Collective Intelligence during Emergency Egress: The Mechanisms Underlying Altruistic Information Exchange

Abstract Understanding the human factors governing effective information exchange is increasingly indispensable for the design of day to day human-computer systems. Moreover, effective information exchange becomes a matter of life or death during emergency egress. The complexity of an unknown environment and the unpredictable locations of hazards often prevent evacuees from identifying safe routes. Successful evacuations from locations impacted by fire or earthquakes may depend on user-generated information to increase the chance of collective survival. The present paper employed multi-user virtual reality experiments and an online survey to investigate the mechanisms underlying social influence and collective intelligence during emergencies. Our results demonstrate that information sharing helps to reduce evacuation time and trajectory length. Participants also shared more when given incentives or when there was a lack of knowledge in the public information pool. This work provides further indications of how collective intelligence can be promoted and deployed during emergencies.


Introduction
With the rapid development of information technology, people increasingly volunteer to share information and data with each other over the internet. This shared information often assists activities in daily life, such as online encyclopedias (Cho et al., 2010) and social media (Ammari & Schoenebeck, 2015;Bi et al., 2019;. These modes of content sharing can facilitate archival or synchronization tasks (Sleeper et al., 2016), such as the sharing of geographic information on collaborative maps (Fechner et al., 2015;Haklay & Weber, 2008). Recently, the power of information sharing is arising for emergency scenarios. Indeed, social media (Neubaum et al., 2014) and other online platforms (Zhou et al., 2018) demonstrate their potential for improving the safety of users during emergencies. The phenomenon that a large group of individuals share their knowledge during an emergency event, can be defined as one form of Collective Intelligence, implying that the emerged collective information from all the members can provide additional help for the whole group. In an evacuation scenario, this type of collective intelligence can manifest as an increase in the map information shared among evacuees when there is more of a need for such information. Because of increasingly frequent and unforeseen disasters (e.g., wildfires in Australia and California), studies of collective intelligence during emergencies (Nayebi et al., 2017;Proulx et al., 1995;Vahidnia et al., 2020;Vieweg et al., 2010;Zook et al., 2010) will be key for supporting real evacuations.
Emerging technologies, such as virtual reality (Moussaïd et al., 2016) and human digitalization (Ludwig et al., 2015;Saffo et al., 2020;Weise et al., 2011), may be extended to study the collective intelligence underlying social influence. Dangerous and stressful scenarios can be simulated and presented to research participants without actually threatening their physical or mental health. Previous research has successfully used virtual environments to investigate safety education (Smith & Ericson, 2009), emergency evacuation (Kinateder et al., 2014), and hazardous working environments (Bliss et al., 1997;Stocker et al., 2011). The utility of multi-user virtual reality for studying emergency scenarios can be further improved by implementing features that allow for spatial information sharing and other social dynamics. However, the mechanisms underlying when and why users are willing to share spatial information during a simulated emergency must be assessed to eventually develop a useful application.
The present paper focuses on the social dynamics underlying altruistic behavior as exhibited by multi-user experiments on spatial collective intelligence during simulated evacuations. We also employed an online survey that studied emergency information sharing from a more general perspective. Based on previous studies of a well-established networked virtual reality setup (Moussaïd et al., 2016;Thrash et al., 2015;Zhao et al., 2018, we provide some of the first indications of how spatial collective intelligence can be elicited via a humancomputer system in a multi-user context. During two virtual reality experiments, participants completed a series of evacuation tasks in which they were asked to locate an exit and evacuate a maze with the help of an information sharing application. In the first study, the participants were given different incentives across trials. In some trials, participants were incentivized for their individual performance (i.e., individual goal), and on other trials, they were incentivized for the overall group's performance (i.e., public goal). In the second study, we replayed the sharing activities and crowd movement from the two conditions of the first study individually to a new set of participants. The independent variable for this second study was the amount of map information provided to individual participants (i.e., high versus low information content). With such a design, we want to answer the following research questions during emergency egress: When incentivized to help the group, do people tend to share more information with the others? How does the group's information-sharing behavior influence individuals' tendency to share? How does information-sharing behavior influence people's evacuation performance?
In three studies, we investigated the "if" and the "why" people would share information during emergencies. In Study 1, the "if" was studied with the independent variable incentive type (public or individual). Study 2 was designed to isolate the "why" by systematically relating participants' sharing behavior to the amount of information provided via sharing. Using an online survey, Study 3 provides further independent evidence regarding "why" participants are more likely to share in scenarios with low information content such as in Study 2. As expected, the results of these experiments revealed that sharing information occurs significantly more often when group performance is incentivized than when individual performance is incentivized. In contrast to our expectations, participants share more when there was less information in the public knowledge pool in the absence of a specific incentive. Study 3 broadens our understanding of these findings by elucidating the motivations and concerns of people related to information sharing. These findings have suggested that applications should be designed to promote information sharing via systematic notifications regarding public benefits that may improve collective evacuation performance. Despite the urgency and danger of such disasters, social media users have demonstrated the possibility of sharing first-hand information during wildfires and earthquakes. The combination of widespread mobile internet and frequent (but unpredictable) disasters highlights the necessity of a crowd-sourcing approach. These results related to social dynamics during emergencies can help policymakers and event organizers to better understand cooperative and altruistic information exchange and design effective humancomputer mediated systems that enhance egress.

