide many everyday decisions.

FSG experiments can be designed for optimizing feedback systems at three levels. First is the optimization of the rating device: Most systems use rating devices that implement Likert scales, e.g., a 1–5 “star” ratings, where the rating device is a trivial “click and submit” radio group. Pooled star ratings can then be made visible to both creators and consumers of content, or in e-commerce, to service providers and potential clients (Figure 4A). There are many concerns about the quality (and honesty) of such rating scores (Luca, 2016). A recent FSG experiment found that even without any conflicts of interest or incentives to cheat, pooled rating scores could explain only about 14% of the variance in the speed of ferries (Tchernichovski et al., 2019). However, pooled rating accuracy was about twice as high when ratings were submitted via a device that imposed time costs of a few seconds on reporting extreme scores. Such an improvement in feedback information quality could be useful in the case of the NGO providing online services presented earlier (Figure 1A): With better rating accuracy the evaluation committee should be able to detect a change in satisfaction with its service outcome much faster. This could have two practical advantages: First, services can respond to the change faster. Second, the sooner a change can be detected the easier it is to identify its cause. In other words, reducing latency in detecting a change is likely to improve reinforcement learning in a multi-agent scenario (Sutton and Barto, 2018).

These results demonstrate the utility of using FSG for testing different designs of rating devices. The FSG game template we shared (https://github.com/oferon/FerryGame) can be used for testing such designs. FSG is coded in the gaming platform Unity, where building arbitrarily complex 3D rating devices (game objects) with physics can be easily implemented. The game can be then compiled as is into either a desktop application for lab experiments, or into a WebGL for online experiments via a browser, or into a phone app, for long-term experiments.

A second manner in which FSG experiments can be used for improving feedback systems is via experimentation with distributed feedback monitoring methods. Performance measurement systems with dashboards are widely used in many industries (Bititci et al., 2000) in order to monitor and coordinate performance benchmarks and optimize institutional learning. In principle, similar dashboard systems can be developed to optimize social learning in a non-commercial setting of online communities. Such dashboards may be particularly useful for online communities that offer services that are used repeatedly by a pool of members. In such cases, it might make sense to present dashboards with trends (Tchernichovski et al., 2017) rather than mean rating scores. FSG experiments can be designed for testing the utility of presenting such trends (Figure 4B). FSG experiments can test, for example, if the presentation of trends in rating scores of ferry services improves social learning, prompting players to adjust their strategies more efficiently. Note that even in a small community that runs a simple operation, social learning may show complex temporal dynamics: the detection of a trend gives both clients and service providers an opportunity for social learning in real time: service providers can “experimentally” adjust their behavior/product in real time in response to trends, while consumers can make choices with more up-to-date information. The challenge is how to tune the parameters of the trend presentation such that each side can learn most effectively from the other? Returning to the case of an NGO providing online services (Figure 1A): should trends in satisfaction with each service type be presented over days, weeks, or months? Presenting short term trends should minimize delay and improve social learning. However, presenting long term trends may promote stability and improve the clarity of social signals. For the NGO, manipulating such parameters can be risky, but FSG experiments can be designed for testing such dashboard calibrations while keeping the temporal dynamics of service provision fixed. For example, one may assign participants into ferry rider and ferry driver groups, and allow them to learn from each other only via a (service provider) dashboard presenting trends. Simulating the dynamics of social learning via feedback in different conditions could then guide the NGO on how to design and tune performance measurement systems in their non-market setting (a monosponistic environment).

Third, FSG experiments can be designed for testing the utility and risks of implementing machine learning for continuous optimization of rating system features. Such features may include, for example, the physics of a rating device (Tchernichovski et al., 2019), or the temporal resolution of a dashboard presenting trends (Tchernichovski et al., 2017), or both. About a decade ago, much of the software industry adopted continuous deployment, such that software is continually released and experimented with (Kevic et al., 2017). Such commercial systems for automated experimentation with platform design are often based on a closed-loop feedback system (Figure 4C). For example, the “style” of a banner in a web page includes many features, such as screen coordinates, size, colors, and fonts, which can be manipulated while monitoring changes in client behavior. Here, the typical feedback is not rating scores, but changes in traffic, clicking on ads, and so on. Using standard machine learning approaches (Mattos et al., 2017; Gauci et al., 2018), such systems can continuously “nudge” the design, with the aim of maximizing outcomes desired by the platform owner. Matias and Mou (2018) suggested that a similar testing approach might be useful in the implementation of online governance policies (Matias, 2019).

One may treat a subset of governance logic in a community as a set of features that can be continuously optimized as in the typical case of an ad banner. This could include gain and delay parameters (how often a certain committee should meet), or setting the threshold of consensus required for votes to enable a petition to pass, etc. In this spirit, we imagine FSG experiments combining continuous rating feedback (which can be seen as continuous voting) with continuous exploration of governance logic (Figure 4C). We emphasize that trying such an approach in real communities could be dangerous and possibly unethical, and that virtual worlds experiments should be regarded as sandboxes where thresholds for tipping points can be safely established. An example of such an FSG experiment would be implementing a machine learning algorithm to find the temporal resolution of a dashboard that maximizes both service usage and satisfaction with service outcomes over time. There are several risks in running such experiments in an online community. For example, people may (wrongly) perceive that information is being manipulated by bad actors, or the algorithm may become unstable, reducing public confidence.

Finally, an experimenter could allow feedback from dashboards to directly guide marginal investments in simulated services via continuous experimentation. The role of such FSG experiments could be to serve as a playground for exploring utility, but even more so, for discovering and then reducing risks prior to deployment in real-world online communities.