Collective intelligence
Collective intelligence refers to the capacity of a group to share intelligence and usually arises from collaboration within a team or competition between individuals. Collective intelligence has been found to be correlated with a group's performance on complex computer tasks (De Liddo et al., 2012;Engel et al., 2015). For example, Malone and colleagues (Malone et al., 2009) found that group decisions were as useful as individual decisions in product development, especially when the knowledge necessary for making a decision was widely distributed among individuals in a group. In addition, Woolley and colleagues (Woolley et al., 2010) found evidence of the existence of collective intelligence within a group using experiments with video games and architectural design. Practitioners have suggested that collective intelligence is also beneficial for applied disciplines such as architecture (Hight & Perry, 2006) and evacuation research (Drury et al., 2009). In the context of evacuation, Moussaid and colleagues (Moussaïd et al., 2016) studied the collective dynamics of herding under high density scenarios in a virtual evacuation task. Similarly, Dodds and Ruddle (2009) found that collaboration among participants leads to significantly more efficient and shorter travel distances during an urban planning task in a collaborative virtual environment. Starbird and Palen have also investigated the collective behavior of self-organizing volunteers during a crisis (Starbird & Palen, 2011). Indeed, the use of social media during a crisis can not only enhance the sharing of information of the disaster (Alam et al., 2018;Anderson et al., 2018;Y. Chen et al., 2018;, but also enabled collective intelligence on the internet within digital volunteer communities (Starbird, 2011(Starbird, , 2013. The investigation of collective behavior during emergencies can also help researchers in human-computer interaction to better design crowd sourcing platforms (Dailey & Starbird, 2014).

Altruistic behaviors and social influence
Altruistic behaviors are the deliberate actions that people take despite knowing that these actions cannot maximize their monetary reward (Charness & Haruvy, 1999). Our understanding of altruistic behavior is primarily derived from observational studies in domains such as psychology and social economics. For example, researchers have observed that people are more likely to behave altruistically when their altruistic behavior is public (Burnham, 2003). Indeed, Andreoni and Rao (2011) suggested that the interactions among participants can be the key to stimulating greater altruism through the power of empathy and external communication. Social influence, which operates through social norms, conformity, and compliance (Cialdini & Trost, 1998), is one of the principles on which relationships are built with others. Similar altruistic behavior of donating (Kang et al., 2014) and sharing knowledge (Ducheneaut & Moore, 2004) has emerged and been studied in massively multiplayer online games.
Another mechanism of establishing relationships is reciprocation. Reciprocation is when people return favors to others for what they had previously received (Cialdini & Goldstein, 2004). For example, in commercial activities (Bodur & Grohmann, 2005) and sharing economics (Habibi et al., 2017), reciprocation can influence decision-making and resource allocation (Burnett Heyes et al., 2015). Reciprocation also requires mutual evaluation and judgment, from which trust and distrust can propagate (Guha et al., 2004). Judgment propagation involves adopting a judgment from a sender before spreading it to others , although such propagation can sometimes diminish collective intelligence (Lorenz et al., 2011). Judgment propagation is also reduced with distance from the original source. For example, Moussaid and colleagues  found that evidence for judgment propagation was limited to a social distance of three to four degrees of separation. The extent of judgment propagation can also be limited by the variety of information sources (Centola, 2010), the social structure between senders and recipients (Aral & Walker, 2014), and participants' subjective perceptions of the information being judged (Moussaïd et al., 2015).
Based on theories of social influence, some human-computer systems have been deployed in the real world to facilitate altruistic behavior (Wang et al., 2016;Yang & Olson, 2002). Recently, social media and personal smart devices (Reilly et al., 2009) have been used to enable helping behavior during hazardous events (Neubaum et al., 2014). In addition, crowdsourced mapping services provide maps that contain location information that is generated and shared by users voluntarily (Ding et al., 2014). Shared information includes information for navigation-assistance (Hyla et al., 2015), route choices (W. Wu et al., 2006), real-time traffic conditions (Fujihara, 2018), and hazard information (Chauhan & Hughes, 2017;Hirata et al., 2018;Michael et al., 2014). Such information can be shared offline (Bae & Montello, 2019;He et al., 2015;Helbing et al., 2015) or online via internet networks (X. Wu & Kuang, 2021;Xu et al., 2012). Warnings that can be shared by users increase the safety and navigational capabilities of these systems (Hyla et al., 2015). In addition, facilitated route planning reduces cognitive load while driving (W. Wu et al., 2006). The benefits of shared information can also extend beyond the interests of individual users. For example, Philipp and colleagues (Philipp et al., 2014) used crowdsourced trajectories from pedestrians to automate the generation of indoor maps of buildings. Similar approaches have generated indoor structures using crowdsourced, sensor-rich videos (S. Chen et al., 2015). The routes shared by travelers can also facilitate the updating of map information, which is often challenging because of frequent changes in road networks (Stanojevic et al., 2018). Despite these benefits of a common information pool, contributions to such pools require individuals to help others (Hirata et al., 2018), which can be cumbersome and time-consuming.

Behavior studies in virtual environments
Most studies on altruistic behavior employ either questionnaires (McCormack & Joseph, 2012;Schwartz et al., 2003) or behavioral games (Eckel & Grossman, 1996;G€ achter, 2004;Heli€ ovaara et al., 2013;Kwak et al., 2020). One potential disadvantage of these methods is that they lack a context with which participants can be engaged during the interaction. Lack of engagement may be problematic for inherently spatial contexts such as emergency egress. Virtual reality may provide the engaging spatial context necessary to study the social factors underlying altruistic behavior during emergency egress, especially with respect to human-computer systems.
Virtual environments can provide a realistic spatial context in which researchers can conduct systematic evaluations of participants' spatial (Khan & Rahman, 2018;Zhao et al., 2022) and social behaviors (Czarnek et al., 2020) without violating safety or ethical rules (Moussaïd et al., 2016). For example, Gillath and colleagues (Gillath et al., 2008) found similarities between people's reactions to a virtual person in need and participants' reactions in a real environment. The costly altruistic behaviors observed in the virtual environment were found to be correlated with greater empathetic concern as reported in a questionnaire (Patil et al., 2018). Virtual environments are also convenient for research on multi-user phenomena such as those that occur during hazardous events . Multi-user experiments provide researchers with an overview of the interactions of a group of participants (Moussaïd et al., 2018), but often require rapid changes to the environment between trials (Thrash et al., 2015). Predefined protocols  and experimental frameworks (Gr€ ubel et al., 2016) support the implementation of these studies. Research on spatial and social behavior can benefit from existing frameworks to conduct experiments that require interaction and collaboration in virtual environments (Kallioniemi et al., 2013). For example, Moussaid and colleagues (Moussaïd et al., 2016) used a networked multi-user experiment to study herding behaviors among human participants during a stressful evacuation. Similarly, Zhao and colleagues  extended this experimental framework to study the effects of map complexity and crowd movement on group navigation.

Study designs
The present paper investigates the mechanisms underlying collective intelligence and altruistic information exchange using two experiments in virtual reality and a comprehensive online questionnaire. In Study 1, participants attempted to find the exit of a multi-user virtual environment with the option of sharing spatial information regarding the location of blocked passages and the location of the exit. In some trials, participants received an bonus to perform selfishly, and on other trials, participants received an bonus to behave cooperatively. To be specific, the independent variable Incentives was defined and used here (Individual -can receive a bonus for evacuating within the time by oneself; Public-can receive a bonus if everyone in the group can evacuate within the time). In Study 2, participants performed a similar task individually without a specified incentive among agents that replayed the trajectories and map sharing behavior from different conditions of Study 1. The map sharing behavior were used as independent variable and defined as Map information (High-more information of the current fire locations displayed in the map; Low-less information of the current fire locations displayed in the map). In Study 3, we deployed an online questionnaire to broaden our understanding of people's motivations and concerns regarding spatial information sharing during emergency scenarios. The experimental design of the three studies is summarized in Table 1. With this design, we seek to examine the following hypotheses: H1. When given clear and specific incentives, participants share more information in trials with a public goal than participants in trials with an individual goal. H2. When given clear and specific incentives, participants in trials with a public goal perform better than participants in trials with an individual goal. H3. Without a specific incentive, participants share more information in trials with high information content than participants in trials with low information content. H4. Without a specific incentive, participants in trials with high information content perform better than participants in trials with low information content.

Study 1-Cooperation during evacuation
We designed the first study to examine the effects of different incentives on cooperative behavior in multi-user virtual reality. Towards this end, we introduced two types of incentives that emphasized either individual or group performance. The terms individual and public correspond to the two different incentive types to which the participants were exposed: whether they were encouraged to help themselves first (individual) or help the others (public). These two types of incentives were also expected to generate the two map information conditions in the latter study, as individual trials would generate the low map information condition and public trials would generate the high map information condition. The experiments from all three studies were approved by the ethics commission of the university (EK2020-N-73).

Participants
Thirty participants (15 men and 15 women) were recruited via the university registration center for participants. In the end, all participants were between 19 and 35 years old (mean ¼ 23.7). Participants were compensated with a basic incentive of 35 Swiss Francs and a bonus between 0 to 10 Swiss Francs, depending on their evacuation performance during the experiment. On trials with individual incentives, participants received a bonus of 1 Swiss Francs for evacuating within the time allotted. On trials with group incentives, participants received a bonus of 1 Swiss Francs if all participants evacuated within the time allotted. The study was conducted in two separate sessions that each involved 15 participants.

Materials
The experiment was conducted in a laboratory specialized for multi-user virtual reality. The laboratory consists of a series of networked desktop computers in separate cubicles that are connected to a server in a control room. From the control room, the experimenters can monitor and manage the experimental procedure. Similar to previous research , the experimental software was implemented using the Unity game engine. Photon Cloud, a realtime multi-player game development framework for Unity, was used to establish the networking mechanisms between the server and the client. Each participant used a mouseand-keyboard control setup to move an avatar through the virtual environment (Thrash et al., 2015).
The virtual environment for each trial was a grid maze (see Figure 1(b) for an overview of one maze, the other the maze designs were included in the section 1 of Supplementary material). For different trials, we designed five different mazes, each with a field size of 40 m Â 40 m. Randomly selected doorways between rooms were blocked by obstacles. The design of the maze was based on the rule that the difficulty of the optimal routes was approximately consistent across the trials. These mazes were also designed in a manner that the Euler distance between the starting point and the exit location was similar, considering the connectivity of the graph generated with the maze and the fire obstacles. These obstacles were visualized as fires to represent a time-sensitive emergency context. The locations of the fires varied across different mazes and trials. In addition, the locations of the fires were distributed along the edge of the maze and inside of the maze to prevent the participants from speculating about the locations of the exits based on the pattern of the fire locations. The starting and exit locations were designed so that all five mazes had a Human Unspecified -Altruism similar level of complexity by designing with the same minimum distance between the starting location and the exit location. The maze and map application were designed to be minimalistic and easy to parameterize to guarantee consistency across trials and participants in terms of the complexity of the evacuation task. Using a grid maze instead of a realworld building ensured that we could systematically vary fire locations, starting locations, and exit locations. Models of real-world buildings also tend to contain visual and geometric details that could distract the participants and add noise to the information sharing behavior in which we were primarily interested. Participants controlled an avatar from the first-person perspective and were able to move through the virtual environment using the arrow or WASD keys. The maximum forward movement speed was the virtual equivalent of 1.3 m/s, and the maximum backward speed was 0.6 m/s. Participants also used the mouse to rotate their field of view, restricted to a maximum angular velocity of 120 /s. These avatars were represented by wooden mannequins with an eye height of 1.8 m and a collision radius of 0.25 m. The participants were able to see the other participants' avatars when located within their field of view (see Figure 1(a)). A training tutorial was provided before the experiment to allow participants to familiarize themselves with the control interface.
Participants could locate themselves in the virtual maze using a virtual map application that represented their location and direction with a green arrow (see Figure 1(a)). The same application also allowed them to exchange information regarding the locations of fires and exits that they viewed themselves. As participants explored the environment, fire obstacles that appeared within the visible area were marked on the map. Participants were able to share these fire locations with the others. By clicking the spacebar on the keyboard, participants could share the three most recently explored fire locations with all other participants. The amount of fire information that can be shared was set based on a compromise between the total amount of fire locations and the number of participants available to share. We have also set it at three blocks based on the common sense that, when placed in one squared block, the immediate maximum amount of fire information a person could share is three. Once a participant walked into the area from where the exit was visible, the participant was presented with a message that they had found the exit and that they could share it with the other participants. While sharing fire or exit locations, the participants' actions were suspended for 5 s to represent the cost of sharing this information. Based on the suggestion from Burnham (2003), system notifications (e.g., when other players had shared fire or exit information) were displayed on the screen when new information had been updated on the map.

Procedure
After entering the laboratory, each participant was first randomly assigned to a seat number, corresponding to a unique desktop computer. Then, the participants read and signed an informed consent form and were shown how to use the mouse-and-keyboard control interface with an interactive tutorial. Subsequently, the main experiment consisted of ten trials in total. For each of the experimental trials, participants were placed in a maze and instructed to evacuate through the hidden exit. The trial ended after a time limit of 5 minutes. The time limits of each trial were set as five minutes, representing the harmful effects of toxic fumes to approximate real-life survival expectancy (Marsar, 2010). Based on our pilot study, we have found that the time limit can allow the participants to finish navigation in the designated mazes. Each maze design was used for two different trials (once for each incentive type), but there were always at least two trials between pairs of trials with the same map design. The experiment was conducted in two separate mutli-user sessions with different groups of participants, and the order of trials for the second session was the reverse of the order of trials for the first session. Otherwise, the order of trials was randomized.

Measurement
The primary independent variable was incentive type (i.e., individual versus public; within-subjects). During each session, we collected data on several dependent variables, including map-sharing behaviors, evacuation success rates, trajectories, and evacuation times. Map-sharing behaviors were divided into two categories. Fire sharing was the number of times that participants shared fire locations, and exit The incentive type is indicated constantly on the top right corner with colored text (e.g., in this case the type of incentive is public). (b) Illustration of one of the mazes. Participants start from the middle of the maze where the user icon is located. Each fire symbol represents an obstacle. In this example, the exit is located on the right edge of the maze.
sharing was the number of times that participants shared the exit locations. Total sharing was the sum of fire and exit sharing. The trajectories were analyzed in terms of length, the difference between individuals' evacuation paths, and the shortest possible evacuation path (in terms of the number of rooms traversed). Two-tailed, paired-sample t-tests were used to analyze the effect of incentive type on these dependent variables. Kernel density estimates (KDE) (Duong, 2007) were then used to compare incentive types in terms of the distributions of locations along the trajectories.

Results
The overall failure rate of evacuations in Study 1 was very low, with a mean failure rate of 1.00% for trials with an individual incentive and 1.33% for trials with a public incentive (see Figure 2). A two-proportion Z-test showed that these failure rates were not different from each other (Z ¼ 0.010, p ¼ 0.992). The data were normally distributed therefore twotailed t-tests were applied. It revealed that participants in trials with individual incentives shared significantly less than participants in trials with public incentives in terms of fire sharing (mean of individual incentive ¼ 0.32; mean of public incentive ¼  Figure 4). Trajectory differences in the shortest path were also significantly different between trials with an individual incentive (mean ¼ 23.900) and trials with a public incentive (mean ¼ 20.440), This finding demonstrates that the participants performed better with information sharing. KDE analyses for the density of the trajectories (see Figure 5) did not reveal a significant difference between the two types of incentives (p ¼ 0.557). In addition to these inferential statistics, Study 1 successfully provided two types of map information (low and high) that are sufficiently distinguishable to be used as stimuli in Study 2.

Study 2-Altruism during evacuation
The results of Study 1 showed that different incentives successfully generated the sharing behavior and two different amounts of map information. In order to further examine the altruistic behavior in the second study, we removed the incentives of group performance and replayed these two levels of map information from the first study. In contrast to Study 1, we focused on the influence of different map information by not specifying different incentives across trials. Because there were no direct benefits for helping the other participants by sharing map information, we can define this sharing as altruistic behavior.

Participants
Thirty different participants (16 men and 14 women) were recruited via the university registration center for participants. All participants were between 18 and 31 years old (mean ¼ 23.7). Each participant was compensated with a basic incentive of 35 Swiss Francs and a bonus between 0 to 10 Swiss Francs, depending on their performance during the experiment. Each successful evacuation added 1 Swiss Francs to the bonus. There were two experimental sessions that each had 15 different participants.

Materials
The same maze designs and interaction mechanisms from Study 1 were applied in Study 2 (all maze designs were included in the section 1 of Supplementary material). In contrast to Study 1, each participant only interacted with computer-controlled agents, who automatically moved and shared the map in the same manner as participants in Study 1. Each agent corresponded to one of the fifteen participants  The asterisks " ÃÃ " denote significant effects (p < 0.001). For all the graphs in the present paper, error bars represent the standard error of the difference between means. During trials with public incentives, participants shared significantly more fire and total locations than during trials with individual incentives.
from the first session of Study 1. All participants' maps were updated simultaneously based on the agents' sharing activities. Therefore, all participants in Study 2 experienced the same kind of stimulus for each trial. The map information shared by human participants in Study 2 was recorded but not broadcast to the other participants.

Procedure and measurement
The procedure and design of Study 2 were similar to the procedure and measurement of Study 1, with the following exceptions. In contrast to Study 1, participants received the same reward for evacuating in all ten trials. The primary independent variable was map information (low versus high; within-subjects). Agent behaviors from trials with individual incentives were considered the low map information, and agent behaviors from trials with public incentives were considered the high map information.

Results
The failure rate of evacuations in Study 2 was higher than in Study 1 but remained relatively low, with a failure rate of 6.67% for trials with low map information and a failure rate of 0.00% for trials with high map information (see Figure 2). A two-proportion Z-test revealed that these failure rates were significantly different (Z ¼ 4.398, p ¼ <0.001). The data were normally distributed therefore two-tailed t-tests were applied. Figure 6 represents how participants with low map information (mean ¼ 3.24) shared significantly more than participants with high map information (mean ¼ 1.88) in terms of  Figure 6. Mean number of sharing events per participant in Study 2. The asterisks " ÃÃ " denote significant effects (p < 0.001). During trials with low map information, participants shared significantly more fire and total locations than during trials with high map information.
fire sharing (mean of low map information ¼ 2.66; mean of high map information ¼ 1.38), t 29 ¼ 6.462, s.e. ¼ 0.  Figure 7. We have also compared the sum of sharing events for agents and human participants and found that there was no significant difference (t 29 ¼ À0.667, s.e. ¼ 0.235, p ¼ 0.510, d ¼ À0.073) between low map information trials (mean ¼ 3.95) and high low map information trials (mean ¼ 3.81). KDE analyses for the distribution of locations along the trajectories did not reveal a significant difference between the two types of contexts (p ¼ 0.132; see Figure 5).

Comparison between different incentives
In order to compare the results of the different incentives between Study 1 and Study 2, a mixed model 2 Â 2 ANOVAs was used to investigate the interaction of the effects of both the incentives (individual versus public) and map information (low versus high). Their effects on the total number of sharing events, time, and trajectory length are presented in Table 2. A significant main effect of specified incentives was found on time, trajectory length, and difference to the shortest path. Effects of map information were found on the number of sharing events. In addition, there was a significant interaction between incentives and map information on the number of sharing events (p < 0.001).

Implications from lab study and motivation for survey study
The results from the two studies have demonstrated the mechanisms of information exchange in a virtual environment. In particular, the outcome of Study 2 rejected the hypothesis that the cooperative behavior was the consequence of behavior propagation beyond the effects of incentive types. We expected that participants would be more altruistic in trials with high map information than in trials with low map information. Regardless of the benefits of an experimental approach, there were many aspects of emergency information sharing behavior that are easier to cover outside of the laboratory. To investigate the real motivation of sharing emergency information and complex aspects of the psychological implications behind the sharing behavior, in the next section we extend the first two studies with an online survey that can address a broader set of questions and connect our laboratory findings with realworld scenarios.

Study 3-Survey on emergency information sharing behavior
This online survey was conducted using Amazon Mechanical Turk (AMT). The primary goal of these questions was to characterize factors that influence social decision making during emergencies. We sampled a different group of participants because presenting the survey to the same group of participants from the previous laboratory studies would have biased their answers in an undesirable way. With separate participants for the VR experiments and the survey, we can ensure that the two sets of evidence were independent.

Participants and design
Participants must have met the following requirements: At least 18 years old, a good command of English, and regular smartphone use. In total, 252 participants were recruited from AMT. Sixty of these participants were excluded because they failed the attention check (16.7%), confirmed that they did not participate seriously (2.8%), or gave answers that did not match with the open questions (6.0%). Additional details regarding these exclusions are presented in the 2.2 Data quality checks section of the Supplementary material. Each participant was compensated with 1 USD (approximately 8 minutes, with converted salary rate of $7.5/ hour, which is above the minimum wage in the US). Consequently, 192 participants (36.5% female) were included in this analysis. All participants were between 18 and 67 years old (mean ¼ 37.5). The questionnaire was implemented using Qualtrics (https://www.qualtrics.com/) and comprised 30 items (see section 2.3 of the Supplementary material). To minimize the need for scrolling, these items were split into several pages. The sequences of the answers and questions were partly randomized to avoid order bias (McFarland, 1981) (see 2.1 Questionnaire randomization section of Supplementary materials). We collected information regarding the participants' background, social media use, geographic information, experience with emergency scenarios, and responses to a hypothetical emergency scenario.
Most of the participants were from North America or Central America (90.63%), with a small amount from South America (7.29%) and Europe (2.08%). 97.40% of the participants were social media users. Among the social media platforms, Facebook (84.38%), Instagram (66.67%), Twitter (66.15%), and Snapchat (20.83%) were the most common, but multiple answers were allowed. 65.10% of the participants were Android phone users, 32.29% were iPhone users, and 2.60% of them had a smartphone other than iPhone or Android. Google Maps (85.34%) was the most common navigation application, followed by Apple Maps (9.42%) and Waze (5.24%). Table 3 demonstrates that a large percentage (66.15%) of the participants had previously experienced emergency accidents such as fire accidents, earthquakes, hurricanes, tornadoes, or other severe natural disasters. This finding supports the external validity of our survey sample. We differentiate between participants who have experienced emergencies and those who have not by dividing the sample into experienced and inexperienced categories.

Current emergency information sharing mechanisms
Across all participants, user-generated content from social media is the most common choice (61.46%) for information seeking during an emergency, probably due to its timeliness. When asked about existing emergency sharing mechanisms, 34.38% of the participants chose Facebook Marked Safe and 14.58% chose the Waze report function. When asked what other media or platforms are the best channels for sharing emergency information, participants mostly chose social media platforms such as Facebook, Twitter, and/or instant messaging. However, traditional media, including television, local news, and radio, are still common answers for questions regarding information sharing channels. The vast majority of the experienced participants (76.38%) have shared emergency information with others during emergency accidents. Almost half of the experienced participants (49.61%) have already used one of the emergency functions (i.e., Facebook Marked Safe, Waze report function) to share emergency information with the general public. Among the inexperienced participants, most of the answers demonstrate similar trends to those of the experienced participants, except for the higher concerns regarding the sharing of unconfirmed information (43.08%).

Factors influencing willingness to share
An imaginary fire scenario and eleven questions were designed to investigate the factors influencing respondents' willingness to share. Each factor was addressed in one question (see section 2.3 of the Supplementary materials), which was formatted as a 7-point scale from very unlikely (¼1) to very likely ( ¼ 7). The Control factor investigates their general tendency to share emergency information. Each of the remaining ten factors consisted of the control question ("how likely … ") plus a specific condition related to the investigated factor. In total, eleven factors were presented to the participants. They are summarized as follows: Effectiveness: … if there were other users nearby that could benefit from this information Notification: … if you had seen a notification that this information could be useful for other users First hand: … if you knew that you were the first one to observe this information  . Evacuation performance in terms of time and path in Study 2. The asterisk " Ã " denotes a significant effect (p < 0.05). Participants required significantly more time to evacuate and used paths further from the shortest path in trials with low map information compared to trials with high map information. The asterisks " Ã " denote a significant effect on the level p < 0.05.
Control: how likely would you be to share the fire information on your map application Social relationship: … if someone you know was also using this map-sharing application in such situations Anonymity: … if others did not know that it was you who shared this information Role model: … if you had seen that other users used the map-sharing function to report other accidents Update: … if you had seen that this information was not updated Confirmation: … if this information had already been reported Credit: … if others knew that it was you who shared this information Risk: … if you knew that it was not safe for you to stay near the fire area The scores of all these factors are presented in Figure 8. Among these factors, the effectiveness factor had the highest average score of 6.0, which suggests that people were more willing to help if they knew there was someone in need. The notification factor and first hand factor had the second and third highest scores. The high value of the importance of the first hand factor was consistent with our findings from Study 2. In Study 2, if people knew they were among the first to observe an emergency situation, they tended to share more often. The risk factor had the lowest score, which demonstrates the rationality of the choice to ensure their own safety before helping others. The credit factor had the second lowest score, which shows their relatively less willingness to share such information when exposing their true identity. Participants who reported having experience with emergency accidents (n ¼ 127) were categorized as "Experienced" for the remainder of this section. "Inexperienced" participants were those who responded "no" to this question (n ¼ 65). Questions marked with (M) can have multiple answers and questions marked with (H) could have been answered hypothetically. If a question could have multiple answers, the value for each answer represents the percentage of the participants from the particular sample (i.e., total, experienced, or inexperienced) who chose that answer. For questions that could have had multiple answers and also contained a "no" choice, the "no" answers were exclusive. In other words, the other options became inactive once the "no" option was chosen.  We also collected participants' general feedback regarding emergency information sharing. Several participants mentioned the ongoing wildfire situations in California. For example, one person said," I would use and could use, something like this as we have fires everywhere here in California" (P54). The severe situation and unpredictable nature of fire disasters suggest to the victims that such an emergency sharing concept could be very useful. The reliability of information is another general concern of some participants. For example, one respondent said," I need to have real confidence in what I am sharing. I'm afraid of giving people bad information and causing panic for no good reason" (P12). Finally, another issue can be the misuse of disaster information. For example, one participant commented," Sharing incidents can be helpful, but my main concern with sharing information is that it could be misused by other people. Not everyone would stay away from an emergency situation. Just as there are tornado chasers that willingly drive themselves into danger" (P167). An information sharing platform might be misapplied by curious internet users or journalists to exploit the unusual situation. Granting access to information exclusively for the target users (e.g., users with geographic proximity) might be another essential feature to be considered.

Discussion and implications
The results of our three studies demonstrate the effects of collective intelligence on evacuation behavior and the impact of incentives and publicly shared map information on sharing behavior. Study 1 presents the influence of different incentives on people's willingness to share, whereas Study 2 focuses on the influence of the publicly shared map information on people's sharing behavior. Both laboratory studies have demonstrated the effect of spatial collective intelligence on evacuation performance. Moreover, the online survey further validates the findings from the laboratory experiments by highlighting participants' motivation to fill the knowledge gap in addition to other factors that influence the willingness to share emergency information, such as concerns for privacy and effectiveness. These findings contribute in several ways to our understanding of the power of sharing behavior and provide a basis for applications that aim to encourage such behavior. In this section, we will discuss our findings in terms of the following three implications: 1. enhancing evacuation performance with collective intelligence; 2. motivating people to share despite an information gap; 3. influencing factors for the willingness to share. Afterwards, we also present the limitations and general discussion.

Implication 1: Enhancing evacuation performance with collective intelligence
Regarding the evacuation performance, Study 1 found that the metric of trajectory length supported the hypothesis (H1) that participants with a public incentive performed better. Although the difference is not significant in terms of time, we observed an interesting phenomenon that participants in trials with public incentives stayed longer to help other participants by walking around the exit to attract other participants' attention even after they have found the exit. This unexpected behavior might contribute to the similar evacuation times for the two types of incentives. The failure rates of these evacuations were low for both types of trials, which indicated that the mazes were not very complicated for the participants when driven by time-sensitive incentives.
In the case of Study 2, the hypothesis (H3) that participants performed better with high map information was supported in terms of evacuation time. When given more information about the fire obstacles, the participants required less time to find the exit. There may have been a significantly higher number of sharing events in trials with low map information because of a reduction in individuals' evacuation efficiency. Another important finding is that all failed players in Study 2 are from the trials with low map information. This difference in failure rates also implies that participants performed better with high map information. This combination of findings provides support for the premise that map information has an influence on the altruistic behavior of individuals. This observation from both Studys 1 and 2 suggests that the effect of collaboration on information sharing during evacuations can help evacuees to shorten their evacuation time and path. These results corroborate the findings of previous work regarding the beneficial effect of group membership on emergency situations (Drury et al., 2009;K. Li et al., 2019;Starbird & Palen, 2011).

Implication 2: Motivating people to share despite an information gap
Another implication of our research regards the motivation to share. Study 1 reveals that, when specific incentives are offered, participants are more cooperative in trials with a public goal than participants in trials with an individual goal (H2). Specifically, participants shared more map information when other players' evacuation performances would improve their own compensation. This confirms previous findings on the positive effects of incentives on cooperation behavior (Goldman et al., 2007;Wageman & Baker, 1997).
However, the outcome of Study 2 conflicted with the possibility that this cooperative behavior was the consequence of behavior propagation beyond the effects of incentive types. In Study 2, we used a similar experimental setup as in Study 1, but we focused on the possibility of altruistic behavior in different information content conditions in the absence of specific incentives. In contrast to earlier findings, the hypothesis (H4) that participants would be more altruistic in trials with high map information than in trials with low map information was rejected. In contrast with the findings of Kang and colleagues (Kang et al., 2014), we did not observe the propagation of these altruistic behaviors among the participants. This rejected hypothesis is also contrary to previous studies that emphasize the role of propagation of trust (Guha et al., 2004) and judgment  in social scenarios.
There are several possible explanations for this result. The absence of information in the public information pool might actually encourage people to share more. Similarities in the sum of sharing events between agents and participants and across different map information also support this conclusion. The overall willingness to fill the empty public information pool is of consistent value for a group. Combined with the higher failure rate of evacuation in trials with low map information, this finding implies that true altruistic behavior emerged in Study 2, especially because the participants were willing to share more despite experiencing less sharing from the others. The online survey in Study 3 also confirms that the factor of having firsthand information is the third biggest factor influencing the willingness to share.
When comparing the results from the first two studies, it appears that there exists an interaction between the presence of an incentive and the amount of publicly shared map information in terms of the number of sharing events. This indicates that sharing behaviors were indeed driven by the participants' personal interests and the behavior they observed from the others. We would also propose that feelings of empathy may be transferred through virtual avatars. In a digitized society in which users commonly interact with each other remotely, this research contributes to important questions on the effects of the virtual appearance of a real human on social behavior. Another interesting finding is the difference in failure rates between Study 1 and Study 2. The failure rate of the low map information trials in Study 2 is higher than that in Study 1 (see Figure 2), which can be explained in part by the time-sensitive nature of the tasks. In Study 1, both types of incentives introduced a certain level of time pressure, while in Study 2, the incentives were less time-sensitive despite the same time limit. Such a reduced time pressure can contribute to the higher number of map sharing events, but it may have resulted in a higher failure rate. The impact of spatial collective intelligence also varied in both studies. Participants who were exposed to more collective intelligence (i.e., more map information within the corresponding condition) took routes that were closer to the shortest path in both studies. In terms of absolute length and time, the differences varied between the two studies. While the participants who were exposed to more spatial collective intelligence traveled less in Study 1, the participants who were exposed to more spatial collective intelligence in Study 2 needed a shorter time to evacuate. Considering the difference in failure rates between the two studies, it can be inferred that spatial collective intelligence is helpful for evacuees, although it can be reflected by different measurements.

Implication 3: Influencing factors for the willingness to share
In Study 3, an online survey contributes to identifying the factors and concerns of emergency information sharing strategies of a broader public. The results revealed that social media and internet remain the major information resource during emergencies. The majority of emergency information sharing is still between acquaintances via instant messages and social media, but the possibility to broadcast information sharing is emerging (e.g., the Waze report function). When participants were asked about what factors would influence their willingness to share in an imaginary fire scenario, the effectiveness of the information appears to be the most important factor. Combined with the prominence of system notifications and the first-hand factor, this finding implies that future applications should emphasize a user interface that can inform their users about the necessity of sharing emergency information. Contrary to the findings from Study 2, in which the laboratory participants demonstrated altruism in helping others despite detrimental effects on their own evacuation performance, survey participants highly valued the importance of ensuring their own safety. This inconsistency may be due to the limited realness and pressure in the virtual environment. Unlike the laboratory participants, survey participants might have had to imagine the danger of accidents, which could have provoked more rational decision-making. The presence of other participants and/or experimenters might have induced the laboratory participants in Study 2 to share more than usual. Unlike previous research on the security of crowdsourced information for public maps (Wang et al., 2018), these findings may help us to emphasize the importance of guaranteeing users' own safety before allowing them to help others when designing such an emergency information sharing mechanism. The survey results can also help application designers to improve the amount of activity elicited from potential information contributors.

Limitations and general discussion
Despite the insights gained from this study, our conclusions are limited by the absence of feelings of real danger. The comfortable laboratory environment may have neutralized the stressful context of the virtual emergency evacuation. Since it was neither possible nor ethical to place the participants in a real fire scenario, it is unknown if the panic and time-sensitive nature of a real fire would lessen altruism in map sharing or other crowd behaviors. Such behaviors may have also unfolded differently in a more complex environment (Li et al., 2022;Liu et al., 2021) compared to a simple virtual maze. Indeed, the overall high success rate (>90%) in the present studies implies that the participants did not find the evacuation task particularly challenging, which is usually not the case in real-world fire evacuations (F. Li et al., 2014). In real-world hazard events, the possibility of the evacuation will highly depend on the context and the level of the threat of danger. Future research can focus on examining more complicated evacuation environments which can further challenge the egress ability of the participants.
As a viable alternative, virtual presentations of key information have been used in the security domain (Lukosch et al., 2015). We used VR to balance realistic social behavior during evacuations in simple environments with the potential harm caused by realistic evacuations in complex environments. Similarly, Kinateder and colleagues' work has demonstrated that VR is critical for investigating decision-making in normally dangerous evacuation scenarios (Kinateder et al., 2018). In addition, prior work has shown that simple evacuations in VR follow similar dynamics as those in real environments (Moussaïd et al., 2016). For the present studies, VR allowed us to precisely measure and systematically vary map sharing behavior with both human participants and computercontrolled agents. Highly controlled environments tend to induce more consistent decision-making across scenarios and improve our ability to infer such processes reliably. Future work should focus on how real-world evacuation studies Xue et al., 2021) may be conducted in a safe and ethical manner. Another limitation of the study is the channel for the information sharing. In our study, the participants were expected to share hazard source information via a virtual digital device. It is designed in such a way that for a similar event but on a larger scale, the digital transmission methods can still be possible. However, in real-world evacuation, faceto-face voice communication with encountered evacuees might be the most likely method to know information for safer evacuation. Future research can examine more diverse applications with multimodal communication for actual use in emergency evacuations.
Overall, our study was able to identify the impact of incentives and publicly shared map information on information sharing behavior. These findings raise a series of intriguing questions regarding the mechanism and utility of collective intelligence during hazardous events. These findings can also help us to understand the role of cooperation and altruism in an emergency evacuation. The cooperative behavior of exchanging spatial information is found to be beneficial for shortening the evacuation path. However, the altruistic behavior of exchanging spatial information is found to be less effective in improving the evacuation success rate when incentives are not time-sensitive. This suggests that future emergency information sharing systems should balance between ensuring the individual safety of the users and encouraging them to contribute to the public information pool. The importance of other factors, such as information effectiveness, privacy, and information reliability, was raised by the online survey. Together, these findings can enable future collaborative evacuation systems with a human-centered design.

Conclusion
The present investigation provides a first step towards establishing a protocol for exchanging collective intelligence in emergency evacuations. Specifically, we used a series of virtual mazes and an interactive map application to study the participants' willingness to share evacuation information with each other. Study 1 revealed that participants tended to share more when given a public incentive than when given an individual incentive. Study 2 extracted the effect of incentives from Study 1 and demonstrated the existence of altruistic behavior. Participants showed more willingness to share even facing lower publicly shared map information by others. The importance of collective intelligence is clearly supported by the trajectory length in Study 1 and the time used to evacuate in Study 2. Both results show that the more people contributed to the collective intelligence information pool, the better they perform during the evacuation. In Study 3, an online survey was conducted to investigate further the emergency information sharing mechanism with a broader choice of questions. We have found that social media is still one of the most important sources for emergency information sharing. Participants also revealed their concerns about information misuse and reliability. The findings of this research provide valuable insights into the mechanism of spatial collective intelligence in emergencies. They suggest that altruistic behavior can also be achieved and promoted in the absence of information in a public resource pool. The approach of using a multiplayer networked environment will prove useful in expanding our understanding of how to conduct behavioral science studies of dangerous situations (e.g., fire evacuation, crowd disaster), which are impossible to implement in real life. The insights gained from this study can be of assistance to establish a standard for the next generation of crowdsourced information applications for emergency evacuations.
Notwithstanding the limitations, this work offers valuable insights into cooperative and altruistic behaviors in emergencies. Collective intelligence plays an increasingly important role in the age of digital disruption and sharing economies. To our knowledge, this research is the first comprehensive investigation of collective intelligence in virtual emergency evacuation. Continued efforts are needed to investigate the interaction design and the underlying mechanisms of sharing spatial collective intelligence with digital technologies, especially in emergencies.

Disclosure statement
No potential conflict of interest was reported by the author(s).

Dirk
Helbing is a Professor of Computational Social Science at the Department of Humanities, Social and Political Sciences at ETH Zurich and affiliate of its Computer Science Department. He is a member of the external faculty of the Complexity Science Hub Vienna and cofounded the Decision Science Laboratory at ETH.
Christoph H€ olscher is a Full Professor of Cognitive Science in the DGESS at ETH Zurich since 2013, with an emphasis on Applied Cognitive Science. Since 2016 Christoph is a Principal Investigator at the Singapore ETH Center Future Cities Laboratory and is the Program Director of Future Resilient Systems.